diff --git a/kernels/sgemm_tcore/sgemm_impl.hpp b/kernels/sgemm_tcore/sgemm_impl.hpp index 9c2b0dc5..d2e9af24 100644 --- a/kernels/sgemm_tcore/sgemm_impl.hpp +++ b/kernels/sgemm_tcore/sgemm_impl.hpp @@ -19,7 +19,7 @@ using float_type = float16_t; // Generate kernel for the Hopper-style SMEM-decoupled tensor core. This uses // asynchronous HGMMA and HGMMA_WAIT instructions. -#define TENSOR_HOPPER 1 +#define TENSOR_HOPPER 0 // Constraints on parameters: // * Memory: