[tests] make test_sdpa_can_compile_dynamic device-agnostic

[faaany]

What does this PR do?

test_sdpa_can_compile_dynamic should not only work on GPU, it also works on XPU, below is an example:

PASSED tests/models/hubert/test_modeling_hubert.py::HubertModelTest::test_sdpa_can_compile_dynamic

@amyeroberts and @ydshieh

[amyeroberts]

Thanks for enabling!

Similar to the other PR - we'll want to move this out of drafts and then run the slow tests before merge. Otherwise LGTM :)

[faaany]

Hi @amyeroberts , thanks for the review, undrafted!

[faaany]

can we merge this PR as well? Thx a lot!

[amyeroberts]

Hi @faaany, apologies for the delay. I can see that adding the [run_all] tag didn't run the sdpa tests -- @ydshieh how can we trigger these?

[ydshieh]

Hi @amyeroberts This test is only for A10 or A100.

            if not torch.version.cuda or major < 8:
                self.skipTest(reason="This test requires an NVIDIA GPU with compute capability >= 8.0")

And so far, .github/workflows/self-pr-slow-ci.yml is only written to run on a T4.

The quick but manual way is to launch a A10 SSH runner

https://github.com/huggingface/transformers/actions/workflows/ssh-runner.yml

and run the tests manually.

[faaany]

Hi @amyeroberts , can we also proceed with merging of this PR?

[amyeroberts]

Hi @faaany, yes, apologies for the delay. As this isn't changing any core logic it shouldn't really affect the slow tests so I think we can merge anyway