Bugs when fine-tune llama-2-7b with instructions using llama-2's conversation template

[YJiangcm]

Thanks for your great work! I met some problems when using fastchat/train/train.py to fine-tune a llama-2-7b by using llama-2's conversation template.

I have changed the get_conversation_template("vicuna") to get_conversation_template("llama-2"), and delated assert conv.sep_style == SeparatorStyle.ADD_COLON_TWO. However, the tokenization mismatch warning was reported, and the training loss was always 0.

WARNING: tokenization mismatch: 78 vs. 80.
#turn = 1. (ignored)

Could you please tell me how to adapt the code to llama-2's conversation template? Thanks a lot!