This is on a devserver? Usually a hang is because your proxy settings are wrong.

(Anyway, I agree: cool build systems don't connect to the network when you run them.)

I think @ezyang is right, it'll try to install these packages before it starts building. But this wasn't buying us much anyways

Which packages is it trying to install? Is it due to the NNPACK dependency?

IIRC, nnpack has a few dependency download here: https://github.com/Maratyszcza/NNPACK/tree/master/cmake

Also, I am a bit curious why we need pytest-runner for setup, since we don't run test during setup time I think...?

This (and also test_requires) is for delegating python setup.py test to pytest: https://docs.pytest.org/en/latest/goodpractices.html#integrating-with-setuptools-python-setup-py-test-pytest-runner. I think it's fine to remove it because IIRC pytest doesn't really work in pytorch.

Good point @ezyang - likely proxy in this case. git should be setup correctly, so hopefully NNPACK is a non-issue. Generally, I would like to avoid network calls when running this, though, since it's a bit weird to do so. @bddppq and @pjh5 should I remove test_requires as well?

A few points:

1. PyTorch does work with pytest, at least if you use pytest to invoke tests individually. I've used it in the past because it's test filtering is superior to the default Python test runner

2. There isn't really any reason we should force pytest-runner on users on install: it's just for tests (so, if a user is not interested in running tests, it will still try to install it), and it's not even essential to run tests (you can always use the default runner)

So, I support this patch as written, just not for the reasons originally proffered.

@ezyang pytest does not work in PyTorch when invoking from the root directory if tests are not invoked individually.

At this point I'm going to remove tests_require too. I'm a bit concerned about all of the additional deps in install_requires as well, especially since we want to have FULL_CAFFE2=1 in all cases shortly. Is it better to just not have these (similar to PyTorch) and just have the user install them as needed? Or, alternatively, have them in requirements.txt. I like that last option, honestly.

I think that install_requires affects wheel installs, causing pip to install the needed dependencies before installing the wheel. It can be moved to requirements.txt as well, though I think requirements.txt is simply convention and not used by pip itself