update test_can_load_with_global_device_set with a hack

[ydshieh]

What does this PR do?

torch.set_default_device has some unexpected impacts on other tests even if we try to restore the origin value by calling it at the end of the test. See #37551 for one example that is fixed there.

This PR uses torch._GLOBAL_DEVICE_CONTEXT to do something more close to the internal mechanism in torch.set_default_device to perform the cleanup, see

https://github.com/pytorch/pytorch/blob/e229ce34c4ab8cd4e2800227615be32fb362b1e6/torch/__init__.py#L1205-L1218

Running

RUN_SLOW=1 python3 -m pytest -v tests/models/mask2former/test_modeling_mask2former.py

will pass on this PR while it fails on the base commit (b33edf1) of this PR.

FAILED tests/models/mask2former/test_modeling_mask2former.py::Mask2FormerModelTest::test_torch_export - AttributeError: module 'torch._tensor' has no attribute 'split'

FAILED tests/models/mask2former/test_modeling_mask2former.py::Mask2FormerModelIntegrationTest::test_export - AttributeError: module 'torch._tensor' has no attribute 'split'

[github-actions]

Hi 👋, thank you for opening this pull request! The pull request is converted to draft by default. The CI will be paused while the PR is in draft mode. When it is ready for review, please click the Ready for review button (at the bottom of the PR page). This will assign reviewers and trigger CI.

[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.

[Cyrilvallez]

This is kind of convoluted (not your fault haha)... Don't you think that simply running in a subprocess would be better?

[ydshieh]

yeah, i didn't see it is decorated with require_torch_accelerator and I tried to avoid overhead added to CircleCI. But now I just use subprocess.

[Cyrilvallez]

Perfect, LGTM! Much simpler and foolproof! Let's just add a small comment about why we're doing it!