Make EosTokenCriteria compatible with mps

[pcuenca]

What does this PR do?

EOS termination was moved to stopped criteria in #29459. It uses torch.isin(), which is not compatible with the mps device: pytorch/pytorch#77764 (comment). This PR uses the old method to detect whether any of the EOS tokens were found in the generated sequences.

Fixes #29459 (comment)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[zucchini-nlp]

Thanks for fixing! Maybe we can leave the old method only and get rid of extra if conditions. Both methods are compile compatible anyway :)

[pcuenca]

Yes, good point, I thought about that too. I opted to keep both because the new method is much more readable and intuitive, so it doesn't really add too much noise in my opinion. It also documents why we are using the long approach, and provides a hint (via the link to the github issue) to remove that case when we have full support of isin() for mps at some point in the future. But sure, we can keep just one if that's preferred :)

[gante]

Thank you for the fix 🙏

[gante]

I'm pro the if/else: it serves as a reminder to revisit the code in the future, when mps supports this more readable code path

[amyeroberts]

I'm not sure we should be doing this - the lack of support is really a torch issue, and not a transformers one. There's many operations which are (unfortunately) unsupported by the mps backend, and I wouldn't want to see a tonne of if/else statements across the library to support this. At the moment, PYTORCH_ENABLE_MPS_FALLBACK=1 is the standard approach to try and use mps, which I believe resolves this issue

[ArthurZucker]

I don't really mind this as it should be mostly for Llama3, but only if that produces the correct outputs. If not, then we should just recommend to use mlx?
@pcuenca does this produce gibberish or same as on cuda?

[amyeroberts]

As discussed internally in slack - https://huggingface.slack.com/archives/C06SW4886G6/p1713567939087889 - happy to merge as it's currently impacting llama 3, but should have an issue created tracking this to make sure it's removed in the future

[MilesCranmer]

Thanks so much for the quick fix. Could this get a release?

[ArthurZucker]

Yes, will do this asap, just waiting to make sure we don't have other failures.