Optimize cutlass int8 gemm kernel for large M on SM89 Ada GPU (#10714)

2025-09-22 08:04:27 +08:00
parent 2f18602f13
commit 0f04a5f428
1 changed files with 2 additions and 2 deletions
--- a/sgl-kernel/csrc/gemm/int8_gemm_kernel.cu
+++ b/sgl-kernel/csrc/gemm/int8_gemm_kernel.cu
@@ -409,8 +409,8 @@ void sm89_dispatch_shape(
    cutlass_int8_scaled_mm<
        ElementOutput,
        ArchTag,
-        cutlass::gemm::GemmShape<32, 64, 128>,
-        cutlass::gemm::GemmShape<16, 64, 64>,
+        cutlass::gemm::GemmShape<128, 128, 64>,
+        cutlass::gemm::GemmShape<64, 64, 64>,
        InstructionShape,
        5>(out, mat_a, mat_b, scales_a, scales_b, bias);
  }