Fail to call cuDNN or cuBLAS routine with explicit Graph APIs in CUDA Graph

I’m tring to create a graph with a cuDNN activation function as a node using explicit Graph APIs, like cudaGraphAddHostNode, however, I got CUDNN_STATUS_EXECUTION_FAILED returned. I’m doing this with Tesla V100, CUDA-11.0 and cuDNN-7.6.5.
I can create the graph correctly using stream capture, but I wonder if there is a right way to add cuDNN or cuBLAS node with explicit Graph APIs because it’s more convenient when the graph structure is very complex.

1 Like