How to specify os-specific wheel files in the Pipfile

Question

I'm trying to use the pipenv and pytorch.

To install pytorch in windows, I have to write following codes into the Pipfile:

[packages]
torch = {file = "http://download.pytorch.org/whl/cpu/torch-0.4.1-cp37-cp37m-win_amd64.whl"}

However, the wheel file is different for the linux.

[packages]
torch = {file = "http://download.pytorch.org/whl/cpu/torch-0.4.1.post2-cp37-cp37m-linux_x86_64.whl "}

How to specify both of them in Pipfile?

Answer 1

This article demonstrates how to structure your Pipfile to use pytorch across multiple platforms. I tweaked their example to deal with whl files on my local file system:

[packages]
pyfoo = {path = "./../pyfoo/dist/pyfoo-1.1.0-cp37-cp37m-linux_x86_64.whl", platform_system = "== 'Linux'"}
pyfoo-win = {path = "./../pyfoo/dist/pyfoo-1.1.0-py3-none-any.whl", platform_system = "== 'Windows'"}

In this case, pyfoo is an internal libarary that has been built using python setup.py build and python setup.py bdist_wheel.

The article uses some kind of pipenv generated hash in place of pyfoo-win. I could not coerce pipenv into generating that hash, so I created my own string. It may just be the prefix of the hash in the lock file.

Note: This approach is working for me, but there is a downer: both whl files need to be present when you install packages. pipenv sync bombs out if one of them is missing, even though it really only needs one of the two. Interestingly, the contents of the irrelevant file does not seem to matter. On my linux machine, I did an echo 'hello' > to the windows whl file, and pipenv was happy with it.