Fix autocast incompatibility in RecurrentGemma

[xplip]

What does this PR do?

Fixes #30830

In modeling_recurrent_gemma.py, cast the key_states and value_states in RecurrentGemmaSdpaAttention._update_cache to the dtype of the cache tensors they are written to. This way, a RuntimeError is prevented when autocasting to bfloat16/float16.

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.

[amyeroberts]

cc @ArthurZucker

[ArthurZucker]

Hey, could you share a reproducer? 🤗

[xplip]

Hey, could you share a reproducer? 🤗

Sure, I added one in the issue #30830 :)

[ArthurZucker]

Thanks for the detailed issue 😉
What's weird for me is that

        with torch.autocast(device_type=device_type, enabled=False):
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos()
            sin = emb.sin()
        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)

is the only place we mess up with autocast. I think casting earlier should help memory, but this is overall the safest way .

Could you run the slow tests?