iandennismiller's picture
map model name to template, map template to prefix/suffix
6cc9d7b
raw
history blame
4.63 kB
#!/bin/bash
function llama_init_environment {
# if conf does not exist, create it
if [ ! -f "$HOME/.config/llama/llama-main.conf" ]; then
mkdir -p "$HOME/.config/llama"
cat <<EOF > "$HOME/.config/llama/llama-main.conf"
LLAMA_MODEL_NAME=teknium/OpenHermes-2.5-Mistral-7B/openhermes-2.5-mistral-7b-f16.gguf
LLAMA_TEMPLATE=chatml
LLAMA_MODELS_PATH=$HOME/.ai/models/llama
LLAMA_PROMPT_PATH=$HOME/.local/share/llama/prompts
LLAMA_CACHE_PATH=$HOME/.cache/llama
LLAMA_CONTEXT_SIZE=4096
LLAMA_TEMPERATURE=1.6
LLAMA_TOP_P=1.0
LLAMA_MIN_P=0.1
LLAMA_TOP_K=0
LLAMA_REPETITION_PENALTY=1.15
EOF
fi
source $HOME/.config/llama/llama-main.conf
if [ ! -d "$LLAMA_CACHE_PATH" ]; then
mkdir -p "$LLAMA_CACHE_PATH"
fi
}
function llama_interactive {
# if 4 arguments are provided...
if [[ $# -eq 2 ]]; then
LLAMA_MODEL_NAME=$1
LLAMA_CONTEXT_SIZE=$2
fi
# if fewer than 4 arguments - but greater than 0 - are provided, then display error message
if [[ $# -lt 2 ]] && [[ $# -gt 0 ]]; then
echo "Error: 4 arguments are required."
echo "Usage: llama <model_name> <context_size>"
echo "Example: llama teknium/OpenHermes-2.5-Mistral-7B/openhermes-2.5-mistral-7b-f16.gguf 4096"
return
fi
LLAMA_TEMPLATE=$(get_template_for_model $LLAMA_MODEL_NAME)
llama \
--n-gpu-layers 1 \
--model "$LLAMA_MODELS_PATH/$LLAMA_MODEL_NAME" \
--prompt-cache "$LLAMA_CACHE_PATH/${LLAMA_MODEL_NAME//[\/\.]/-}-${LLAMA_CONTEXT_SIZE}.cache" \
--file "$(get_model_prompt $LLAMA_MODEL_NAME)" \
--in-prefix "$(get_model_prefix $LLAMA_TEMPLATE)" \
--in-suffix "$(get_model_suffix $LLAMA_TEMPLATE)" \
--reverse-prompt "$(get_model_prefix $LLAMA_TEMPLATE)" \
--reverse-prompt "<|im_end|>" \
--reverse-prompt "</s>" \
--threads 1 \
--temp 0 \
--top-p 1.0 \
--top-k 40 \
--min-p 0.5 \
--repeat-penalty "$LLAMA_REPETITION_PENALTY" \
--ctx-size "$LLAMA_CONTEXT_SIZE" \
--batch-size 1024 \
--n-predict -1 \
--keep -1 \
--instruct \
--no-mmap \
--color \
--escape \
--log-disable
# --temp "$LLAMA_TEMPERATURE" \
# --top-p "$LLAMA_TOP_P" \
# --top-k "$LLAMA_TOP_K" \
# --min-p "$LLAMA_MIN_P" \
}
function get_model_prefix {
case $1 in
chatml)
printf "<|im_start|>user\\\n"
;;
instruct)
printf "[INST] "
;;
alpaca)
printf "### Instruction: "
;;
vicuna)
printf "USER: "
;;
*)
printf "Input: "
;;
esac
}
function get_model_suffix {
case $1 in
chatml)
printf "<|im_end|>\n<|im_start|>assistant\\\n"
;;
instruct)
printf " [/INST]"
;;
alpaca)
printf "### Response:"
;;
vicuna)
printf "ASSISTANT:"
;;
*)
printf "Output:"
;;
esac
}
function get_template_for_model {
case $1 in
*dolphin*)
echo "chatml"
;;
*mixtral-8x7b-instruct*)
echo "instruct"
;;
*upstage*|*airoboros*|*hermes*)
echo "alpaca"
;;
*wizardlm*|*samantha*|*scarlett*|*capybara*)
echo "vicuna"
;;
*)
echo "chatml"
;;
esac
}
function get_model_prompt {
case $1 in
*guanaco*)
echo "$LLAMA_PROMPT_PATH/guanaco.txt"
;;
*alpaca*|*upstage*|*airoboros*|*hermes*)
echo "$LLAMA_PROMPT_PATH/alpaca.txt"
;;
*dolphin*)
echo "$LLAMA_PROMPT_PATH/chatml-gpt.txt"
;;
*vicuna*|*wizardlm*)
echo "$LLAMA_PROMPT_PATH/vicuna-v11.txt"
;;
*scarlett*)
echo "$LLAMA_PROMPT_PATH/scarlett.txt"
;;
*samantha*)
echo "$LLAMA_PROMPT_PATH/samantha.txt"
;;
*based*)
echo "$LLAMA_PROMPT_PATH/based.txt"
;;
*capybara*)
echo "$LLAMA_PROMPT_PATH/capybara.txt"
;;
*yi*)
echo "$LLAMA_PROMPT_PATH/yi.txt"
;;
*orca*)
echo "$LLAMA_PROMPT_PATH/orca.txt"
;;
*)
echo "$LLAMA_PROMPT_PATH/idm-gpt-lite.txt"
;;
esac
}
llama_init_environment
llama_interactive $*