DeepSpeed

DeepSpeed is a user-friendly, all-in-one package for deep learning optimization. It makes it possible to run Deep Learning Training and Inference on massive models with billions or trillions of parameters. It also offers superior system throughput, scalability to thousands of GPUs, and a low latency and high throughput for inference. Additionally, DeepSpeed provides extreme compression for model size reduction and cost reduction.

DeepSpeed is an integral part of Microsoft’s AI at Scale initiative, which intends to bring next-generation AI capabilities to the forefront. To learn more about this project, you can visit their website here.