`_name_or_path` is not stored in llama config.json any more

[nrailg]

https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/convert_llama_weights_to_hf.py#L237

but it is still used in src/alpaca_farm/common.py:get_pretrained_model_name_with_model_name_or_path to obtain model family.

At the sft stage, I used a local directory path as --model-name-or-path, and the local dir path is stored as _name_or_path in config.json. Then get_pretrained_model_name_with_model_name_or_path loaded llama config.json and failed to lookup '_name_or_path'.

[lxuechen]

Thanks for raising this issue. This has been resolved in a recent PR.

[nrailg]

Thanks for raising this issue. This has been resolved in a recent PR.

thank u .