Skip to content

[blackwell] sm100: Fused INT4 grouped + SiLU Triton kernel (K2.5 MoE FFN) #231

Description

@Andrewxu313

Parent: #204 — Blackwell Phase 2 kernel port
Branch: tairan/blackwell-02-08-int4-fused-silu
Depends on: sub-task 7
Spec: blackwell-kernel-port-v1.md § Sub-task 8

What

Write batchgen_kernels/triton/fused_int4_grouped_silu.py:
V1 approach (split + fused silu_mul):
gate = int4_grouped_gemm(x, Wg)
up = int4_grouped_gemm(x, Wu)
out = silu(gate) * up (via small Triton silu_mul kernel)

Wire into batchgen/moe/fused_int4_wgmma_grouped.py for arch == 'sm100'.

Acceptance

  • max_abs_err < 1.0 vs float32 reference
  • K2.5 MoE FFN forward completes on B200 without error

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