clean up device initialization in test

[ksimpson-work]

I suspect the issue was related to the fact that we had a dangling device initialization. In any case we should test against our actual guarantee which is that Device() is current device when device_ordinal is unsupplied

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[ksimpson-work]

/ok to test

[leofang]

I think we miss an import.

[rwgk]

I only have a minute right now. Just quick, I saw this in one of the failing tests:

______________________________ test_device_alloc _______________________________

deinit_cuda = None

    def test_device_alloc(deinit_cuda):
        device = Device()
        device.set_current()
        buffer = device.allocate(1024)
        device.sync()
        assert buffer.handle != 0
        assert buffer.size == 1024
>       assert buffer.device_id == cuda.cudaGetDevice()

tests/test_device.py:53: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

name = 'cudaGetDevice'

    def __getattr__(name):
        if name == "__version__":
            import warnings
            warnings.warn("accessing cuda.__version__ is deprecated, "
                          "please switch to use cuda.bindings.__version__ instead",
                          DeprecationWarning, stacklevel=2)
            from . import bindings
            return bindings.__version__
    
>       raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
E       AttributeError: module 'cuda' has no attribute 'cudaGetDevice'

/opt/hostedtoolcache/Python/3.9.21/x64/lib/python3.9/site-packages/cuda/__init__.py:10: AttributeError

I see this elsewhere:

handle_return(runtime.cudaGetDevice())

We have the runtime and handle_return imports already in test_device.py

Maybe @leofang saw something else?

[ksimpson-work]

@rwgk You're totally right, I didn't test locally before running the CI since it was so small but I missed some obvious issues

[ksimpson-work]

/ok to test

[rwgk]

Looking at this with a fresh eye, I'm wondering:

• We're in cuda.core but we're reaching back into a cuda.bindings API (runtime). Do we have to?
• Could/should we do this instead? (does it matter?)

    assert buffer.device_id == int(device)

[ksimpson-work]

Yeah that also works. I'll add it, and also a new test for the device id against our guarantee

[ksimpson-work]

/ok to test

[rwgk]

Reapproving, assuming tests pass.

The new test looks great.

[github-actions]

Doc Preview CI
Preview removed because the pull request was closed or merged.