docker-compose.gpu.yaml

Rama dev · réplica de lectura del repositorio original.

# Docker Compose for GPU-Accelerated Worker Testing
#
# Usage:
#   # NVIDIA GPU testing
#   docker compose -f docker-compose.gpu.yaml --profile nvidia up --build
#
#   # Intel Arc GPU testing
#   docker compose -f docker-compose.gpu.yaml --profile intel up --build
#
#   # CPU-only testing (with GPU image)
#   docker compose -f docker-compose.gpu.yaml --profile cpu up --build
#
# Prerequisites:
#   - For NVIDIA: nvidia-container-toolkit installed
#   - For Intel: /dev/dri accessible
#   - VLOG_WORKER_API_KEY environment variable set

services:
  # NVIDIA GPU worker
  worker-nvidia:
    build:
      context: .
      dockerfile: Dockerfile.worker.gpu
    environment:
      - VLOG_WORKER_API_URL=${VLOG_WORKER_API_URL:-http://host.docker.internal:9002}
      - VLOG_WORKER_API_KEY=${VLOG_WORKER_API_KEY}
      - VLOG_WORKER_HEARTBEAT_INTERVAL=${VLOG_WORKER_HEARTBEAT_INTERVAL:-30}
      - VLOG_WORKER_POLL_INTERVAL=${VLOG_WORKER_POLL_INTERVAL:-10}
      - VLOG_HWACCEL_TYPE=nvidia
      - VLOG_HWACCEL_PREFERRED_CODEC=${VLOG_HWACCEL_PREFERRED_CODEC:-h264}
      - NVIDIA_VISIBLE_DEVICES=all
      - NVIDIA_DRIVER_CAPABILITIES=compute,video,utility
    extra_hosts:
      - "host.docker.internal:host-gateway"
    volumes:
      - worker-data-nvidia:/tmp/vlog-worker
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu, video, compute]
    restart: unless-stopped
    healthcheck:
      # Check the worker's /ready endpoint which validates FFmpeg and API connectivity
      test: ["CMD", "curl", "-f", "http://localhost:8080/ready"]
      interval: 30s
      timeout: 10s
      retries: 3
      # Allow time for worker to start and establish API connection
      start_period: 30s
    profiles:
      - nvidia

  # Intel Arc GPU worker
  worker-intel:
    build:
      context: .
      dockerfile: Dockerfile.worker.gpu
    environment:
      - VLOG_WORKER_API_URL=${VLOG_WORKER_API_URL:-http://host.docker.internal:9002}
      - VLOG_WORKER_API_KEY=${VLOG_WORKER_API_KEY}
      - VLOG_WORKER_HEARTBEAT_INTERVAL=${VLOG_WORKER_HEARTBEAT_INTERVAL:-30}
      - VLOG_WORKER_POLL_INTERVAL=${VLOG_WORKER_POLL_INTERVAL:-10}
      - VLOG_HWACCEL_TYPE=intel
      - VLOG_HWACCEL_PREFERRED_CODEC=${VLOG_HWACCEL_PREFERRED_CODEC:-av1}
      - LIBVA_DRIVER_NAME=iHD
    extra_hosts:
      - "host.docker.internal:host-gateway"
    volumes:
      - worker-data-intel:/tmp/vlog-worker
    devices:
      - /dev/dri:/dev/dri
    group_add:
      - video
      - render
    restart: unless-stopped
    healthcheck:
      # Check the worker's /ready endpoint which validates FFmpeg and API connectivity
      test: ["CMD", "curl", "-f", "http://localhost:8080/ready"]
      interval: 30s
      timeout: 10s
      retries: 3
      # Allow time for worker to start and establish API connection
      start_period: 30s
    profiles:
      - intel

  # CPU-only worker (using GPU image but forcing CPU encoding)
  worker-cpu:
    build:
      context: .
      dockerfile: Dockerfile.worker.gpu
    environment:
      - VLOG_WORKER_API_URL=${VLOG_WORKER_API_URL:-http://host.docker.internal:9002}
      - VLOG_WORKER_API_KEY=${VLOG_WORKER_API_KEY}
      - VLOG_WORKER_HEARTBEAT_INTERVAL=${VLOG_WORKER_HEARTBEAT_INTERVAL:-30}
      - VLOG_WORKER_POLL_INTERVAL=${VLOG_WORKER_POLL_INTERVAL:-10}
      - VLOG_HWACCEL_TYPE=none
    extra_hosts:
      - "host.docker.internal:host-gateway"
    volumes:
      - worker-data-cpu:/tmp/vlog-worker
    deploy:
      resources:
        limits:
          cpus: '4'
          memory: 8G
        reservations:
          cpus: '1'
          memory: 2G
    restart: unless-stopped
    healthcheck:
      # Check the worker's /ready endpoint which validates FFmpeg and API connectivity
      test: ["CMD", "curl", "-f", "http://localhost:8080/ready"]
      interval: 30s
      timeout: 10s
      retries: 3
      # Allow time for worker to start and establish API connection
      start_period: 30s
    profiles:
      - cpu

volumes:
  worker-data-nvidia:
  worker-data-intel:
  worker-data-cpu: