Docker-compose for Ollama / Open WebUI / ComfyUI

# Docker Compose stack for local AI services:
# - `ollama` serves local LLMs
# - `open-webui` provides a browser UI for Ollama
# - `comfyui` provides image-generation workflows
services:
  ollama:
    volumes:
      # Persist downloaded models and Ollama state between restarts.
      - ollama:/root/.ollama
    container_name: ollama
    pull_policy: always
    tty: true
    restart: unless-stopped
    image: ollama/ollama:${OLLAMA_DOCKER_TAG-latest}
    environment:
      # Keep models loaded briefly to reduce reload latency.
      - OLLAMA_KEEP_ALIVE=5m
      - OLLAMA_FLASH_ATTENTION=1
      # Large context window; increase only if you have enough VRAM.
      - OLLAMA_CONTEXT_LENGTH=128000
      # Limit Ollama to the listed GPU devices.
      - CUDA_VISIBLE_DEVICES=0,1
      # Keep concurrency conservative to avoid VRAM exhaustion.
      - OLLAMA_MAX_LOADED_MODELS=1
      - OLLAMA_NUM_PARALLEL=1
    ports:
      # Ollama API endpoint.
      - 11434:11434
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              # Reserve both GPUs for this container.
              count: 2
              capabilities:
                - gpu

  open-webui:
    build:
      context: .
      args:
        OLLAMA_BASE_URL: "/ollama"
      dockerfile: Dockerfile
    image: ghcr.io/open-webui/open-webui:${WEBUI_DOCKER_TAG-main}
    container_name: open-webui
    pull_policy: always
    volumes:
      # Persist Open WebUI data (users, settings, chats).
      - open-webui:/app/backend/data
      # Mount TLS certificate/key from the host.
      - /etc/ssl/openui:/ssl
    depends_on:
      - ollama
    ports:
      # Host port is configurable with OPEN_WEBUI_PORT; container listens on 8080.
      - ${OPEN_WEBUI_PORT-3001}:8080
    environment:
      - "OLLAMA_BASE_URL=http://ollama:11434"
      - "WEBUI_URL=https://openui.local"
      # Set this to a strong non-empty value in production.
      - "WEBUI_SECRET_KEY="
      - "WEBUI_SSL_CERT=/ssl/openui.crt"
      - "WEBUI_SSL_KEY=/ssl/openui.key"
      # Wide-open CORS is convenient locally but risky outside trusted networks.
      - "CORS_ALLOW_ORIGIN=*"
    extra_hosts:
      - host.docker.internal:host-gateway
    restart: unless-stopped

  comfyui:
    # Prebuilt ComfyUI image with CUDA 12.8 / Torch 2.8 support.
    image: smyshnikof/comfyui:full-torch2.8.0-cu128
    container_name: comfyui
    environment:
      - NVIDIA_VISIBLE_DEVICES=all
      #- INSTALL_SAGEATTENTION=True  #sageattention is a custom node for comfyui that allows you to use the sage attention mechanism
    ports:
      # Main ComfyUI web interface.
      - "8188:3000"
      # Additional forwarded ports for custom nodes or side services.
      - "8081:8081"
      - "8082:8082"
      - "8083:8083"
      - "8888:8888"
    volumes:
      # Bind local model, output, and user data directories into the container.
      - ./comfyui/models:/workspace/ComfyUI/models
      - ./comfyui/output:/workspace/ComfyUI/output
      - ./comfyui/user:/workspace/ComfyUI/user/default
      # Preset file used by the Preset Download Manager custom node.
      - ./comfyui/preset-download-manager-presets.json:/workspace/ComfyUI/custom_nodes/ComfyUI-PresetDownloadManager/presets.json
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              # Reserve both GPUs for ComfyUI workloads.
              count: 2
              capabilities:
                - gpu
    restart: unless-stopped

volumes:
  # Named volumes keep app data across container recreation.
  ollama: {}
  open-webui: {}

mkdir -p comfyui/models comfyui/output comfyui/user comfyui/flows