From be9c4050b09ebb0cfc96dc36c73375f0fbacb160 Mon Sep 17 00:00:00 2001 From: "Dimitar (Mitko) Asenov" Date: Thu, 8 Feb 2024 07:34:54 -0800 Subject: [PATCH] Remove `--xla_gpu_simplify_all_fp_conversions` in commands where it is set to the default value of `true`. PiperOrigin-RevId: 605311817 --- paxml/contrib/gpu/scripts_gpu/benchmark_gpt_multinode.sh | 2 +- paxml/contrib/gpu/scripts_gpu/run_lambada_singlenode.sh | 2 +- paxml/contrib/gpu/scripts_gpu/run_pile_multinode.sh | 2 +- paxml/contrib/gpu/scripts_gpu/run_pile_singlenode.sh | 2 +- 4 files changed, 4 insertions(+), 4 deletions(-) diff --git a/paxml/contrib/gpu/scripts_gpu/benchmark_gpt_multinode.sh b/paxml/contrib/gpu/scripts_gpu/benchmark_gpt_multinode.sh index 6421a049d..467c970b5 100644 --- a/paxml/contrib/gpu/scripts_gpu/benchmark_gpt_multinode.sh +++ b/paxml/contrib/gpu/scripts_gpu/benchmark_gpt_multinode.sh @@ -26,7 +26,7 @@ LOG_DIR=${5:-"test_logdir"} export VOCAB_PATH=None export XLA_PYTHON_CLIENT_MEM_FRACTION=${XLA_PYTHON_CLIENT_MEM_FRACTION:-0.85} BASE_XLA_FLAGS=${BASE_XLA_FLAGS:-"--xla_gpu_enable_latency_hiding_scheduler=true --xla_gpu_enable_triton_gemm=false - --xla_gpu_simplify_all_fp_conversions --xla_gpu_enable_async_all_gather=true + --xla_gpu_enable_async_all_gather=true --xla_gpu_enable_async_reduce_scatter=true --xla_gpu_enable_highest_priority_async_stream=true --xla_gpu_enable_triton_softmax_fusion=false --xla_gpu_all_reduce_combine_threshold_bytes=51200 --xla_gpu_graph_level=0 --xla_gpu_enable_async_all_reduce=true"} diff --git a/paxml/contrib/gpu/scripts_gpu/run_lambada_singlenode.sh b/paxml/contrib/gpu/scripts_gpu/run_lambada_singlenode.sh index 119fcdc8e..055c1670b 100755 --- a/paxml/contrib/gpu/scripts_gpu/run_lambada_singlenode.sh +++ b/paxml/contrib/gpu/scripts_gpu/run_lambada_singlenode.sh @@ -27,7 +27,7 @@ LOG_DIR=$6 export VOCAB_PATH=$VOCAB_PATH BASE_XLA_FLAGS=${BASE_XLA_FLAGS:-"--xla_gpu_enable_latency_hiding_scheduler=true --xla_gpu_enable_triton_gemm=false - --xla_gpu_simplify_all_fp_conversions --xla_gpu_enable_async_all_gather=true + --xla_gpu_enable_async_all_gather=true --xla_gpu_enable_async_reduce_scatter=true --xla_gpu_enable_highest_priority_async_stream=true --xla_gpu_enable_triton_softmax_fusion=false --xla_gpu_all_reduce_combine_threshold_bytes=51200 --xla_gpu_graph_level=0 --xla_gpu_enable_async_all_reduce=true"} diff --git a/paxml/contrib/gpu/scripts_gpu/run_pile_multinode.sh b/paxml/contrib/gpu/scripts_gpu/run_pile_multinode.sh index b40444449..48cf130b9 100644 --- a/paxml/contrib/gpu/scripts_gpu/run_pile_multinode.sh +++ b/paxml/contrib/gpu/scripts_gpu/run_pile_multinode.sh @@ -27,7 +27,7 @@ LOG_DIR=${6:-"test_logdir"} export VOCAB_PATH=$VOCAB_PATH export XLA_PYTHON_CLIENT_MEM_FRACTION=${XLA_PYTHON_CLIENT_MEM_FRACTION:-0.85} BASE_XLA_FLAGS=${BASE_XLA_FLAGS:-"--xla_gpu_enable_latency_hiding_scheduler=true --xla_gpu_enable_triton_gemm=false - --xla_gpu_simplify_all_fp_conversions --xla_gpu_enable_async_all_gather=true + --xla_gpu_enable_async_all_gather=true --xla_gpu_enable_async_reduce_scatter=true --xla_gpu_enable_highest_priority_async_stream=true --xla_gpu_enable_triton_softmax_fusion=false --xla_gpu_all_reduce_combine_threshold_bytes=51200 --xla_gpu_graph_level=0 --xla_gpu_enable_async_all_reduce=true"} diff --git a/paxml/contrib/gpu/scripts_gpu/run_pile_singlenode.sh b/paxml/contrib/gpu/scripts_gpu/run_pile_singlenode.sh index 30e8db08a..df8a91e54 100644 --- a/paxml/contrib/gpu/scripts_gpu/run_pile_singlenode.sh +++ b/paxml/contrib/gpu/scripts_gpu/run_pile_singlenode.sh @@ -27,7 +27,7 @@ LOG_DIR=${6:-"test_logdir"} export VOCAB_PATH=$VOCAB_PATH BASE_XLA_FLAGS=${BASE_XLA_FLAGS:-"--xla_gpu_enable_latency_hiding_scheduler=true --xla_gpu_enable_triton_gemm=false - --xla_gpu_simplify_all_fp_conversions --xla_gpu_enable_async_all_gather=true + --xla_gpu_enable_async_all_gather=true --xla_gpu_enable_async_reduce_scatter=true --xla_gpu_enable_highest_priority_async_stream=true --xla_gpu_enable_triton_softmax_fusion=false --xla_gpu_all_reduce_combine_threshold_bytes=51200 --xla_gpu_graph_level=0 --xla_gpu_enable_async_all_reduce=true"}