# docker-compose.ollama.yml
#
# Ollama opt-in overlay — local LLM inference via Ollama.
# Can be used independently of vLLM.
#
# USAGE:
#   pdavid --mode up --ollama
#
# Or manually:
#   docker compose -f docker-compose.yml -f docker-compose.ollama.yml up -d
#
# REQUIREMENTS:
#   - NVIDIA GPU with drivers installed
#   - NVIDIA Container Toolkit (https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html)
#
# IMAGE PINNING
# ─────────────
# Pinned: 2026-04-04 | ollama: 0.20.0 | See PINNED_IMAGES.md
# NOTE: ollama digest is also present in docker-compose.gpu.yml.
#       Update both files when bumping version.

services:
  ollama:
    # 0.20.0 — upgrade tag to ollama/ollama:0.20.0 on next bump
    image: ollama/ollama@sha256:0455f166da85b1d07f694c33ba09278ca649603c0611ba8e46272b16eed7fccd
    container_name: ollama
    restart: unless-stopped
    runtime: nvidia
    environment:
      - NVIDIA_VISIBLE_DEVICES=all
    volumes:
      - ollama_data:/root/.ollama
    ports:
      - "11434:11434"
    networks:
      - my_custom_network
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

  # Override api to point at the real Ollama internal URL
  # instead of the localhost stub in the base compose file.
  api:
    environment:
      - OLLAMA_BASE_URL=http://ollama:11434/v1
    depends_on:
      ollama:
        condition: service_started

volumes:
  ollama_data: