Skip to content
FindOpenSource

DeepSpeed

Deep learning optimization library from Microsoft for training and running very large models efficiently across multiple GPUs.

Overview

DeepSpeed implements memory- and communication-efficient techniques - like ZeRO optimizer state partitioning - that let teams train models far larger than would otherwise fit in available GPU memory, and speeds up training across multiple GPUs and machines. It integrates with PyTorch (also in this catalog) rather than replacing it.

DeepSpeed is a fit for teams training large models that push the limits of available hardware, a specialized optimization layer on top of PyTorch rather than a general-purpose framework - most teams reach for it specifically when standard single-GPU or naive multi-GPU training isn't enough.

Categories
AI & Machine Learning
Keywords
distributed-trainingdeep-learningmodel-optimization
Languages
Python
License
Apache-2.0

Spotted an error? Suggest an edit on GitHub.