The TensorRT result is incorrect when the kernel_size and stride of Conv2D are both [14, 14]

Description

The TensorRT result is incorrect when the conv2d's kernel_size and stride are both [14, 14] .

Environment

TensorRT Version: 10.14.1.48

NVIDIA GPU: RTX 3090

NVIDIA Driver Version: 570.133.07

CUDA Version: 12.8

CUDNN Version: 9.1.0.70

ONNX Version: 1.17.0

ONNXRuntime Version: 1.19.0

Operating System: Linux

Python Version (if applicable): 3.10.14

Tensorflow Version (if applicable):

PyTorch Version (if applicable): 2.6.0

Baremetal or Container (if so, version):

Relevant Files

test_image.jpg

Steps To Reproduce

You can reproduce my issue by running the following python code, my result is: Max Diff: tensor(2.8298, device='cuda:0')

import os
import torch
import onnxruntime
import tensorrt as trt
from PIL import Image
from torchvision.transforms import v2

if __name__ == "__main__":
    # you can download this image in (https://github.com/user-attachments/assets/18dd13ae-c030-42f4-86f7-1ae4e1cfe422)
    image = Image.open('test_image.jpg')  
    width, height = image.size
    image = v2.functional.to_image(image)
    transform = v2.Compose([
        v2.ToDtype(torch.uint8, scale=True),
        v2.Resize(size=(1008, 1008)),
        v2.ToDtype(torch.float32, scale=True),
        v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])])
    image = transform(image).unsqueeze(0).to('cuda').to(torch.float32)
    
    model = torch.nn.Sequential(torch.nn.Conv2d(3, 1024, kernel_size=(14, 14),stride=(14, 14))).to('cuda').eval()
    torch.onnx.export(model, image, "conv.onnx",input_names=['image'], output_names=['output'])

    os.system("trtexec --onnx=conv.onnx --saveEngine=conv.engine")
    assert os.path.exists('conv.engine')
    
    with open('conv.engine', "rb") as f, trt.Runtime(trt.Logger()) as runtime:
        engine = runtime.deserialize_cuda_engine(f.read())
        context = engine.create_execution_context()
    trt_input = [image.clone()]
    trt_output = [torch.empty(1, 1024, 72, 72, dtype=torch.float32, device='cuda')]
    bindings = [v.data_ptr() for v in (trt_input + trt_output)]
    context.execute_v2(bindings=bindings)

    sess = onnxruntime.InferenceSession('conv.onnx', providers=["CUDAExecutionProvider"]) 
    onnx_output = sess.run(['output'], {'image': image.clone().cpu().numpy()})
    
    trt_output, onnx_output = trt_output[0], torch.from_numpy(onnx_output[0]).to(trt_output[0])
    assert trt_output.shape == onnx_output.shape
    print("Max Diff: ", (trt_output - onnx_output).abs().max())

Commands or scripts:

Have you tried the latest release?: Yes

Attach the captured .json and .bin files from TensorRT's API Capture tool if you're on an x86_64 Unix system

Can this model run on other frameworks? It can be run.

Start by running the supplied Python reproduction with the listed TensorRT, CUDA, PyTorch, and GPU versions, then compare the TensorRT and ONNX Runtime outputs for conv.onnx. Done means the Conv2D case with kernel_size and stride [14, 14] no longer produces the reported maximum difference.