# High-throughput OCR with MPS multi-instance scaling.
# Requires NVIDIA MPS: sudo nvidia-cuda-mps-control -d
#
# TRT engines are auto-built from ONNX on first startup by the C++ binary.
# Cached in the shared volume — built once, used by all instances.
#
# Usage:
#   docker compose up -d                # Start 7 instances
#   docker compose up -d --scale ocr=5  # Start 5 instances
#   docker compose down                 # Stop all

volumes:
  trt-cache:  # Shared TRT engine cache

services:
  ocr:
    build:
      context: ..
      dockerfile: docker/Dockerfile.gpu
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
      replicas: 7
    # See README.md#configuration for all tunables (OCR_LANG, MAX_BODY_MB,
    # MAX_IMAGE_DIM, DET_MAX_SIDE, etc.).
    environment:
      - PIPELINE_POOL_SIZE=3
    ports:
      - "8001-8007:8000"
    volumes:
      - trt-cache:/home/ocr/.cache/turbo-ocr
    restart: unless-stopped
    healthcheck:
      test: ["CMD", "curl", "-sf", "http://localhost:8000/health"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 180s

  # gRPC mode (alternative)
  ocr-grpc:
    build:
      context: ..
      dockerfile: docker/Dockerfile.gpu
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    environment:
      - PIPELINE_POOL_SIZE=5
    ports:
      - "50051:50051"
    volumes:
      - trt-cache:/home/ocr/.cache/turbo-ocr
    command: ["./build/paddle_grpc_server"]
    restart: unless-stopped
    profiles:
      - grpc