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

software dependencies updated #209

Closed allaffa closed 9 months ago

jychoi-hpc commented 9 months ago

Can we try with the following version set? Those are the currently working set on Frontier: pytorch_scatter 2.1.2 pytorch_sparse 0.6.18 pytorch_cluster 1.6.3 pytorch_spline_conv 1.2.2 pytorch_geometric 2.5.0

allaffa commented 9 months ago

@jychoi-hpc @pzhanggit

I tried to find new pre-compiled wheels that automatically rely on binary files to combine PyTorch>=2.0 with PyG.

All the wheels can be found at this links: https://data.pyg.org/whl/

There are wheels that supposedly make torch-scatter-2.1.2 work with torch-2.1 and torch-2.2. However, when you try to install them, they all return the error “torch not found”, which seems to be a well-known problem https://github.com/rusty1s/pytorch_scatter/issues/365

The most update wheel that I could make work is the one that where torch-scatter uses torch<=1.13.

Please see PR#209 https://github.com/ORNL/HydraGNN/actions/runs/8031604060/job/21940095437?pr=209

I think that we need to maintain a very detail requirement.txt and reuirement-torchdep.txt where we clearly specify all the version of the packages.

If we say package >= X.y this may lead to individually install the latest version of a package, but that version may be incompatible with other packages.