docker-compose.gpu.yml

# =============================================================================
# Docker Compose — Stack 2.9 GPU Deployment
# =============================================================================
# Usage:
#   Start:  docker compose -f docker-compose.gpu.yml up --build -d
#   Logs:   docker compose -f docker-compose.gpu.yml logs -f
#   Stop:   docker compose -f docker-compose.gpu.yml down
#   Restart: docker compose -f docker-compose.gpu.yml restart
#
# Prerequisites:
#   1. NVIDIA Container Toolkit installed:
#        https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html
#   2. docker run --gpus all working on the host
#   3. Model files present at ./base_model_qwen7b (or path set below)
# =============================================================================

services:
  stack-2.9:
    build:
      context: .
      dockerfile: Dockerfile.gpu
      target: runtime
      args:
        UID: ${UID:-1000}
        GID: ${GID:-1000}

    image: stack-2.9-gpu:latest
    container_name: stack-2.9-api

    # ---------------------------------------------------------------------
    # GPU access — requires nvidia-container-toolkit on the host.
    # ---------------------------------------------------------------------
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all          # "1" for a specific GPU
              capabilities: [gpu]

    # ---------------------------------------------------------------------
    # Environment
    # ---------------------------------------------------------------------
    environment:
      - MODEL_PATH=/model
      - DEVICE=cuda
      - PORT=8000
      - HOST=0.0.0.0
      - CUDA_VISIBLE_DEVICES=0
      - PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512
      - TRANSFORMERS_CACHE=/model/.cache
      - HF_HOME=/model/.cache
      # Optional tuning — increase if you have ample GPU VRAM
      - DEFAULT_MAX_TOKENS=512
      - DEFAULT_TEMPERATURE=0.2
      - DEFAULT_TOP_P=0.95

    # ---------------------------------------------------------------------
    # Port mapping — REST API
    # ---------------------------------------------------------------------
    ports:
      - "${STACK_PORT:-8000}:8000"

    # ---------------------------------------------------------------------
    # Volume mounts
    # ---------------------------------------------------------------------
    volumes:
      # ── Model weights (read-only, essential) ──────────────────────────
      # Mount your fine-tuned or base Qwen-7b model directory here.
      # Example:  ./base_model_qwen7b  →  /model
      - ${MODEL_PATH:-./base_model_qwen7b}:/model:ro

      # ── HuggingFace cache (optional, speeds up rebuilds) ──────────────
      # Uncomment if you want to persist the HF hub cache:
      # - ./hf_cache:/model/.cache

      # ── Inference data / logs (optional) ───────────────────────────────
      # Mount a directory for additional prompt templates or static files:
      # - ./data:/data:ro

    # ---------------------------------------------------------------------
    # Restart policy
    # ---------------------------------------------------------------------
    restart: unless-stopped

    # ---------------------------------------------------------------------
    # Healthcheck (also defined in Dockerfile; repeated here for compose)
    # ---------------------------------------------------------------------
    healthcheck:
      test: ["CMD", "curl", "-sf", "http://localhost:8000/health"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 120s   # Model loading can take 60–90 seconds

    # ---------------------------------------------------------------------
    # Resource limits (tune to your GPU VRAM)
    # ---------------------------------------------------------------------
    # Uncomment and adjust if you want to cap resource usage:
    # mem_limit: 16g
    # shm_size: 4g

    # ---------------------------------------------------------------------
    # Logging
    # ---------------------------------------------------------------------
    logging:
      driver: json-file
      options:
        max-size: 50m
        max-file: "3"