ONNX Runtime with CUDA Execution Provider mismatch with CPU Execution Provider results

Dear CuDNN team,

I find a mismatch of ONNX Runtime with CUDA Execution Provider and with CPU Execution Provider on op AveragePool with asymmetric padding (pads=[0, 0, 2, 2]).

libcudnn8=8.5.0.96-1+cuda11.7

Can you help to take a look? Thanks!

Hi @user9611 ,
I doubt if this issue is related to cudnn.
I may suggest you to reach out to onnxruntime issues to get better assistance.

thanks