DataLoader vs DataListLoader

[mdanb]

The docs mention that DataListLoader "should be used for multi-gpu support via torch_geometric.nn.DataParallel." Why exactly is that?

[rusty1s]

Yes, this is a restriction since the DataParallel module will take care of splitting the batches to multiple GPUs, and therefore needs access to each individual example. A more elegant way is provided via DistributedDataParallel, where you can simply set up a data loader for each process/GPU, see here.