DeepSpeed is a deep learning optimization library for PyTorch. It is designed to reduce computing power and memory use and to train large distributed models with better parallelism on existing computer hardware. DeepSpeed is also optimized for low latency, high throughput training and inference. It offers a suite of system innovations that make deep learning training and inference faster, cheaper and easier.