# GEPARD — Cartesia-compatible TTS server.
#
#   cp .env.example .env   # then fill in DATABASE_URL + HF_TOKEN
#   docker compose up      # builds locally, or set `image:` to pull from Docker Hub
#
# Requires the NVIDIA driver + NVIDIA Container Toolkit on the host.

services:
  gepard:
    # Build locally from the Dockerfile...
    build: .
    # ...or pull a published image instead (comment out `build:` above):
    # image: YOUR_DOCKERHUB_USER/gepard:latest
    ports:
      - "8000:8000"
    env_file:
      - .env
    volumes:
      # Persist the Hugging Face model download across runs.
      - hf-cache:/root/.cache/huggingface
    # Expose the GPU (needs a recent Docker Compose). On older Compose, replace
    # the line below with:
    #   deploy:
    #     resources:
    #       reservations:
    #         devices:
    #           - driver: nvidia
    #             count: all
    #             capabilities: [gpu]
    gpus: all
    # Run an init (tini) as PID 1 so SIGTERM is forwarded and the vLLM EngineCore
    # subprocess is reaped on stop/restart — otherwise it can survive and keep
    # ~80 GB of VRAM pinned, OOMing the next start (see README "Stopping").
    init: true
    # Model load + warmup takes a bit; don't kill it on a slow first boot.
    stop_grace_period: 30s
    restart: unless-stopped
    # Mirrors the image HEALTHCHECK; start_period covers first-boot model download
    # + compile so a slow cold start isn't flagged unhealthy.
    healthcheck:
      test: ["CMD", "curl", "-fsS", "http://localhost:8000/health"]
      interval: 30s
      timeout: 5s
      start_period: 10m
      retries: 3

volumes:
  hf-cache: