qlm

#function qlm {
#This function body is to be autoloaded, along with its completion companion _qlm.
#Place this file in your fpath and `autoload -Uz qlm _qlm` from your .zshrc
#or just 'autoload -Uz full/path/to/qlm (same for the completion function _qlm)
#Do not forget to register the completion function in .zshrc with `compdef _qlm qlm`, (after compinit). 
#The Large language model names and filenames to select from can be declared/added in qlm.cfg.

#-----Local Configuration, adjust as/when needed:
#local inference engine, assuming llama.cpp (`llam` is a symbolic link to `llama-cli`), options preset for it: 
local inferengine='llam'
#Location of local language models:
#in $LLMDIR, used here and defined in qlm.cfg 
#Default model name (a valid key in llmodels, defined in qlm.cfg) when no model name provided:
local defaultllm="Qwen-14B"
#We can fill this prompt file from other programs, or populate it with the (alias) function `promf`:
#TPROMPTF='/dev/shm/promf'  DO NOT uncomment here: Used in other functions, so declared globally in qlm.cfg.
#NOTE: If $1 and TPROMPTF don't exist, the code will look for any selected text and place it in the context.
#-----End Local Configuration

#Source model names and config. parameters:
source $ZDOTDIR/.zfunc/qlm.cfg

#This collects any (overriding) options that the user may want to pass directly to llama-cli
#These must always be followed by "--" to delimit from model name and prompt, which will follow.
local posit=${argv[(Ie)--]}
if (( $posit )); then
   local llamacli_opts=${argv:1:$posit-1} 
   shift $(( $posit ))
fi

if [[ $# -gt 2 ]]; then
    echo "Usage: $0 [-flags [args] -- ] [ModelName] ['Prompt']"
    return 1 
elif [[ $# -ge 1 ]]; then
   # Check if the first argument is a valid LLM.
   if [[ -n "${llmodels[$1]}" ]]; then
       local choice=$1
       shift 
   else
       # Raise an error and exit if not found and there are 2 command-line arguments:
       [[ -n "$2" ]] && { echo "Error: '$1' is not a valid large language model." ; return 1 ; }
       #If the only argument is not a model name, it must be a prompt. Load the default model:
       local choice=$defaultllm
   fi
else
   #No arguments. Load the default LLM:
   local choice=$defaultllm
fi

(( 2 - ${#funcstack} )) && echo -e "${choice}\n" > /dev/shm/reqlm
local userinput="${1:-$(cat $TPROMPTF 2>/dev/null || xsel -op)}"
echo -e "${userinput}\n" >> /dev/shm/reqlm
#The local inference engine (defined above) must be in the PATH.
#The 'UIUIUI' placeholder is used to keep config in a single file, that is sourced on startup. 
$inferengine -t 8 --temp $temps[$choice] -fa --top-p 0.95 -mli --no-mmap --mlock --no-display-prompt --no-perf --no-warmup 2>/dev/null -c $ctxsize[$choice] -ngl $gpulayers[$choice] -s $(date +%s) ${1:+-no-cnv} -m $LLMDIR/$llmodels[$choice] -p "${${llmprompts[$choice]}/UIUIUI/$userinput}" ${(z)llamacli_opts} | tee -a /dev/shm/reqlm
#}