docs/DeepSpeed.md

An alternative way of reducing the GPU memory usage of models is to use the DeepSpeed ZeRO-3 optimization.

With this, I have been able to load a 6b model (GPT-J 6B) with less than 6GB of VRAM. The speed of text generation is very decent and much better than what would be accomplished with --auto-devices --gpu-memory 6.

As far as I know, DeepSpeed is only available for Linux at the moment.

How to use it

1. Install DeepSpeed:

conda install -c conda-forge mpi4py mpich
pip install -U deepspeed

2. Start the web UI replacing python with deepspeed --num_gpus=1 and adding the --deepspeed flag. Example:

deepspeed --num_gpus=1 server.py --deepspeed --chat --model gpt-j-6B

Learn more

For more information, check out this comment by 81300, who came up with the DeepSpeed support in this web UI.