🐛 [Bug] torch_tensorrt.compile does not work with nn.ConvTranspose2d and output_padding

[xavierjimenezp]

Bug Description

Trying to use torch_tensorrt.compile to compile a model using nn.ConvTranspose2d with output_padding = 1 raises the following error:

RuntimeError: Target aten.convolution.default does not support `transposed=True` 

While executing %convolution : [num_users=1] = call_function[target=torch.ops.aten.convolution.default](args = (%x, %deconv_weight, %deconv_bias, [2, 2], [0, 0], [1, 1], True, [1, 1], 1), kwargs = {_itensor_to_tensor_meta: {<tensorrt_bindings.tensorrt.ITensor object at 0x72fa0452d6b0>: ((4, 256, 296, 296), torch.float32, False, (22429696, 87616, 296, 1), torch.contiguous_format, False, {})}})

Note: using default value output_padding = 0 works fine. I have not tried with other values.

To Reproduce

The followig code allows to reproduce the error

import torch
import torch.nn as nn
import torch_tensorrt


class ToyModel(nn.Module):
    def __init__(self) -> None:
        super(ToyModel, self).__init__()

        self.deconv = nn.ConvTranspose2d(
            256,
            128,
            kernel_size=3,
            output_padding=1,
            stride=2,
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.deconv(x)
        return x


def main():
    # Create a toy model instance
    model = ToyModel().eval().cuda()

    # Create dummy input
    input_tensor = torch.randn(4, 256, 296, 296).cuda()

    # Compile the model with torch_tensorrt
    trt_model = torch_tensorrt.compile(
        model,
        inputs=[input_tensor],
        enabled_precisions={torch.float32},
        min_block_size=1,
    )


if __name__ == "__main__":
    main()

Environment

• Torch-TensorRT Version (e.g. 1.0.0): 2.5.0
• PyTorch Version (e.g. 1.0): 2.5.1+cu124
• CPU Architecture: x86_64
• OS (e.g., Linux): Ubuntu 22.04
• How you installed PyTorch (conda, pip, libtorch, source): pip
• Build command you used (if compiling from source):
• Are you using local sources or building from archives: No
• Python version: 3.11.11
• CUDA version: 12.6
• GPU models and configuration: NVIDIA GeForce RTX 4090
• Any other relevant information:

[zewenli98]

seems to related to #3343