Move CUDA device probe in test_symbol_exposure.py from import time to test execution

[coderabbitai]

Summary

Module-level Device(0) initialization in numbast/src/numbast/experimental/mlir/tools/tests/test_symbol_exposure.py (lines 8–12) runs at pytest collection time, which blocks collection on machines without a CUDA runtime and prevents graceful skip behavior.

Details

# Current problematic code (lines 8-12)
from cuda.core.experimental import Device

dev = Device(0)
cc = dev.compute_capability

The codebase already uses fixture-based device queries elsewhere (e.g., conftest.py:arch_str() and test_shim_writer.py:cc()). The probe should be moved inside the test function (or a fixture), with a pytest.skip fallback if CUDA is unavailable.

Suggested fix

Move Device(0) into the test body and wrap with try/except:

def test_symbol_exposure(run_in_isolated_folder, arch_str):
    try:
        cc = Device(0).compute_capability
    except Exception as e:
        pytest.skip(f"CUDA device unavailable for symbol-exposure test: {e}")
    ...

Or alternatively, reuse the existing arch_str fixture (defined in conftest.py) which already derives the compute capability at test execution time.

References

• Raised in PR: Migrate experimental MLIR backend #346
• Review comment: Migrate experimental MLIR backend #346 (comment)
• Requested by: @isVoid