Any Plan for releasing the latest version of DGL on Pip?

[kenko911]

Hi DGL developers,

I am the developer of MatGL (https://github.com/materialsvirtuallab/matgl), a framework built on DGL and PyTorch. Recently, we encountered an issue updating our PyTorch version to 2.4.0 via bot because the latest DGL version available on Pip, 2.2.1, does not support PyTorch 2.4.0. I am wondering if there are any plans to release the latest version on Pip. Thank you!

Best
Kenko

[Bin-cc]

Hi Kenko,

Have you successfully solved this problem? I also encountered such situation, then I tried to install torch==2.1.0 and torchdata==0.7.0 while remaining dgl==2.2.1.

It was indeed useful for importing dgl, but when I tried to add edges to graph, a new error would rise: RuntimeError: Could not infer dtype of numpy.int64. Thus, the problem has been bothering me. It will be grateful if you have any new ideas and share them with me.

Best,
Bin

[rudongyu]

Have you tried with pip install dgl -f https://data.dgl.ai/wheels/torch-2.4/cu124/repo.html? (https://www.dgl.ai/pages/start.html) It contains wheels for dgl 2.4.0 based on pytorch 2.4

[DanielYang59]

@kenko911 Great to see you here. Looks like dgl has their own way to distribute bdist and sdist. I'm not a dgl user at all, but they only ship 2.2.1 wheels for windows only, 2.2.0 wheels for macos only and so on.

For the time being we might need to build from source ourselves, or perhaps we might need to store a pre-built one somewhere as a temporary solution (as the build is pretty costly).