It covers W4A4 NVFP4 activations with per-tensor or per-token scaling and W4A16 with BF16 activations.
The change replaces a standalone native finalizer with a shared CuTe kernel, adjusts dependent-launch hints, and wires the --no-fused-finalize benchmark flag through to CUTLASS. The stated scope is six CuTe DSL and benchmark-wiring files, with no C++ changes; public MoE defaults and fused-finalize arithmetic are unchanged.
The PR records a historical single-sweep benchmark on one NVIDIA B300 GPU: all 15 tested token counts improved for each of W4A4 per-tensor, W4A4 per-token, and W4A16. After a September 30 rebase, the author reports validation checks but says GPU benchmarks were not rerun. The source does not establish that the PR has merged, shipped in a FlashInfer release, or that its historical performance results apply to current upstream code.