A newer version of the Gradio SDK is available:
5.9.1
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
- Install DeepSpeed:
conda install -c conda-forge mpi4py mpich
pip install -U deepspeed
- Start the web UI replacing
pythonwithdeepspeed --num_gpus=1and adding the--deepspeedflag. 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.