Non-zero status code returned while running ConvTranspose node.

[Jerry-Master]

Describe the issue

When running my onnx model on C++ in CPU everything works perfectly. However, when running it with the Cuda provider it throws this error:

2024-06-13 16:21:29.9651844 [E:onnxruntime:InferenceEngineORT, cuda_call.cc:118 onnxruntime::CudaCall] CUDNN failure 3: CUDNN_STATUS_BAD_PARAM ; GPU=0 ; hostname=008C00014 ; file=C:\a\_work\1\s\onnxruntime\core\providers\cuda\nn\conv_transpose.cc ; line=318 ; expr=cudnnAddTensor(GetCudnnHandle(context), &alpha, s_.b_tensor, b_data, &alpha, s_.y_tensor, y_data);
2024-06-13 16:21:29.9907653 [E:onnxruntime:, sequential_executor.cc:516 onnxruntime::ExecuteKernel] Non-zero status code returned while running ConvTranspose node. Name:'_inlfunc_torch_nn_modules_conv_ConvTranspose2d_p_m_up3_0_1_ConvTranspose_7' Status Message: CUDNN failure 3: CUDNN_STATUS_BAD_PARAM ; GPU=0 ; hostname=008C00014 ; file=C:\a\_work\1\s\onnxruntime\core\providers\cuda\nn\conv_transpose.cc ; line=318 ; expr=cudnnAddTensor(GetCudnnHandle(context), &alpha, s_.b_tensor, b_data, &alpha, s_.y_tensor, y_data);

To reproduce

The model has as input l_x_ tensor of shape 1,3,128,256 and l_k_ tensor of shape 1,1,32,32 and the output is called p_8 with shape 1,3,128,256. The model is uploaded to drive here. It was exported from usrnet repo using dynamo exporter and opset 18.

The following python code helps reproduce the issue:

import argparse
import cv2
import onnxruntime
import numpy as np

def _create_parser():
    parser = argparse.ArgumentParser()
    parser.add_argument('--onnx-path', type=str, default="usrnet_128x256_32x32.onnx")
    return parser


def main():
    parser = _create_parser()
    args = parser.parse_args()

    x = np.ones((1, 3, 128, 256)).astype(np.float32)
    k = np.ones((1, 1, 32, 32)).astype(np.float32)

    # compute ONNX Runtime output prediction
    ort_session = onnxruntime.InferenceSession(args.onnx_path, providers=['CUDAExecutionProvider'])  
    ort_inputs = {'l_x_': x, 'l_k_': k}
    ort_outs = ort_session.run(['p_8'], ort_inputs)

    out = ort_outs[0].squeeze(0).transpose(1, 2, 0)
    out_norm = np.zeros_like(out)
    out_norm = cv2.normalize(out, -1, 0, 255, norm_type=cv2.NORM_MINMAX)
    cv2.imwrite('out.png', out_norm)
    

if __name__ == '__main__':
    main()

Urgency

Yes

Platform

Windows

OS Version

11

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.18.0

ONNX Runtime API

C++

Architecture

X64

Execution Provider

CUDA

Execution Provider Library Version

CUDA 11.6

[github-actions]

This issue has been automatically marked as stale due to inactivity and will be closed in 30 days if no further activity occurs. If further support is needed, please provide an update and/or more details.