diff --git a/example/65_gemm_multiply_multiply/moe_gemm1_xdl_fp8.cpp b/example/65_gemm_multiply_multiply/moe_gemm1_xdl_fp8.cpp index 66825edcf9..0633cc3eab 100644 --- a/example/65_gemm_multiply_multiply/moe_gemm1_xdl_fp8.cpp +++ b/example/65_gemm_multiply_multiply/moe_gemm1_xdl_fp8.cpp @@ -356,7 +356,9 @@ int main(int argc, char* argv[]) KBatch, a_element_op, b_element_op, - cde_element_op); + cde_element_op, + N, + K); if(!device_op.IsSupportedArgument(argument)) { diff --git a/example/65_gemm_multiply_multiply/moe_gemm1_xdl_pk_i4.cpp b/example/65_gemm_multiply_multiply/moe_gemm1_xdl_pk_i4.cpp index c06e595c0f..b79889b991 100644 --- a/example/65_gemm_multiply_multiply/moe_gemm1_xdl_pk_i4.cpp +++ b/example/65_gemm_multiply_multiply/moe_gemm1_xdl_pk_i4.cpp @@ -438,7 +438,9 @@ int main(int argc, char* argv[]) KBatch, a_element_op, b_element_op, - cde_element_op); + cde_element_op, + N, + K); if(!device_op.IsSupportedArgument(argument)) { diff --git a/example/65_gemm_multiply_multiply/moe_gemm2_xdl_fp8.cpp b/example/65_gemm_multiply_multiply/moe_gemm2_xdl_fp8.cpp index 0d12441016..5abac7c8cf 100644 --- a/example/65_gemm_multiply_multiply/moe_gemm2_xdl_fp8.cpp +++ b/example/65_gemm_multiply_multiply/moe_gemm2_xdl_fp8.cpp @@ -365,7 +365,9 @@ int main(int argc, char* argv[]) KBatch, a_element_op, b_element_op, - cde_element_op); + cde_element_op, + N, + K); if(!device_op.IsSupportedArgument(argument)) { diff --git a/example/65_gemm_multiply_multiply/moe_gemm2_xdl_pk_i4.cpp b/example/65_gemm_multiply_multiply/moe_gemm2_xdl_pk_i4.cpp index c80b01d8c5..d3bf67ca86 100644 --- a/example/65_gemm_multiply_multiply/moe_gemm2_xdl_pk_i4.cpp +++ b/example/65_gemm_multiply_multiply/moe_gemm2_xdl_pk_i4.cpp @@ -405,7 +405,9 @@ int main(int argc, char* argv[]) KBatch, a_element_op, b_element_op, - cde_element_op); + cde_element_op, + N, + K); if(!device_op.IsSupportedArgument(argument)) { diff --git a/include/ck/tensor_operation/gpu/device/impl/device_moe_gemm.hpp b/include/ck/tensor_operation/gpu/device/impl/device_moe_gemm.hpp index 950fe0236d..3d334c0fcf 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_moe_gemm.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_moe_gemm.hpp @@ -391,7 +391,9 @@ struct DeviceMoeGemm : public DeviceGemmMultipleDSplitKBPreShuffle(p_sorted_token_ids), static_cast(p_sorted_expert_ids), @@ -432,7 +434,9 @@ struct DeviceMoeGemm : public DeviceGemmMultipleDSplitKBPreShuffle(nullptr, nullptr,