SpeedTorch

Very useful! Thanks! > Although PyTorch has offered a series of tutorials on distributed training, I found it insufficient or overwhelming to help the beginners to do state-of-the-art PyTorch distributed training Can confirm. Am relatively junior, spent way too long trying to understand the documentation. Tangentially related that bothers me: distributed training is is poorly documented, distributed memory management is barely documented and not fully supported. If your models are too large to fit in the memory of a single GPU (like really big word embeddings, which are not exotic by any stretch) things get wacky. Distributing model parts across different GPUs is supported but poorly documented and a bit hacky. If your models are too large for all of your GPUs you're stuck with an outside package like SpeedTorch (https://github.com/Santos