Description
Pipeline parallelism is a technique employed in distributed deep learning that enhances model execution by dividing the model into distinct segments, or "stages." As large language models and other memory-intensive models become more common, pipeline parallelism has grown increasingly important for several key areas: - Executing large-scale training jobs. - Enhancing performance in bandwidth-limited clusters. - Supporting large model inference. In this talk, we will introduce the torch.distributed.pipelining package which provides users a seamless way of applying pipeline parallelism. We will demonstrate the following features: - Splitting of model code based on simple specification. - Support for pipeline schedules, including GPipe, 1F1B, Interleaved 1F1B and Looped BFS, and providing the infrastructure for writing customized schedules. - Composability with other PyTorch parallel techniques such as data parallel (DDP, FSDP) or tensor parallel. - Out of the box integration with Hugging Face models for efficient inference.