All pull requests

Pull requests

• Speed up the fused MoE kernel for models with many small experts: cheaper scheduler fences, local combine fused into GEMM1, a faster quantizing dispatch, and an in-kernel task tracer

  #31·KernelFlow-ops opened Oct 6, 2026·

  ·
  6

• Add MXFP4/NVFP4 expert paths

  #30·KernelFlow-ops opened Oct 6, 2026·

  ·
  1

• Add MXFP8 expert path (SM120)

  #29·KernelFlow-ops opened Oct 6, 2026·

  ·
  7

• Support SM120/SM121 (RTX 50)

  #28·KernelFlow-ops opened Oct 6, 2026·

  ·
  4

• Optimize Hopper dispatch with warp-split copies

  #27·syedshazli opened Sep 24, 2026·

  ·
  1

• Python cutedsl nvshmem port

  #26·benhg opened Jun 25, 2026·

  ·
  11

• add mathdx version

  #25·syedshazli opened May 12, 2026·

  ·
  3