diff --git a/vllm/models/qwen4_exp/nvidia/low_latency_gemm.py b/vllm/models/qwen4_exp/nvidia/low_latency_gemm.py index acc5b3e6bd..2ee1c7cb5a 100644 --- a/vllm/models/qwen4_exp/nvidia/low_latency_gemm.py +++ b/vllm/models/qwen4_exp/nvidia/low_latency_gemm.py @@ -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): {