Microsoft DeepSpeed

Deep learning optimization library for training and serving massive models efficiently

AI Infrastructure & MLOpsFree

DeepSpeed is an open-source deep learning optimization library from Microsoft Research that enables training and inference of extremely large AI models (hundreds of billions of parameters) more efficiently through techniques including ZeRO memory optimization, pipeline parallelism, and kernel fusion. DeepSpeed dramatically reduces the hardware required to train large models and has been used to train models including Megatron-Turing NLG and BLOOM. AI research labs, enterprises training custom large models, and teams scaling up ML training use DeepSpeed to overcome GPU memory and compute constraints.

Key Features

• ✓ZeRO memory optimization
• ✓Pipeline parallelism
• ✓Large model training
• ✓Inference optimization
• ✓Open-source