trunk/7b450b9f6b869896ba6bce8e6ab0ba5fbc85c0f9: flex_gemm: support variable-length-K grouped GEMMs (#199808)
- PyTorch: 2301 events in the last 90 days
- PyTorch: 2271th Release in the last 90 days
- Previous: earlier the same day · trunk/10dc63912abee5b1c6b942e670872e515cbf9210: [CPP FakeTensor] Don't look up the op in register_fake (#200798)
What happened
Human Note Add varlen k support for quack and grouped_gemm Agent note Expose QuACK's existing varlen-K mainloop through FlexGEMM for BF16/FP16 F.grouped_mm weight gradients: [M,T] @ [T,N] with expert end offsets produces [E,M,N] . Carry the K prefix through lowering, generated calls, EpiMod caches and compilation, and legal configuration selection. Autotuning uses boundaries over T rather than output batch E. No new…
Summary assembled by rule from the sources below