services:
  # serve a GGUF (llama.cpp's model file format) already on disk
  # instead of pulling one from Hugging Face. start.sh layers this
  # in when LLAMACPP_MODEL_FILE is set, it's only the mount,
  # docker/llamacpp-entrypoint.sh does the rest
  llamacpp:
    volumes:
      - "${LLAMACPP_MODELS_DIR}:/models:ro"