🐛 [Bug] Missing directory and file in FX path (Python only) installation

[HolyWu]

Bug Description

1. torch_tensorrt.fx.converters.impl was not added to the packages and package_dir list, hence the impl directory was not installed.
2. _Input.py was not installed even though you had added it to package_data because torch_tensorrt was not in the packages list in FX_ONLY. Instead, you have to use py_modules to copy modules that aren't part of a package.
3. _Input.py imported _enums.py which was also not installed in FX_ONLY. And _enums.py imported stuff from torch_tensorrt._C which was unavailable in FX_ONLY.
4. torch_tensorrt/dynamo/__init__.py imported compile from torch_tensorrt/dynamo/backend, but the backend directory was not installed in FX_ONLY.

To Reproduce

python3 setup.py install --fx-only

Environment

• Torch-TensorRT Version (e.g. 1.0.0): 1.4.0
• PyTorch Version (e.g. 1.0): 2.0.1+cu118
• CPU Architecture: x86_64
• OS (e.g., Linux): Windows
• 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:
• Python version: 3.11.3
• CUDA version: 11.8
• GPU models and configuration:
• Any other relevant information:

[gs-olive]

Thank you for reporting this - regarding the Dynamo comments, our torch_tensorrt.dynamo path still has a few C++ dependencies, and so it is currently not able to be imported without the full python setup.py install, though we have open issues to resolve this: #1943, #1993.

[apbose]

Hi, thanks for reporting the issue. I wanted to clarify. I have added torch_tensorrt.fx.converters.impl to package list and I do see it in the site_packages of torch_tensorrt installed, however even if I do not mention the torch_tensorrt.fx.converters.impl in package list, the impl is still installed. So I wanted to clarify on what issue you are facing currently. Thanks!

[HolyWu]

https://colab.research.google.com/drive/1hJwftiDT1U7rzGu_YeuVA9rrffikCvR6?usp=sharing

[apbose]

Ok, looking into this.

[apbose]

Thanks for the repo steps above.
So when I install using python setup.py install --fx-only --release and do import torch_tensorrt.fx it shows ModuleNotFoundError: No module named 'torch_tensorrt._C. As mentioned earlier since dynamo still has C++ dependencies these won't be imported until we do the full python setup.py install

[apbose]

From torch_tensorrt._compile import * ts package is getting imported from torch_tensorrt/_compile.py .
--fx-only is not proper at the moment. PEP517- #2056 changes should address the cross C++ and python dependencies