itod/mistral-7B-instruct-v0.2-q8

• Source Mistral 7B model:
  https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2/

• This model is converted from Bfloat16 datatype to Int8 datatype with convert tool from:
  https://github.com/ggerganov/llama.cpp

• Deployment on CPU:
  Pull the ready-made llama.cpp container:

docker pull ghcr.io/ggerganov/llama.cpp:server

Assuming mistral-7B-instruct-v0.2-q8.gguf file is downloaded to /path/to/models directory on local machine, run the container accesing the model with:

docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server -m /models/istral-7B-instruct-v0.2-q8.gguf --port 8000 --host 0.0.0.0 -n 512

• Test the deployment accessing the model with the browser at http://localhost:8000
• llama.cpp server also provides OpenAI compatible API
• Deployment on CUDA GPU:
  

docker pull ghcr.io/ggerganov/llama.cpp:server-cuda

docker run --gpus all -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server-cuda -m /models/mistral-7B-instruct-v0.2-q8.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 50

• If CUDA GPU with 16GB RAM is available, the version of the model converted to float16 may be interesting, available in this repo:
  https://huggingface.co/itod/mistral-7B-instruct-v0.2-f16
• More details about usage is avalable in llama.cpp documentation:
  https://github.com/ggerganov/llama.cpp/tree/master/examples/server