Unable to add `dgl` as Python dependency in TOML specifications

[laserkelvin]

Unable to add `dgl` as Python dependency in TOML specifications

[jermainewang]

Hi @laserkelvin , we followed the practice of PyTorch to host different packages for different CUDA versions. I wonder how did you resolve the problem for PyTorch?

[laserkelvin]

I see what you mean, having taken a second look at the PyTorch instructions. However running pip install torch still works, albeit I think the wheels being put on PyPI default to being CUDA 12.1. For packaging I would've thought a CPU target as the default would make more sense, but I can understand their reason for doing so. 🤷🏼‍♂️

For DGL, however, is installation from PyPI just being dropped completely? Like I said, I found it a good user experience to be able to pip install dgl and just have a working version to start development out of the box not needing to worry about accelerators or where to source the wheels from. I'd ask the DGL team to consider maintaining the PyPI distribution, even if it's just CPU only.

A potential workaround for pyproject.toml is to point to the exact wheel, but it means that I have to pin the package to a specifiy Python version and distribution - that's also not ideal IMHO.

e.g.

dependencies = [
   "torch",  # installs from PyPI, but just adds CUDA bloat if you're not using CUDA which I can live with because it works
   "dgl==2.0.0",  # doesn't work because there are no >1.1.3 wheels on PyPI
   "dgl @ https://data.dgl.ai/wheels/dgl-2.0.0-cp310-cp310-manylinux1_x86_64.whl",   # works, but forces Python 3.10
]

At the end of the day I'm aiming to have a complete out of the box experience, and making the most performant configuration optional for the end-user - they can choose to rebuild DGL from source with whatever libraries they want, but as a package maintainer I'm more interested in providing an environment that at least works.