rusty1s/pytorch_cluster

setup.py depends on `torch`

dyollb opened this issue · 4 comments

dyollb commented

I would like to create a python package that installs its dependencies properly using pip.
Unfortunately, the install fails because pytorch-cluster expects torch to already be installed.

Simply trying to install torch-cluster using pip install torch-cluster fails. First, you need to install pytorch, and only then install pytorch-cluster.

image

Originally I hoped it would simply suffice to add torch here. But the setup.py depends on torch to build. It seems that pytorch C++ extensions are done in this way.

Is there any workaround (using pip)?

rusty1s commented

The only workaround for this right now is to install from the wheels page (i.e., by specifying -f in the pip install command). This way, torch is not required to install the package.

dyollb commented

hmm. I did that but it doesn't work. I tried installing (copy pasted from the README) via

pip install torch-cluster -f https://data.pyg.org/whl/torch-2.1.0+cpu.html

which gives me:
image

System specs:

  • Python 3.10.11
  • Apple M2
  • Sonoma 14.0
  • Darwin Kernel Version 23.0.0

Btw, independent of the issue above, surely torch-cluster should explicitly depend on torch. Currently it doesn't (see here):
image

  • 👍1
rusty1s commented

That's because you are trying to install on Apple M2 for which do not exist pre-built wheels (only for Intel-Mac at the moment) :(

dyollb commented

Okay, understood. Thanks!