Open
Description
Is parallel GPU training support possible? We would like to try this with a fairly large (multi-GB) dataset, but to make training time reasonable it would need to be done in parallel. Single node parallelism with DataParallel() would probably work for our use case, although the PyTorch documentation suggests that DistributedDataParallel() is preferred even for a single node.
Part of the motivation for this is that a large dataset needs a lot of memory, which in a cloud environment means a large, multi-GPU instance. It is very expensive to run such a large instance for weeks with all but one of the GPUs idle.