Inconsistent inference results between PyTorch and converted TensorRT model with Pad operator

[hongliyu0716]

Description:

I am encountering an issue when converting Torch models that solely consist of the Pad operator to TensorRT. The problem is that the results obtained after the conversion process are inconsistent with the original Torch implementation.

Reproduce

This issue can be reproduced by the following script:

import torch
from torch.nn import Module
from torch2trt import torch2trt

para_0 = torch.randn([1, 2, 2, 2], dtype=torch.float32).cuda()
para_1 = (2, 2, 2, 2)
para_2 = 'replicate'
class pad(Module):
    def forward(self, *args):
        return torch.nn.functional.pad(args[0], para_1,para_2,)
model = pad().float().eval().cuda()
model_trt = torch2trt(model, [para_0])

output = model(para_0)
output_trt = model_trt(para_0)
print(torch.max(torch.abs(output - output_trt)))

The output is:

tensor(2.1606, device='cuda:0')

Environment

• torch: 2.1.1
• torch2trt: 0.4.0
• tensorrt: 8.6.1