Data-parallel distributed training of very large models beyond GPU capacity
2018·Arxiv