ciflow/torchtitan/194364: Inductor: fix NVGEMM EFC reduction fallback
- PyTorch: 1011 events in the last 90 days
- PyTorch: 1000th Release in the last 90 days
- Previous: earlier the same day · viable/strict/1789514701: [Inductor] Eliminate no-op self slice_scatter after split (#195608)
What happened
Cutlass Operators 0.2 moved the shared EFC module, so import it independently from the optional dense EFC class. Preserve the identity finalizer needed by generated reductions. Direct EFC supports one generated reduction. Defer only selected composite reduction plans, which otherwise produce invalid results, and validate the regular NVGEMM fallback. Update tests to assert supported behavior rather than unsupported o…
Summary assembled by rule from the sources below