Inconsistent inference results between PyTorch and TensorRT using torch2trt with ELU operator

[Thrsu]

Description:

I'm experiencing a discrepancy between the inference results of my PyTorch model and the TensorRT model obtained by converting it using the torch2trt tool.

Reproduce

This can be reproduced by the following script:

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

model = torch.nn.ELU(inplace=True,).cuda()
input_data=torch.randn([1, 3, 10, 10], dtype=torch.float32).cuda()
model_trt = torch2trt(model, [input_data])
y = model(input_data)
y_trt = model_trt(input_data)

# check the output against PyTorch
print(torch.max(torch.abs(y - y_trt)))

The output is:

tensor(0.0909, device='cuda:0')

Environment

• torch: 1.11.0
• torch2trt: 0.4.0
• tensorrt: 8.6.1.6

[Thrsu]

Moreover, I noticed the inference results for LeakyRelu operator are also inconsistent between PyTorch and TensorRT. The script is as below:

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

model = torch.nn.LeakyReLU(inplace=True,).cuda()
input_data = torch.randn([3, 2, 5], dtype=torch.float32).cuda()
model_trt = torch2trt(model, [input_data])
y = model(input_data)
y_trt = model_trt(input_data)

# check the output against PyTorch
print(torch.max(torch.abs(y - y_trt)))