services:
  unsloth-train:
    image: rocm/pytorch:latest
    container_name: unsloth-train
    restart: unless-stopped
    devices:
      - /dev/kfd:/dev/kfd
      - /dev/dri/renderD128:/dev/dri/renderD128
    group_add:
      - video
    ipc: host
    shm_size: 16g
    volumes:
      - unsloth-workspace:/workspace
      - unsloth-cache:/root/.cache/huggingface
    working_dir: /workspace
    stdin_open: true
    tty: true
    # keeps the container running in the background so you can exec in, start a long training run, and detach without killing it. Use tmux or nohup inside the container for runs that should outlive your SSH session.
    command: sleep infinity 
    labels:
      - homepage.group=Development
      - homepage.name=Unsloth
      - homepage.icon=unsloth.png
      - homepage.description=LLM Training
      - rackpeek.tags=ai
    networks:
      - ollama-net

networks:
  ollama-net:
    external: true

volumes:
  unsloth-workspace:
  unsloth-cache: