Fix: Disable torch.autocast in RotaryEmbedding of Gemma and LLaMa for MPS device

[currybab]

What does this PR do?

Fixes #29431

The issue on MPS devices was caused by the merge of #29285 in version 4.38.2.

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?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[ArthurZucker]

That is indeed a problem. Was not aware that autocast is not available for mps.
We probably need to do a patch for this!
I think we can use cpu device even if the tensors are not on CPU no?

[ArthurZucker]

LGTM thank you for the prompt fix

[ArthurZucker]

FYI @fxmarty and @gante !

[ArthurZucker]

I have not tested this with compile but the dtype should be alright to check / we can always check self.dtype to not be input dependant

[Tran-Thanh1991]

Hi everybody. I'm having the same error, does anyone know how to fix it?

RuntimeError: User specified an unsupported autocast device_type 'cuda:0'

[ArthurZucker]

Do you have a reproducer?