can padding_side be 'right'?

[ry-sun]

🐛 Describe the bug

First off, I'd like to say that this is an excellent repo.

However, I did notice that the default tokenizer config sets padding_side=left, which seems to be different from the common default of right in most NLP tokenizers. While I understand that setting padding_side to right might be more convenient when calculating advantages, I think it could potentially lead to bugs if users are allowed to modify it. Would it be safer to prevent users from modifying the padding side?

Which trlX version are you using?

newest

Additional system and package information

No response

[Chen9154]

@rockmagma02 For some frequently-used gpt-like LLMs such as BLOOM, the padding_side is set to left by default. Meanwhile, you could easily change the padding_side to right after initialization.

rw_tokenizer = AutoTokenizer.from_pretrained("bigscience/bloom-1b1")
rw_tokenizer.padding_side = 'right'

[ItGirls]

for glm, the padding side can only be left, so that the reward model is not for glm