Closed puririshi98 closed 7 months ago
@rusty1s anything needed to get this merged? works great for us at nvidia
the main change was to stop using PyTorch Lightning which was not working for us for multigpu. Instead i updated the example to use vanilla PYG neighborloading and it works great. the original example did not work for multigpu at all on any NVIDIA system we tested and therefore should be replaced by my version ASAP. this version is tested to be working well on many NVIDIA systems
moved any changes to the training scripts into PyG itself. lets just add the pyg_to_hetero_data with this PR