Bug: _select_device_and_dtype does not work on virtualized hardware

[ibevers]

Description

I ran into this issue because MacOS tests were passing locally, but not on GitHub Actions. It is because _select_device_and_dtype selected MPS on the MacOS GitHub Actions runner, but the runner virtualizes MacOS, so no MPS device is available.

See details here:
#215

Steps to Reproduce

Run code with DeviceType.MPS on a GitHub Actions MacOS runner (ARM).

Expected Results

The tests will segfault.

Actual Results

The tests segfaulted.

Additional Notes

No response

[fabiocat93]

Isn't this more of a GitHub limitation? https://docs.github.com/en/actions/using-github-hosted-runners/using-github-hosted-runners/about-github-hosted-runners#limitations-for-arm64-macos-runners

[ibevers]

@fabiocat93 This could come up in other contexts though too. Perhaps we could add a try/except that tries to run something on MPS in the function? We could do the same thing for GPU.

if torch.backends.mps.is_available():
    try:
        torch.empty(0, device="mps")
        available_devices.append(DeviceType.MPS)
    except Exception as e:
        print(f"MPS is available but encountered an error: {e}")

[fabiocat93]

@fabiocat93 This could come up in other contexts though too. Perhaps we could add a try/except that tries to run something on MPS in the function? We could do the same thing for GPU.

if torch.backends.mps.is_available():
    try:
        torch.empty(0, device="mps")
        available_devices.append(DeviceType.MPS)
    except Exception as e:
        print(f"MPS is available but encountered an error: {e}")

Have you tested this on the GitHub runner? Does it solve the issue you describe above? If yes, it looks a good solution to me. I agree with you that we should apply the same approach to GPU. Also, I would recommend adding a comment explaining the rationale behind these edits

[ibevers]

@fabiocat93 I haven't tested it yet--I just wanted to see what you thought of that approach in concept. I will make proper PR and test. Thanks for your feedback!