DeepSpeed Chat

Microsoft has recently released DeepSpeed-Chat, a cost-effective, open-source tool which can be used to create advanced ChatGPT-style models. It is possible to train models of up to 13 billion parameters on a single GPU, or with a low-cost of $300 using Azure Cloud. DeepSpeed-Chat leverages DeepSpeed-Inference and DeepSpeed-Training to provide faster training, and supports model sizes that are up to 7.5x bigger than other solutions. In this way, DeepSpeed-Chat makes complex RLHF training accessible, simple, and affordable to the AI community.

The initial version of DeepSpeed-Chat includes three main features:

• Easy-to-use Training and Inference Experience for ChatGPT-like Models.
• DeepSpeed-RLHF Pipeline.
• DeepSpeed-RLHF System