[yeliu-oss]

Description

Human Reviewers

The positive grouped-MLP dprob determinism test can reach a cuDNN Frontend wrapper that accepts deterministic=True but explicitly refuses its configured kernel. On Rubin, its first backward raises the known NotImplementedError before the test reaches its byte comparisons.

Recognize that exact refusal around the first run and report the configuration as skipped. A successful first run continues through the existing fusion assertions and repeated byte comparisons. Other exception messages, exception types, and failures during later runs still fail the test.

flowchart LR
  subgraph before["Before"]
    B1["First deterministic run"] -->|"Known FE refusal"| B2["Test fails"]
    B1 -->|"Success"| B3["Repeated byte comparisons"]
  end
  subgraph after["After"]
    A1["Same first run"] -->|"Exact known refusal"| A2["Skip with FE reason"]
    A1 -->|"Success"| A3["Same repeated byte comparisons"]
  end

Loading

This changes test classification for an unsupported configuration. The full SM107 grouped-MLP replay preserves every passing case and changes only the known GLU failure to a reasoned skip. Full L0 coverage remains outstanding.

Agent Reviewers

• Target: main, base eb3c860a4d3394b7eb5e21dba3021fa10782a52e; head 04a275acb1084a476cee6d8cffcb06487ac5aeaa.
• Scope: tests/pytorch/test_grouped_mlp.py, TestGroupedMLPDeterminism.test_dprob_is_bit_exact_across_runs.
• Match only NotImplementedError with the complete reason deterministic dprob is implemented only for the SM100 kernel without dbias. The handler surrounds the initial _run() assignment only. After unwrapping that handler, the complete file's AST matches the base.
• FE retains ownership of per-configuration support, consistent with [PyTorch] Allow deterministic dprob for the GLU fused grouped MLP #3635 and the adjacent runs-or-refuses test. This avoids adding architecture policy to TE's cached wrapper-signature query.
• Full grouped-MLP SM107 replay: 1,256 passed, 4,113 skipped, exit 0. Baseline: 1,256 passed, 4,112 skipped, one failure. All 5,369 unique JUnit identities match; only the GLU positive case changes from failed to skipped with the exact FE reason. SReLU bytewise determinism and both runs-or-refuses cases stay passed. The new skip is a waiver of an unsupported configuration.
• The candidate Python test ran against installed TE 33f3b3561943fcd9dd2f544f67dc2dfd62efdc9e production binaries. The affected common/PyTorch production sources match the target base. FE 1.32.0, candidate source identity, native-library identities and process exit were checked in the test environment. This Python-only change required no native rebuild.
• Candidate repeats were pinned to four. The baseline value is inferred from the unchanged default and configuration; its known refusal occurs on the first call.
• Eleven actual-test-tail CPU probes and eight independent FE-guard probes cover successful execution, exact refusal, unrelated errors, later refusal, nonfinite output and byte mismatch. These probes use CPU stand-ins. Independent code and runtime evidence reviews passed.
• Pinned pre-commit hooks, the PyTorch Python lint launcher and the license check passed. Full L0_* and PR CI coverage remain outstanding.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

• Skip the first-run, exact known FE determinism refusal in the positive dprob test.
• Preserve supported-path numerical checks and unrelated failure propagation.

Checklist

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

The affected full grouped-MLP suite passed as recorded above; the complete L0/PR matrix remains outstanding. This test-only correction does not change the documented product API. The PR remains draft pending human author review.