Pytorch Lightning: Support of DDPStrategy in PyG

[ZikangZhou]

Hi PyG team,

I'm trying to use PyTorch Lightning for multi-gpu training. The data in my task are multiple small graphs, not a single large graph.
I looked at the examples of PyG + PyTorch Lightning and found that all examples use DDPSpawnStrategy instead of DDPStrategy. I'm wondering whether I can switch to DDPStrategy without modifying my code in my case?

[rusty1s]

You should be able to use DDPStrategy but keep in mind that each GPU process will create a replica of the dataset in CPU memory. Therefore, we generally recommend to use DDPSpawnStrategy to be able to leverage shared memory across processes. Another alternative is to not make use of InMemoryDataset and only load in the graphs from disk that are actually used within a mini-batch (I/O might be a bottleneck here).