Deploy LLM

This is a repository for a production-ready LLM deployment (in time)

Components

• LLM serving with vLLM
• Extracting and serving statistics with Prometheus
• Exposing GPU metrics with dcgm-exporter
• Displaying statistics with Grafana dashboard

Installation

To install all the requirements you need the following dependencies on your machine

• Cuda
• Nvidia drivers
• Docker
• Nvidia container toolkit

You can install all of these dependencies on a fresh ubuntu machine using ./install-script. Be careful running this command if you do not have a clean machine. Reboot after installing for the changes to take effect.

Configuration of secrets

You will want to change some settings before using this configuration in production, like changing the VLLM secret key and the grafana password. Configuration of secrets is handled in an .env file in the repository root. This file is ignored by git to ensure secrets are not accidentally committed. You can initalize a .env from our provided example by running

cp .env.example .env

Running

Start the deployment with docker compose up -d