Runtime dependencies are not (correctly) specified

[sisp]

torch-sparse has a number of runtime dependencies, but only scipy (without a version specifier) is specified in setup.py via install_requires. This breaks dependency resolution and makes it hard to use torch-sparse in a properly set up project.

I've identified the following dependencies:

• numpy
• scipy
• torch
• torch-scatter
• typing-extensions>=3.7.4; python_version < '3.8' (because Final was added to the typing package in Python v3.8 and to typing-extensions in v3.7.4)

You should also determine proper version specifiers for the other dependencies, so that a dependency resolver can correctly determine the versions of all packages to be installed.

[rusty1s]

Thanks for creating the issue. We sadly cannot add torch nor torch-scatter as a dependency, as torch needs to be installed in advance to make installation succeed, and torch-scatter is best installed from wheels.

As such, the newly provided conda installation procedure fixes those problems and is now the recommended way to install torch-sparse.

[sisp]

[...] as torch needs to be installed in advance to make installation succeed [...]

If we could solve #156, then torch could be added as a dependency, right?

[...] and torch-scatter is best installed from wheels.

So the problem is that torch-scatter is only published as a source distribution package on pypi.org, right? Is there any reason why you don't publish wheels there?

As such, the newly provided conda installation procedure fixes those problems and is now the recommended way to install torch-sparse.

Conda is becoming a bit of an issue for users engaged in commercial activities (i.e. enterprises) because they've changed their terms of service. I'd be nice to have a proper setup path using standard Python packages. 🙂

[rusty1s]

I'm not a big fan of adding torch as dependency, as it is totally unclear which PyTorch version (CUDA, CPU) gets installed (the same holds true for torch-scatter as its version needs to align with the PyTorch/CUDA version. That's also the reason why we cannot simply push wheels to PyPi, as binaries depend both on PyTorch and CUDA version.

[sisp]

Ah, yes, I see. 🤔

[github-actions]

This issue had no activity for 6 months. It will be closed in 2 weeks unless there is some new activity. Is this issue already resolved?