Skip to content

[blackwell] sm100: MXFP4 grouped/expert dispatch -> Triton fused_mxfp4_gemm #227

Description

@Andrewxu313

Parent: #204 — Blackwell Phase 2 kernel port
Branch: tairan/blackwell-02-04-mxfp4-triton-dispatch
Spec: blackwell-kernel-port-v1.md § Sub-task 4

What

Wire arch == 'sm100' branch in fused_wgmma_grouped.py and fused_wgmma_expert.py to call the existing pure-Triton fused_mxfp4_gemm.py (no WGMMA intrinsics, JIT-compiles for sm100). Tune BLOCK_M/N for B200 128KB SMEM.

Acceptance

  • MXFP4 grouped and expert forward: max_abs_err < 5e-3 vs float32 reference
  • No _C_expert_mxfp4_wgmma / _C_grouped_mxfp4_wgmma import at runtime on sm100

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions