Quick question on the need for `KeywordsStoppingCriteria`

Hello Team,

First off, thanks for your work! It is truly a great project!

Quick question on the `KeywordsStoppingCriteria` [implementation](https://github.com/DAMO-NLP-SG/VideoLLaMA2/blob/main/videollama2/mm_utils.py#L326-L357). I see that it is a custom stopping criteria appended to other forms like EOSTokenCriteria,MaxLengthCriteria etc..

As far as I understood, The `EOSTokenCriteria` used [here](https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/generation/utils.py#L902) looks if the predicted token is an EOS token which is `151645` which I believe is the same as the `KeywordsStoppingCriteria` functionality where the `['<|im_end|>']` is tokenized as `151645`. I feel they do the same thing. Return `True` if the predicted token is `151645`.

I tried passing `None` to `stopping_criteria` [here](https://github.com/DAMO-NLP-SG/VideoLLaMA2/blob/main/videollama2/__init__.py#L108) and it did not change the behavior. Is there a reason we need both?

Compare the KeywordsStoppingCriteria implementation in videollama2/mm_utils.py with the generation stopping criteria referenced in Transformers, then inspect how stopping_criteria is passed in videollama2/__init__.py. Determine whether both mechanisms are needed and document the confirmed behavior or scope any cleanup. Done means the redundancy question has a clear, tested answer.

Written by the indexing model from the issue text.