pytorch - 💡(How to fix) Fix DISABLED detectron2_fasterrcnn graph breaks (dynamic_cpu_max_autotune_inductor_amp_freezing_torchbench) [1 comments, 2 participants]

TL;DR

Investigate and address the graph break count regression in the Detectron2 models to resolve the benchmark failures.

Guidance

• Review the changes introduced in PR #181360 to identify the potential cause of the graph break count regression.
• Compare the expected and actual graph break counts for each model to understand the nature of the regression.
• Check the CSV file benchmarks/dynamo/ci_expected_accuracy/dynamic_cpu_max_autotune_inductor_amp_freezing_torchbench_inference.csv for more detailed information on the benchmark results.
• Consider re-running the benchmarks with additional logging or debugging to gather more insights into the graph break count regression.

Notes

The issue seems to be related to a specific change introduced in PR #181360, and resolving the graph break count regression is likely to fix the benchmark failures.

Recommendation

Apply workaround: Revert or modify the changes introduced in PR #181360 to mitigate the graph break count regression, and re-run the benchmarks to verify the fix.