diff --git a/src/hydrogen/device/rocBLAS_API.cpp b/src/hydrogen/device/rocBLAS_API.cpp index 8be85e6aa..75c5ddcdb 100644 --- a/src/hydrogen/device/rocBLAS_API.cpp +++ b/src/hydrogen/device/rocBLAS_API.cpp @@ -75,6 +75,10 @@ class ResultMgr H_CHECK_HIP(hipFreeAsync(device_, stream)); #endif // HYDROGEN_HAVE_CUB + // Sync stream to match cuBLAS behavior (cuBLAS docs here: + // https://docs.nvidia.com/cuda/cublas/#scalar-parameters) + H_CHECK_HIP(hipStreamSynchronize(stream)); + // Reset pointer mode H_CHECK_ROCBLAS(rocblas_set_pointer_mode(handle_, rocblas_pointer_mode_host)); }