[BUG]: AttributeError: 'LlamaActor' object has no attribute 'resize_token_embeddings'

[ZhangMaoTai]

🐛 Describe the bug

When run the Stage3 code https://github.com/hpcaitech/ColossalAI/blob/main/applications/Chat/examples/train_prompts.py using LLaMA, this bug is encountered in line 137: tokenizer = prepare_llama_tokenizer_and_embedding(tokenizer, actor)

The detailed of this bug:

Traceback (most recent call last):
  File "experience.py", line 183, in <module>
    tokenizer = prepare_llama_tokenizer_and_embedding(tokenizer, actor)
  File "/XXX/XXX/ColossalAI/applications/Chat/coati/utils/tokenizer_utils.py", line 37, in prepare_llama_tokenizer_and_embedding
    smart_tokenizer_and_embedding_resize(
  File "/XXX/XXX/ColossalAI/applications/Chat/coati/utils/tokenizer_utils.py", line 68, in smart_tokenizer_and_embedding_resize
    model.resize_token_embeddings(len(tokenizer))
  File "/root/anaconda3/envs/coati/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1269, in __getattr__
    raise AttributeError("'{}' object has no attribute '{}'".format(
AttributeError: 'LlamaActor' object has no attribute 'resize_token_embeddings'

This bug is the same as this one: https://github.com/hpcaitech/ColossalAI/issues/3389

Environment

No response

[JerryYao80]

Use the SFT LLaMa-7B model in stage 1 , not the original LLaMa-7B

[ZhangMaoTai]

Thanks! By the way, if I use tokenizer.pad_token = tokenizer.eos_token to avoid this bug, will it make a big difference in training?