Summary
test_unittests_v2[unittest/_torch/executor] fails intermittently in pre-merge single-GPU stages with AssertionError: failure reported in unittests while all ~953 inner executor tests report PASSED in the merged results.
Mechanics (from the wrapper stack): the parallel phase exits non-zero after tests complete, the --lf retry re-runs and passes everything, but the session still exits non-zero — i.e. a crash/failure in session teardown (thread-leak check or native crash at shutdown), not a test failure.
Occurrences (all on unrelated PR #15730, current-main bases)
- L0_MergeRequest_PR 47040 — B300-PyTorch-2
- L0_MergeRequest_PR 47191 — DGX_H100-PyTorch-3
- L0_MergeRequest_PR 47357 — B300-PyTorch-2
- L0_MergeRequest_PR 47366 — B300-PyTorch-2
That is 4 of ~10 runs — the single most frequent pre-merge blocker we hit while babysitting that PR. The NCCL shutdown fix (#15892) did not eliminate it.
Ask
Deflake the executor unittest group teardown (e.g. identify the leaking/crashing shutdown path, or apply the pytest.mark.threadleak(enabled=False) pattern used elsewhere if it is the threadleak gate).
Summary
test_unittests_v2[unittest/_torch/executor]fails intermittently in pre-merge single-GPU stages withAssertionError: failure reported in unittestswhile all ~953 inner executor tests report PASSED in the merged results.Mechanics (from the wrapper stack): the parallel phase exits non-zero after tests complete, the
--lfretry re-runs and passes everything, but the session still exits non-zero — i.e. a crash/failure in session teardown (thread-leak check or native crash at shutdown), not a test failure.Occurrences (all on unrelated PR #15730, current-main bases)
That is 4 of ~10 runs — the single most frequent pre-merge blocker we hit while babysitting that PR. The NCCL shutdown fix (#15892) did not eliminate it.
Ask
Deflake the executor unittest group teardown (e.g. identify the leaking/crashing shutdown path, or apply the
pytest.mark.threadleak(enabled=False)pattern used elsewhere if it is the threadleak gate).