This is an automated email from the ASF dual-hosted git repository. masahi pushed a commit to branch unity in repository https://gitbox.apache.org/repos/asf/tvm.git
commit 6e3c8d34cdc0b73ab1dcb6ccf47ae0d3d42993b3 Author: Masahiro Masuda <masahi...@gmail.com> AuthorDate: Mon Nov 6 21:04:30 2023 +0000 use fp32 accum in cublas --- src/runtime/contrib/cublas/cublas.cc | 1 + 1 file changed, 1 insertion(+) diff --git a/src/runtime/contrib/cublas/cublas.cc b/src/runtime/contrib/cublas/cublas.cc index 10db3b1c50..f4b28d29a8 100644 --- a/src/runtime/contrib/cublas/cublas.cc +++ b/src/runtime/contrib/cublas/cublas.cc @@ -163,6 +163,7 @@ void CallCublasLt(cublasLtHandle_t hdl, cudaStream_t stream, const DLTensor* A, } if (TypeMatch(C->dtype, kDLFloat, 16)) { + // Disable fp16 accum to align with PT c_type = CUDA_R_16F; compute_type = CUBLAS_COMPUTE_16F; scale_type = CUDA_R_16F;