search

ORNL / HydraGNN

Distributed PyTorch implementation of multi-headed graph convolutional neural networks
BSD 3-Clause "New" or "Revised" License
68 stars 29 forks source link

Minimum Requirement for Scatter #182

Closed JustinBakerMath closed 1 year ago

JustinBakerMath commented 1 year ago

Automated CI fails because torch_sparse dropped their support for torch==1.10 (see v6.13 vs v6.17) in favor of torch.sparse_csc_tensor introduced in torch==1.13.

It should be possible to version control torch_sparse<=6.13 instead, but the torch_geometric, torch_scatter, ... bundle can be trickier to control.

JustinBakerMath commented 1 year ago

I need to make sure that this dependency requirement does not cause problems on the supercomputers, where unfortunately the software stack is not alway the most up-to-date

I completely understand. Let me know if going forward it makes sense for me to add all three of you as reviewers whenever I make version related PRs.

The brute force alternative to this PR is to hard code scatter when I need it for DimeNet.