AutoDDL, a distributed training framework, automatically explores and exploits new parallelization schemes with near-optimal bandwidth cost to facilitate the training of large-scale deep learning models.
