Skip to content

[blackwell] sm100: recompile 9 generic CUDA kernels #225

Description

@Andrewxu313

Parent: #204 — Blackwell Phase 2 kernel port
Spec: blackwell-kernel-port-v1.md § Sub-task 2
Branch: tairan/blackwell-02-02-recompile-generic
Depends on: sub-task 1

What

Populate _sm100_extensions with 9 kernels that use only generic CUDA or SM80 mma.sync (safe to recompile for -arch=sm_100a without source changes):
marlin_grouped_gemm, fp8_blockwise_ops, marlin_transform, routing generics (4×, excluding fused_gate.cu which uses WGMMA), dispatch_scatter_3d, fused_kv_norm_rope, fused_q_absorb, fused_q_split.

Acceptance

  • All 9 compile with ptxas reporting sm_100a code
  • Import smoke tests pass

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

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