[Perf][Qwen4Exp] Retune H200 M=4 merged QSA LL-GEMM plan (#60776)

Signed-off-by: zigzagcai <zigzagcai@users.noreply.github.com>
Co-authored-by: zigzagcai <zigzagcai@users.noreply.github.com>
Co-authored-by: Claude Code <noreply@anthropic.com>
This commit is contained in:
Zheng Cai
2026-10-09 07:18:50 +00:00
committed by GitHub
co-authored by zigzagcai Claude Code
parent 19c0bed584
commit 82e89e9944
@@ -79,7 +79,7 @@ QWEN4_EXP_SM90_GEMM_PLANS: dict[tuple[int, int], dict[int, SkinnyGemmConfig]] =
(4224, 2560): {
1: SkinnyGemmConfig(1, 128, 2, k_unroll=6, vector_width=2, static_k=2560),
2: SkinnyGemmConfig(2, 128, 2, k_unroll=3, vector_width=2, static_k=2560),
4: SkinnyGemmConfig(4, 64, 4, k_unroll=2, vector_width=4),
4: SkinnyGemmConfig(4, 64, 4, k_unroll=1, vector_width=8),
},
# GDN and QSA output projections, TP=4.
(2560, 1536): {