Powering ML Services with Ollama: A Complete Setup for Self-Hosted LLMs

Our ML Services need LLMs to process the large documents and data. We started using Ollama since 2024 and the learnings below:

Ollama, an open-source tool that packages LLMs into a simple CLI and REST server, has emerged as a game-changer for developers and organizations seeking privacy, cost-efficiency, and offline capabilities

*What is Ollama and Why Use It? *

Ollama is often described as "Docker for LLMs" – you pull a model by name, and Ollama handles quantization, memory mapping, and inference runtime behind the scenes. The core reasons developers choose Ollama over cloud APIs include:

• Privacy: Your prompts and data never leave your machine
• Cost: Zero per-token charges after the one-time model download
• Latency: No network round-trip, especially fast on modern GPUs
• Offline use: Works without any internet connection once models are downloaded

Our ML Services Environment

Our setup runs a variety of models optimized for different tasks, including:

• Qwen3 series (8B, 4B-instruct, 0.6B) for general-purpose inference
• Llama3.2 and custom variants for specialized tasks
• Gemma3:4b and Gemma2:2b for lightweight processing
• BGE-M3 for embedding generation
• llm – a custom model created from GGUF format

We also maintain two dedicated services:

• ner: For Named Entity Recognition in forms
• summarizer-llm: For summarizing signals and communications

Complete Setup Walkthrough

1. Install Dependencies

Start by updating your system and installing essential packages:

apt update && apt upgrade -y
apt install git docker-compose curl

1. Install NVIDIA Drivers and Container Toolkit

For GPU acceleration (which provides 10-100x speedup over CPU inference) , install NVIDIA drivers and the NVIDIA Container Toolkit:

distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add -
curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | sudo tee /etc/apt/sources.list.d/nvidia-docker.list

sudo apt-get install -y nvidia-docker2
sudo systemctl restart docker
sudo ubuntu-drivers autoinstall

Reboot the system after driver installation.

1. Clone and Configure the Ollama Repository

git clone https://github.com/sujithrpillai/ollama.git
cd ollama

Create a docker-compose.yml file with GPU support:

version: '3.8'

services:
  ollama:
    image: ollama/ollama:latest
    hostname: ollama
    ports:
      - "11434:11434"
    volumes:
      - ./models:/root/.ollama/models
    networks:
      - genai-network
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    runtime: nvidia
    restart: always

  open-webui:
    image: ghcr.io/open-webui/open-webui:main
    ports:
      - "3000:8080"
    environment:
      - OLLAMA_BASE_URL=http://ollama:11434
    volumes:
      - ./backend/data:/app/backend/data
    networks:
      - genai-network
    restart: always

networks:
  genai-network:
    driver: bridge
    name: genai-network

The GPU configuration in the deploy section enables hardware acceleration for supported NVIDIA GPUs .

1. Start the Services

docker-compose up -d

1. Pull Models

Access the Ollama container and download models:

docker exec -it ollama ollama pull qwen3:8b
docker exec -it ollama ollama pull gemma2:2b
docker exec -it ollama ollama pull llama3.2:latest

1. Create Custom Models from GGUF Files

Ollama allows creating custom models from GGUF format files :

sudo apt-get install git-lfs
git clone https://huggingface.co/org/llm-v3-270k-GGUF
curl -fsSL https://ollama.com/install.sh | sh

Create a Modelfile:

FROM /home/RatnaDveLinga-v3-270k-GGUF/RatnaDveLinga-v3-270k.gguf

Then create the model:

ollama create llmname -f Modelfile

1. Verify Setup

Check models available:

root@ollama:/ ollama list

NAME                                ID              SIZE      MODIFIED
qwen3:8b                            500a1f067a9f    5.2 GB    8 months ago
qwen3:4b-instruct                   0edcdef34593    2.5 GB    8 months ago
qwen2.5:1.5b                        65ec06548149    986 MB    9 months ago
gemma3:4b                           a2af6cc3eb7f    3.3 GB    13 months ago
llama3.2:latest                     a80c4f17acd5    2.0 GB    13 months ago
bge-m3:latest                       790764642607    1.2 GB    13 months ago
RatnaDveLinga:latest                decc819a53b1    1.7 GB    21 months ago

Model Management Commands

Essential CLI commands for managing your Ollama setup :

Command	Description
ollama list	List all downloaded models
ollama pull <model>	Download a model without running it
ollama run <model>	Run a model interactively
ollama stop <model>	Stop a running model
ollama rm <model>	Remove a model from disk
ollama ps	Show currently running models
ollama show <model>	Display model metadata

Performance Considerations

GPU acceleration can provide 10-100x faster inference compared to CPU-only operation . To verify GPU usage:

docker exec -it ollama nvidia-smi

The Open WebUI interface available at http://localhost:3000 provides a user-friendly way to interact with models, pull new ones, and manage your setup without using the command line .

Conclusion

This setup provides a robust, self-hosted LLM environment with GPU acceleration, a comprehensive model library, and a web interface for easy interaction. Whether you're running specialized NER services, summarization tasks, or general-purpose inference, Ollama combined with Open WebUI delivers a powerful and private AI platform.

Happy prompting!