diff --git a/src/nvfp4_megamoe_kernel/cutlass_nvfp4_gemm/cutlass_nvfp4_gemm.cu b/src/nvfp4_megamoe_kernel/cutlass_nvfp4_gemm/cutlass_nvfp4_gemm.cu index 566444b3..1ce952b7 100644 --- a/src/nvfp4_megamoe_kernel/cutlass_nvfp4_gemm/cutlass_nvfp4_gemm.cu +++ b/src/nvfp4_megamoe_kernel/cutlass_nvfp4_gemm/cutlass_nvfp4_gemm.cu @@ -135,8 +135,6 @@ __global__ void remap_sf_to_cutlass_kernel( constexpr int LayoutRank = cute::rank_v; - int dst_idx = 0; - // Use crd2idx to map flat logical coordinate to CUTLASS physical index. // Decompose flat (mn, k_sf) into hierarchical coordinates matching the atom layout: // Shape: ((32, 4, mn_tiles), (SFVecSize, 4, k_tiles), ...) // First group: mn = m0 + 32*m1 + 128*mt where m0 in [0,32), m1 in [0,4), mt = mn/128