🐛 [Bug] Segmentation Fault in TorchTensorRTModule with PyTorch 2.5.0.dev Due to full_like Converter Replacement

[chohk88]

Bug Description

We are encountering a segmentation fault when running the test case for the full_like converter. We have identified some potential causes, but a definitive solution has not yet been found.

Code Context

File: py/torch_tensorrt/dynamo/runtime/_TorchTensorRTModule.py

outputs: List[torch.Tensor] = torch.ops.tensorrt.execute_engine(
    list(input_tensors), self.engine
)

This line is part of the forward method, defined as follows:

def forward(self, *inputs: Any) -> torch.Tensor | Tuple[torch.Tensor, ...]:

Class: TorchTensorRTModule(torch.nn.Module)

Behavior Based on PyTorch Version

• PyTorch 2.4.0a: The function executes without any issues.

• PyTorch 2.5.0.dev20240807+cu124: When running the full_like converter test case, the program terminates abruptly with a segmentation fault, and no error message is displayed.

Additional Details

This issue consistently reproduces both in the CI environment and in a local setup using torch 2.5.0.dev20240807+cu124.
When converting, the shape information is extracted from the inputs, but the input tensor itself is lost.

To Reproduce

With PyTorch 2.5.0, run the following command:

python -m pytest tests/py/dynamo/lowering/test_decompositions.py -k test_lowering_full_like_module -s

This will reproduce the segmentation fault.

Expected Behavior

With PyTorch 2.4.0a, running the same test case should complete successfully without any issues:

python -m pytest tests/py/dynamo/lowering/test_decompositions.py -k test_lowering_full_like_module -s

Environment

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

• pytorch-triton: 3.0.0+dedb7bdf33
• torch: 2.5.0.dev20240807+cu124
• torch_tensorrt: 2.5.0.dev0+cf390fa55 /TensorRT/py
• torchvision: 0.19.0.dev20240617+cu124