test: skip system NVML tests when the symbol is absent in older bindings

[rluo8]

Description

This is to fix bug 6179497, and it's a test only issue.
Tests failed when using cuda.core v1.1.0 and cuda bindings 12.9.7.

FAILED system/test_system_device.py::test_addressing_mode
FunctionNotFoundError: function nvmlDeviceGetAddressingMode is not found
FAILED system/test_system_device.py::test_repair_status
FunctionNotFoundError: function nvmlDeviceGetRepairStatus is not found

This issue was originally fixed in pr #1898, but the fix in cuda.bindings wasn't inculded in cuda 12.9.x.
This fix back-stops the skip on the cuda.core test side (independent of the cuda.bindings version): a missing NVML symbol is skipped rather than failed.

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-2390/
https://nvidia.github.io/cuda-python/pr-preview/pr-2390/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-2390/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-2390/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[mdboom]

unsupported_before is meant to indicate that a feature isn't available on a specific hardware architecture. It's not meant to indicate that a function isn't available on a specific CTK version.

We should skip each of the failing tests individually, based on the CTK version.

For example:

@pytest.mark.skipif(get_user_mode_driver_version() < (13, 0))

[leofang]

I think we should check the KMD version not UMD version?

[rluo8]

Thanks @mdboom , @leofang. I updated the patch to skip the two cases when kernel mode driver < R580. Verified on a forward compatibility system (UMD 13.0 + KMD 575), the tests would still fail with FunctionNotFoundError if using UMD check. After switching to use KMD version check, the two tests could be skipped correctly.

[mdboom]

Thanks, @rluo8. LGTM.

[rluo8]

Hi @mdboom , the CI check have all passed now. It seems that there's something wrong with Github and it disabled the auto-merge. Could you please help merge this PR?
Thanks!

[rluo8]

Closing as no longer needed.
PR #1898 already made unsupported_before skip on a missing NVML symbol, but the helper shipped inside the cuda.bindings wheel and not included in version 12.9.7.
After #2384 moved arch_check.py into the standalone cuda_python_test_helpers package, which travels with the test suite rather than with the installed cuda.bindings. Both test_addressing_mode and test_repair_status now skip on their own.