🐛[BUG]: physicsnemo Graphcast and GNN layers depend on now dead DGL

[djgagne]

Version

1.1.1

On which installation method(s) does this occur?

No response

Describe the issue

One of my team members @kanz76 has implemented a graph neural network architecture using the gnn_layers module from physicsnemo. That module depends on the Deep Graph Library (DGL), which had more robust functionality than pytorch-geometric and scaled more efficiently. The problem is that DGL appears to have stopped all development in September 2024 with no notice, and no updates have been made to the master branch since then along with no responses from developers to issues. DGL is not listed as a dependency (required or optional) in physicsnemo but is imported with no exception handling in gnn_layers. Are there any plans to fork or restart development of DGL or to migrate the gnn_layers to pytorch geometric, which is still being actively developed?

Minimum reproducible example

Relevant log output

Environment details

[coreyjadams]

Hi @djgagne - check out #995 from @Alexey-Kamenev. He has been developing PyG as an alternative backend to DGL for many of the gnn layers, which ought to enable the layers you need.

[coreyjadams]

Last week, #995 was merged into physicsnemo which enables PyG support in MeshGraphNet as well as the layers in gnn_layers. If you encounter specific problems, or need other enhancements/support, please open another issue.

To be clear:

• DGL will be maintained for some time in physicsnemo in "deprecation" status.
• PyG will be simultaneously supported during this time.
• Eventually, we will remove the DGL versions and maintain only PyG.

Thanks for bringing this up!