pytorch - 💡(How to fix) Fix DISABLED DistillGPT2 (dynamic_cpu_max_autotune_inductor_amp_freezing_huggingface) [1 comments, 2 participants]

TL;DR

The benchmark regression in CI after PR #181360 may be resolved by investigating and addressing the graph break count increase in DistillGPT2.

Guidance

• Review the changes introduced in PR #181360 to identify potential causes of the graph break count regression.
• Compare the current graph break count with the expected value in the CSV file benchmarks/dynamo/ci_expected_accuracy/dynamic_cpu_max_autotune_inductor_amp_freezing_huggingface_inference.csv to understand the extent of the regression.
• Investigate the unique break type introduced in the regression to determine its impact on the benchmark.
• Collaborate with the listed maintainers and contributors to discuss potential fixes or workarounds for the regression.

Notes

The exact cause of the regression is unclear, and further investigation is needed to determine the root cause.

Recommendation

Apply workaround: Investigate and address the graph break count increase in DistillGPT2, as it is the most likely cause of the benchmark regression.