docker-compose.yml

version: "3.9"

volumes:
  models_cache:
  outputs:
  storage:

services:
  redis:
    image: redis:7-alpine
    ports:
      - "6379:6379"
    restart: unless-stopped
    healthcheck:
      test: ["CMD", "redis-cli", "ping"]
      interval: 10s
      timeout: 5s
      retries: 5

  worker:
    build: .
    # #20 — consume all three queues; KD runs isolated from SFT so an OOM doesn't kill SFT jobs
    command: celery -A workers.celery_app worker --loglevel=info --concurrency=1 --queues sft,dpo,kd
    volumes:
      - models_cache:/app/models_cache
      - outputs:/app/outputs
      - storage:/app/storage
    environment:
      - REDIS_URL=redis://redis:6379/0
      - HF_HOME=/app/models_cache
    depends_on:
      redis:
        condition: service_healthy
    # #21 — cap memory so a runaway rank-256 LoRA job can't OOM-kill the container
    mem_limit: 20g
    memswap_limit: 20g
    restart: on-failure
    healthcheck:
      test: ["CMD", "celery", "-A", "workers.celery_app", "inspect", "ping", "--timeout=5"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 30s
    deploy:
      resources:
        reservations:
          devices:
            - capabilities: [gpu]   # remove if no GPU

  app:
    build: .
    command: reflex run --env prod
    ports:
      - "3000:3000"
      - "8000:8000"
    volumes:
      - outputs:/app/outputs
      - storage:/app/storage
    environment:
      - REDIS_URL=redis://redis:6379/0
    depends_on:
      redis:
        condition: service_healthy
      worker:
        condition: service_healthy
    restart: unless-stopped
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:8000/api/health"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 60s