[Perf][Qwen4Exp] Add SM121 TP=2 merged QSA projection plans (#61437)

Signed-off-by: TimRen01 <HolaTim@hotmail.com>
Co-authored-by: Thien Tran <gau.nernst@yahoo.com.sg>
This commit is contained in:
Tim002
2026-10-13 22:58:44 +00:00
committed by GitHub
co-authored by Thien Tran
parent 78b32b3fdc
commit 96fc2da315
@@ -240,12 +240,20 @@ QWEN4_EXP_SM121_GEMM_PLANS: dict[tuple[int, int], dict[int, SkinnyGemmConfig]] =
8: SkinnyGemmConfig(8, 128, 1, vector_width=2, static_k=6144),
16: SkinnyGemmConfig(16, 32, 1, static_k=6144),
},
# QSA fused QKV/gate projection, TP=2.
# QSA fused QKV/gate projection, TP=2 (LoRA keeps the indexer projection separate).
(6656, 2560): {
1: SkinnyGemmConfig(1, 128, 4, k_unroll=2, vector_width=4, static_k=2560),
2: SkinnyGemmConfig(2, 128, 4, k_unroll=2, vector_width=4, static_k=2560),
4: SkinnyGemmConfig(4, 64, 2, k_unroll=4, vector_width=4, static_k=2560),
},
# QSA fused QKV/gate + replicated indexer Q/K projection, TP=2.
(7296, 2560): {
1: SkinnyGemmConfig(1, 32, 1, vector_width=2, static_k=2560),
2: SkinnyGemmConfig(2, 32, 1, vector_width=2, static_k=2560),
4: SkinnyGemmConfig(4, 32, 1, vector_width=2, static_k=2560),
8: SkinnyGemmConfig(8, 32, 1, vector_width=2, static_k=2560),
16: SkinnyGemmConfig(16, 32, 4, vector_width=2, static_k=2560),
},
# GDN fused QKVZ projection, TP=2.
(8192, 2560): {
1: SkinnyGemmConfig(1, 128, 2, vector_width=4, static_k=2560),