Cuda-compatible installation without explicit use of --global-option

[fuhuifang]

We are trying to use cuda-compatible pytorch-extension in a conda/docker environment, with a conda environment file like this:

dependencies:
  - python=3.7
  - pip=20.2.4
  - pip:
    - numpy==1.21.2 
    - torch==1.7.1+cu110 
    - typing-extensions==3.10.0.2 
    - pytorch-extension==0.2 --global-option "cpp_ext" --global-option "cuda_ext"

However, we got UserWarning of disabling all use of wheels due to use of --global-option and it could only use .tar.gz for all dependencies. Many packages (including numpy, cuda-compatible pytorch, etc.) do not have sdist support, so we saw "no matching distribution" error for those packages.

My question: is there a way to build cuda-compatible pytorch-extension without explicit use of --global-option "cpp_ext" --global-option "cuda_ext"?

[artitw]

Thanks for posting this issue. Have you gotten the setup working when manually installing the dependencies in a local environment? If so, I think the issue might be limitations with conda, and would therefore recommend using docker instead.

[fuhuifang]

Update: we’re unblocked by building pytorch-extension in the docker image and installing the remaining packages in a conda environment on top of it.

[artitw]

What a splendid solution to isolate the conda issues from pytorch-extension. Thanks for sharing