trunk/d25057513dc155ee32bfdf4e760da11652269423: [inductor][NVGEMM] Fuse NVFP4 output scaling (#194651)
- PyTorch: 1835 events in the last 90 days
- PyTorch: 1811th Release in the last 90 days
- Previous: earlier the same day · trunk/bfd8c2497aecccb5e32f08446ac644b87ee2e4d2: [inductor][NVGEMM] Improve NVFP4 autotuning (#194652)
What happened
Human commentary: ModelOpt NVFP4 linear layers multiply _scaled_mm output by a scalar dequantization factor. Folding that factor before QKV fan-out removes one standalone scale launch per layer and lets autotuning consider the GEMM implementation that applies it directly. AI-assisted content: Problem Fixes #198173 ModelOpt NVFP4 linears materialize a separate pointwise multiply after _scaled_mm . That launch is espe…
Summary assembled by rule from the sources below