pytorch - 💡(How to fix) Fix DISABLED vision_maskrcnn graph breaks (dynamic_cpu_max_autotune_inductor_amp_freezing_torchbench) [1 comments, 2 participants]

TL;DR

The benchmark regression in graph break count may be resolved by investigating and addressing the cause of the additional graph break in the vision_maskrcnn model.

Guidance

• Review the changes introduced in PR #181360 to identify potential causes of the graph break count regression.
• Compare the current graph break count of 31 with the expected count of 30 to understand the nature of the regression.
• Analyze the CSV file at benchmarks/dynamo/ci_expected_accuracy/dynamic_cpu_max_autotune_inductor_amp_freezing_torchbench_inference.csv for more detailed information on the benchmark results.
• Investigate the 5 unique break types that remain unchanged to determine if they are related to 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 cause of the additional graph break in the vision_maskrcnn model, as it is likely the primary cause of the benchmark regression.