viable/strict/1788771294: [dynamo] Preserve is_inference in after-AOT repros (#193565)
- PyTorch: 621 events in the last 90 days
- PyTorch: 615th Release in the last 90 days
- Previous: earlier the same day · trunk/8aa0ae4b6773be9b0ca06dc4f7dcd254cefa56ad: [torchcomms hash update] update the pinned torchcomms hash (#196181)
What happened
After-AOT repro generation receives is_inference through the compiler kwargs but drops it while dumping and replaying the graph. When an inference compilation fails with repro_after="aot", the generated repro therefore recompiles the graph with compile_fx_inner's default is_inference=False, potentially changing its behavior and preventing the original failure from reproducing. Propagate is_inference through graph du…
Summary assembled by rule from the sources below