Mark CV-CUDA tests with needs_cuda

[zy1git]

Mark CV-CUDA tests with needs_cvcuda.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/9305

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit b195ef2 with merge base 96e7797 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[NicolasHug]

We'll need to use these two marks on every single CV-CUDA test, so instead of duplicating this everywhere, let's simply define a global mark on top of the file as

CV_CUDA_TEST = (
                    pytest.mark.skipif(not CVCUDA_AVAILABLE, reason="CVCUDA is not available"),
                    pytest.mark.needs_cuda,
 )

and then use marks=CV_CUDA_TEST for all of these.

Let's also update the previous test classes from @AntoineSimoulin:

vision/test/test_transforms_v2.py

Lines 6859 to 6861 in 96e7797

	@pytest.mark.skipif(not CVCUDA_AVAILABLE, reason="test requires CVCUDA")
	@needs_cuda
	class TestCVDUDAToTensor:

and

vision/test/test_transforms_v2.py

Lines 6797 to 6799 in 96e7797

	@pytest.mark.skipif(not CVCUDA_AVAILABLE, reason="test requires CVCUDA")
	@needs_cuda
	class TestToCVCUDATensor:

to use this mark as well.

I suspect we might be able to also remove the CVCUDA_AVAILABLE global variable now (not sure).

[AntoineSimoulin]

Maybe we can create a @needs_cvcuda decorator (similar to @needs_cuda)? We will also have to modify the pytest_configure function to add this decorator. This way we can identify easily tests requiring cvcuda?

[justincdavis]

This significantly reduces the boilerplate!

I have gone ahead and updated the branch implementing CV-CUDA for the ToDtype transform and I will preemptively update the others as well.

As a note, I defined CV_CUDA_TEST as a list so we can use it with the pytestmark class attribute for TestCVCUDAToTensor and TestToCVCUDATensor. Example here

[NicolasHug]

Drive-by suggestion:

Suggested change

	class TestCVDUDAToTensor:
	class TestCVCUDAToTensor:

[NicolasHug]

Some tests in TestToCVCUDATensor are failing because cvcuda is now removed.

It's good that it's removed, but we should update the parts in TestToCVCUDATensor to now use _import_cvcuda() instead

[NicolasHug]

Having the new needs_cvcuda is great, but I realize now we're in this odd situation where we really have to make all CV-CUDA tests also have the needs_cuda mark, otherwise we risk skipping them when we don't want to. That's reflected here by having to add both pytest.mark.needs_cvcuda and pytest.mark.needs_cuda, but also below when parametrizing an entire class, we have to add both marks:

@needs_cuda
@needs_cvcuda
class TestCVDUDAToTensor:

I'd suggest instead to only use the needs_cvcuda mark for the cv-cuda test, and leave out the needs_cuda mark. That is:

            pytest.param(
               # ...,
                marks=pytest.mark.needs_cuda,
            )


# and:

# @needs_cuda  # <-- removed!
@needs_cvcuda
class TestCVDUDAToTensor:

In order to still mark the cv-cuda tests with needs_cuda, all we need to do is to add this logic in conftest.py:

needs_cvcuda = item.get_closest_marker("needs_cvcuda") is not None
if needs_cvcuda:
    item.add_marker(pytest.mark.needs_cuda)
needs_cuda = item.get_closest_marker("needs_cuda") is not None

Let me know if I can clarify anything!

[NicolasHug]

That looks great. Thanks @zy1git !

Importantly we should make sure none of the CVCUDA tests are being skipped anymore. To check, we have to look at the logs https://github.com/pytorch/vision/actions/runs/20149139270/job/57837488003 and it's sometimes easier to check the "raw logs"

I did a manual check, things look good, but I encourage you to check on your own before merging! :)