← Back to events
ActiveAIRelease

trunk/0e7f619f994b5a7675e48d07b0da1c11424c6e99: Integrate scaled addmm with torch.compile (#195410)

What happened

Human Note Keep compiler support in a focused top-of-stack change so Dynamo and Inductor behavior can be reviewed independently from the CUDA implementation. Agent note This allows the functional and mutating scaled-addmm bindings in Dynamo graphs, lowers both through Inductor's external-kernel fallback, and teaches reinplacing to turn the functional operator back into _scaled_addmm_ when its accumulator can be reus…

Summary assembled by rule from the sources below

Why it's spreading

Sources