diff --git a/vllm/models/qwen4_exp/nvidia/low_latency_gemm.py b/vllm/models/qwen4_exp/nvidia/low_latency_gemm.py index 2ee1c7cb5a..ea56750702 100644 --- a/vllm/models/qwen4_exp/nvidia/low_latency_gemm.py +++ b/vllm/models/qwen4_exp/nvidia/low_latency_gemm.py @@ -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),