Hello. I am trying to run a simple test using

pants test ::

, but it is failing. The failure it shows is this

ProcessExecutionFailure: Process 'Building requirements.pex with 2 requirements: detectron2@ git+<https://github.com/facebookresearch/detectron2.git>, torch' failed with exit code 1

And

ModuleNotFoundError: No module named 'torch'

The requirements file is this

--extra-index-url <https://download.pytorch.org/whl/cu118>
torch
detectron2 @ git+<https://github.com/facebookresearch/detectron2.git>

Have you tried adding

torch

explicitly as a dependency in the

python_test

or

python_tests

target?

Yes, I have tried adding torch explicitly as a dependency in the python_test or python_tests target and it throws the same error

is this a module mapping thing?

pytorch

==

torch

? Or is

torch

the pip install command as well?

Somewhat confusingly, the Python package for PyTorch on PyPi is just called torch. Fortunately, if you'd try to pip install pytorch (a dummy empty package), you'll get a clear error: You tried to install "pytorch". The package named for PyTorch is "torch".

So, I can't even get that working via

pip install

outside of pants - complains about torch. Is this supposed to be installed with an older pip or something?

I think that Pants won't use the arguments found in requirements.txt files. I think you might want to set/add to the

[python-repos].indexes

config setting (doc on using custom repositories)

Yeah, there was more pants stuff to look into here - but the requirements in isolation didn't work with the same error, so there wasn't a hope and a prayer of pants working

We worked on it some more, and it seems like it's an issue with detectron, documenting our findings: • someone else ran into it: https://pantsbuild.slack.com/archives/C046T6T9U/p1685543367756089 . Seems the problem is that detectron needs

torch

to be installed when it's building the wheels (

import torch

in

setup.py

) • the proposed answer the last time this came up, to use a fork to patch it to specify that it needs

torch

during setup, makes sense (not much change if you're already using git as a source) • You could also pre-build the wheels yourself targeting matching versions of torch and cuda. A complication is building these for all torch and cuda you'll use, but there might not be that many. You could then: ◦ host them in a way

[python-repos].find_links

can find them, which could just be some files in a storage bucket (it looks like detectron used to do this, but not for their more recent versions) ◦ build them and push to a python package registry and use

[python-repos].indexes

(if you've already got an internal one, this might be pretty easy)