A simple friendly llama-cli executor

llama-cpp.sh

#!/usr/bin/env bash

usage_exit() {
	cat <<!
usage: $0 -p path model

The path to GGUF models can be the root/base path like \$HOME/.models or \$HOME/.cache/huggingface.
The model name can be an unambiguous prefix like qwen-3-30b-a3b.
Pass any flags to llama-cli through LLFLAGS:
    LLFLAGS="-ngl 99 -fa" llama-cpp -p ~/.cache/huggingface gemma-4
!
    exit 1
}

args=`getopt p: $*`
if [ $? -ne 0 ]; then
	usage_exit $0
fi
set -- $args

while :; do
	case "$1" in
	-p)
		gguf_path="$2"
		shift; shift
		;;
	--)
		shift; break
		;;
	esac
done

if [ $# -ne 1 ]; then
	usage_exit $0
fi

llama_cli=`type llama-cli`
if [ $? -ne 0 ]; then
	echo 'llama-cli is not found in path' 1>&2
	exit 2
fi

llama_cli=`echo $llama_cli | cut -d' ' -f3` # path to llama-cli
model=`echo $1 | tr A-Z a-z`
nm=0
matches=""
matched=""

 while read m; do
	m=`echo $m | tr A-Z a-z`
	r=`expr "$m" : ".*$model"`
	if [ $? -eq 0 ]; then
		nm=`expr 1 + $nm`
		matched=$m
		matches="$matches `basename $matched .gguf`"
	fi
done < <(find $gguf_path -name '*.gguf' -print)

if [ $nm -eq 0 ]; then
	echo "no models matched $model" 1>&2
	exit 3
elif [ $nm -gt 1 ]; then
	echo "$model is ambiguous, multiple matches -$matches" 1>&2
	exit 4
fi

if [ -n "$LLFLAGS" ]; then
	llama_cli="$llama_cli $LLFLAGS"
fi
$llama_cli -m $matched