docker-compose.nvidia.yml

# NVIDIA GPU worker overlay. Requires a Linux host with an NVIDIA GPU and the
# nvidia-container-toolkit installed and configured for Docker.
#
# Run alongside the base stack:
#   docker compose -f docker-compose.yml -f docker-compose.nvidia.yml up -d --build
#
# Verify GPU access first:  docker run --rm --gpus all nvidia/cuda:12.4.1-base nvidia-smi
#
# This overlay only adds the GPU worker. The datastore/API port hardening and
# reverse-proxy guidance live in docker-compose.yml (see its PRODUCTION NOTES);
# they apply unchanged here. The worker is compute-heavy — we do NOT cap its
# CPU/memory (that would throttle OCR); the GPU is granted via the reservations
# block below. Set an explicit mem limit only if the host is memory-constrained.

services:
  nvidia-worker:
    image: ghcr.io/dim145/subtitleextractor-worker-nvidia:${IMAGE_TAG:-latest}
    hostname: nvidia-worker
    container_name: nvidia-worker
    build:
      context: ./worker
      dockerfile: Dockerfile.nvidia
    restart: unless-stopped
    environment:
      API_BASE_URL: http://api:8080
      INTERNAL_API_TOKEN: ${INTERNAL_API_TOKEN}
      WORKER_CLASS: gpu-nvidia
      PPOCR_USE_GPU: "1"
      WORKER_DECODER: ffmpeg
      WORKER_HWACCEL: cuda
      NVIDIA_VISIBLE_DEVICES: all
      NVIDIA_DRIVER_CAPABILITIES: compute,utility,video
    depends_on:
      api:
        condition: service_healthy
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]