[Gemma4] Investigate flaky test_generation_beyond_sliding_window_1_eager (#48236)
* [Gemma4] Isolate test_generation_beyond_sliding_window in integration test class Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [Gemma4] Keep only eager sliding window test for investigation Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [Gemma4] Log full output_text before assertion for CI visibility Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [Gemma4] Run sliding window test 10x in fresh subprocesses to expose non-determinism Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [Gemma4] Log stdout/stderr from each subprocess run Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [Gemma4] Remove image processing and processing tests; tweak modeling test Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * try * try * [Gemma4] Log free -g before subprocess loop Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [Gemma4] Add commented-out torch.use_deterministic_algorithms for reference Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [Gemma4] Use torch.use_deterministic_algorithms to fix flaky sliding window test Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix * [Gemma4] Use require_deterministic_for_accelerator decorator; add it to testing_utils Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [Gemma4] Add require_deterministic_for_accelerator; apply to sliding window test Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Revert utils/pr_slow_ci_models.py to main Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [Gemma4] Fix decorator order: @parameterized.expand must be outermost Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [Gemma4] Update eager expected value after torch.use_deterministic_algorithms Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [Gemma4] Skip test_sdpa_can_dispatch_on_flash; add TODO for modeling fix Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [gemma4] Update stale expected value for test_model_text_only_multigpu on sm86 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [gemma4] Clarify comment: block-overlay mask forces materialization by design Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>
Y
Yih-Dar committed
50adfa191b91d2a9e6c21ae302a57c291517cd75
Parent: 2b4e2b3
Committed by GitHub <noreply@github.com>
on 8/25/2026, 2:59:08 PM