Allow top-k and top-p to coexist

[jkrukowski]

This PR is based on the comment #23 (comment)

@pcuenca please let me know if that's what you had in mind.

This PR:

• adds LogitsWrapper based on https://github.com/huggingface/transformers/blob/42017d82baa083da2bee3055fdac80c81ee97b8a/src/transformers/generation/logits_process.py#L62
• adds LogitsProcessor based on https://github.com/huggingface/transformers/blob/42017d82baa083da2bee3055fdac80c81ee97b8a/src/transformers/generation/logits_process.py#L73
• changes generation to use LogitsProcessor
• adds tests

[pcuenca]

Great job, @jkrukowski, sorry for taking long to review 🙏

I think the PR is in very good shape and can be merged quickly. I like that it follows the transformers pattern and is super clean and easy to follow. I only have two comments:

• We should always use the nomenclature warper in all file, class, and variable names.
• I think it's a bit more expressive if the LogitsWarper protocol has a method called warp, and a default callAsFunction method that simply calls warp. We'd use the implicit call most of the time for symmetry with transformers, but users may also choose to call warp if they prefer, and the semantics would be clear.

Thanks a lot for another great contribution!

[pcuenca]

Thanks a lot for adding tests too! 🙌

[jkrukowski]

@pcuenca applied the requested changes, please take a 2nd look

[pcuenca]

Thanks a lot!