test: patch pytest-run-parallel worker context setup

[rwgk]

Alternative to #2228 (so we can decide which direction we prefer)

Summary

Add a cuda-python-local workaround for pytest-run-parallel's current fixture-sharing model so CUDA context setup can run inside each worker thread.

The shared helper patches pytest_run_parallel.plugin.wrap_function_parallel when pytest-run-parallel is installed, attaches optional per-worker context factories to selected pytest items, and keeps the behavior a no-op when pytest-run-parallel is absent. The bindings and core test conftests use that helper to set up CUDA context state in the worker thread before the test body runs.

Background

pytest-run-parallel executes the same pytest item in multiple worker threads, but fixture values are created by pytest before the worker calls and are shared across those threads. That is problematic for CUDA tests because the current CUDA context is thread-local. Until pytest-run-parallel has an upstream extension hook for per-worker setup, this PR monkeypatches the wrapper function as a localized workaround.

This is intended as an alternative to wrapping CUDA tests directly in each conftest. The monkeypatch centralizes the pytest-run-parallel-specific behavior and keeps package conftests focused on package-specific CUDA setup.

Changes

• Add cuda_python_test_helpers.pytest_run_parallel with:
  • idempotent installation of the pytest-run-parallel wrapper patch
  • signature validation for the upstream wrapper entry point
  • isolated kwargs copies for worker and iteration calls
  • per-worker context factory support attached to pytest item functions
  • no-op behavior when pytest-run-parallel is not importable
• Wire cuda_bindings/tests/conftest.py to run per-worker CUDA driver/context setup for CUDA-context-dependent bindings tests.
• Exclude cuDLA tests from the bindings worker-context selection because they intentionally override ctx as a no-op to avoid CUDA initialization.
• Wire cuda_core/tests/conftest.py to run per-worker init_cuda setup and replace direct CUDA fixture kwargs that need worker-local values.
• Add unit coverage using a fake pytest-run-parallel module so the monkeypatch mechanics can be tested without requiring pytest-run-parallel or CUDA hardware.

Notes

This workaround should be removable once pytest-run-parallel exposes an upstream hook that runs inside each worker thread before the test function is invoked.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[rwgk]

/ok to test

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-2282/
https://nvidia.github.io/cuda-python/pr-preview/pr-2282/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-2282/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-2282/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.