conda-python-scikit-learn-tests hanging in CUDA 11.8 Python 3.10

[csadorf]

The conda-python-scikit-learn-tests regularly hang for the CUDA 11.8, Python 3.10 configuration. The test suite completes most of its execution (approximately 98%) but then hangs indefinitely until the CI job times out.

The impact is limited to nightlies, because we do not run the CUDA 11.8 config for PRs.

Current Status

• A 1-hour timeout limit has been added (PR Add timeout to scikit-learn test suite #6591) to catch these issues earlier
• This is a recurring issue affecting multiple CI runs

Examples

• CI Run #14703638056
• CI Run #14689699930
• CI Run #14679054310
• CI Run #14725764519

[csadorf]

After analyzing the test results from three CI runs (41223147980, 41258408547, and 41199348400), I've found the following patterns that might help diagnose the issue:

Consistent Test Failures

The following tests failed consistently across multiple runs with similar progress percentages:

• test_ridge_vs_lstsq (appears in all three runs)
• test_transform_target_regressor_error (appears in all three runs)
• test_bagging_with_metadata_routing[model0] (30-32% in two runs)
• test_calibration_multiclass[0-True-sigmoid] (28-30% in two runs)
• test_search_with_estimators_issue_29157 (28-30% in two runs)

Key Observations

1. Early failures consistently occur between 19-45% completion
2. Some workers progress much further (90%+ completion)
3. Each run shows at least one worker with a significantly delayed timestamp:
   • Run 1: gw7 at 13:31 (vs 07:3X for others)
   • Run 2: gw12 at 14:32 (vs 08:3X for others)
   • Run 3: gw16 at 13:30 (vs 07:3X for others)
4. Tests fail at similar progress percentages despite running on different workers
5. The third run shows more workers reaching higher completion percentages (up to gw20)
6. XFAIL results appear consistently across runs

The pattern of delayed timestamps in each run (always one worker running 5-6 hours later than others) and the fact that some workers consistently progress much further than others strongly suggests a resource contention or deadlock issue. The consistent early failures of certain tests might be contributing to this by leaving resources in an inconsistent state.

Context

The issue might be exacerbated by the recent switch to parallel test execution introduced in PR #6553, which added pytest-xdist support for faster test runs. The resource contention patterns we're seeing could be related to multiple workers trying to access CUDA resources simultaneously.

[csadorf]

The most recent test failure appears to be an OOM issue has as the most notable error Reason=1:CUBLAS_STATUS_NOT_INITIALIZED: https://github.com/rapidsai/cuml/actions/runs/14725764519/job/41327957214

[csadorf]

The most recent CI run did not hang, but fail with a ton of RuntimeError: cuBLAS error errors.

[csadorf]

I suspect that the root-cause for the hangs is #6617 and that the cuBLAS and the cuSolver issues are unrelated.

[csadorf]

Mitigated via #6619 .