🐛 [Bug] aten::_convolution output shape no longer matches Pytorch after TensorRT 10 upgrade

[matthewfl]

Bug Description

The output shape of aten::_convolution no longer matches pytorch after the TensorRT 10 upgrade.

I have noticed that the output shape is correct when I pass in the weight matrix as a constant instead of as a input as done in the reproducer below.

To Reproduce

import torch, torch_tensorrt

class Model(torch.nn.Module):
    def forward(self, x,y):
        return torch.ops.aten._convolution(x, y, None, [1,1], [1,1], [1,1],False,[0,0],144,True,True,True,True)

model = Model().cuda()
input = torch.rand(20,144,55,55).cuda(), torch.rand(144,1,3,3).cuda()

compiled = torch_tensorrt.ts.compile(torch.jit.trace(model,input),input)

assert compiled(*input).shape == model(*input).shape

Expected behavior

The program should work without the assert failing.

Environment

Build information about Torch-TensorRT can be found by turning on debug messages

• Torch-TensorRT Version (e.g. 1.0.0): 2.4.0
• PyTorch Version (e.g. 1.0): 2.4.1
• CPU Architecture:
• OS (e.g., Linux): Linux
• How you installed PyTorch (conda, pip, libtorch, source):
• Build command you used (if compiling from source):
• Are you using local sources or building from archives:
• Python version:
• CUDA version:
• GPU models and configuration:
• Any other relevant information:

Additional context

[matthewfl]

Hi @zewenli98, is there any update on this?

[zewenli98]

@matthewfl It seems to be the issue of TS path that had no longer been supported. Can you try replacing

compiled = torch_tensorrt.ts.compile(torch.jit.trace(model,input),input)

with

exp_program = torch.export.export(model, input)
compiled = torch_tensorrt.dynamo.compile(exp_program, input, min_block_size=1)

[matthewfl]

seems like this issue was related to the output/post padding being handled differently between the two code paths https://github.com/pytorch/TensorRT/blob/main/core/conversion/converters/impl/conv_deconv.cpp#L201

[zewenli98]

seems like this issue was related to the output/post padding being handled differently between the two code paths https://github.com/pytorch/TensorRT/blob/main/core/conversion/converters/impl/conv_deconv.cpp#L201

right. dynamo path uses pre-padding by default.