Files
open-instruct/scripts/weights/convert_llama_weights_to_hf.sh
Nathan LambertandGitHub 7250a01932 Docs for scripts (#221)
* up

* clean

* up

* add example for eval

* up
2024-08-08 09:01:45 -04:00

9 lines
393 B
Bash
Executable File

LLAMA_FOLDER=/net/nfs.cirrascale/allennlp/jacobm/llama/llama/models
for MODEL_SIZE in 7B 13B 30B 65B; do
echo "Converting Llama ${MODEL_SIZE} to HuggingFace format"
python -m transformers.models.llama.convert_llama_weights_to_hf \
--input_dir $LLAMA_FOLDER/ \
--model_size $MODEL_SIZE \
--output_dir /net/nfs.cirrascale/allennlp/yizhongw/hf_llama_models/${MODEL_SIZE}
done