services:
  nemotron-parse:
    build: .
    image: nemotron-parse:latest
    working_dir: /workspace

    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

    volumes:
      # Project source — live-edits on the host are reflected immediately.
      - .:/workspace
      # Shadow the host .venv (wrong Python version) so uv uses /opt/venv inside
      # the container rather than the host-side virtual environment.
      - /workspace/.venv
      # HuggingFace model cache — persists the RADIO encoder weights across
      # container restarts so they are not re-downloaded on every run.
      - hf-cache:/root/.cache/huggingface

    environment:
      HF_HOME: /root/.cache/huggingface

    # Keep the container alive for interactive use (exec, attach, etc.).
    stdin_open: true
    tty: true

volumes:
  hf-cache: