diff --git a/src/nvfp4_megamoe_kernel/cutlass_nvfp4_gemm/setup.py b/src/nvfp4_megamoe_kernel/cutlass_nvfp4_gemm/setup.py index d813b700..f82528e0 100644 --- a/src/nvfp4_megamoe_kernel/cutlass_nvfp4_gemm/setup.py +++ b/src/nvfp4_megamoe_kernel/cutlass_nvfp4_gemm/setup.py @@ -53,7 +53,8 @@ setup( CUDAExtension( name="cutlass_nvfp4_gemm._C", sources=[ - "cutlass_nvfp4_gemm/pytorch_binding.cpp", + "pytorch_binding.cpp", + "cutlass_nvfp4_gemm.cu", ], extra_include_paths=extra_include_paths, extra_cuda_cflags=extra_cuda_cflags,