services:
  llama-chat:
    build: .
    container_name: uncensored-llama
    # Run as root so we have permission to write the download to the host's folder
    user: root

    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]

    volumes:
      # 1. Mount the project directory
      - .:/app

      # 2. Mount the Cache
      # If the user HAS the model, it reads it.
      # If the user DOES NOT have the model, it downloads it here (saving it for next time).
      - ${HOME}/.cache/huggingface:/root/.cache/huggingface

    environment:
      # Tell the library where to look
      - HF_HOME=/root/.cache/huggingface
      # Clean visuals
      - TERM=xterm-256color
      # IMPORTANT: We REMOVED 'HF_HUB_OFFLINE=1'.
      # Now, if the model is missing, Unsloth is allowed to go fetch it.

    stdin_open: true
    tty: true
    ipc: host