Should I use 'spawn' method to start multi-processing?

Let’s continue discussions in DistributedDataParallel causes Dataloader workers to utilize GPU memory