Hi all,
We’re working on a ML app that uses pytorch, but we’re having trouble specifying the gpu version of pytorch as a dependency for the build. Our project uses pyproject.toml to specify all dependancies and setuptools for the build. Our goal is to allow both CPU and GPU (if available) runs of pytorch after a user pip install’s our app without any further configuration needed.
we want to specify torch==2.0.1+cu118 for windows and ubuntu users- so that if they have a GPU we will be able to use gpu pytorch. (+cu118 will just default to cpu if there is no gpu available). We also want to specify the cpu-only torch==2.0.1 for osx users since there is not a CUDA build for OSX.
I dont believe there is a good way to do this solely using pyproject.toml and setuptools, since we have to specify the index-url for the +cu118 version of pytorch which is not supported by PEP and recent versions of python.
Ideally we dont have to bring in other dependency managers like poetry or pdm- but if there is a good solution with those I will consider it. The perfect solution would work with pyproject.toml and setuptools.
Thanks for the help