version: "3.9"

services:
  backend:
    build:
      context: .
      dockerfile: Dockerfile
    ports:
      - "7860:7860"
      - "8001:8000"
    environment:
      - WORK_DIR=/tmp/elevnclip
      - VLLM_BASE_URL=http://localhost:8000/v1
      - VLLM_MODEL=Qwen/Qwen2.5-VL-7B-Instruct
      - VLLM_PORT=8000
      - VLLM_DOCKER_CONTAINER=
      - REMOTE_BACKEND_URL=
      - NEXT_PUBLIC_API_URL=http://localhost:7860
    volumes:
      - /tmp/elevnclip:/tmp/elevnclip
      - huggingface_cache:/root/.cache/huggingface
    # AMD GPU passthrough for ROCm
    devices:
      - /dev/kfd
      - /dev/dri
    group_add:
      - video
    security_opt:
      - seccomp:unconfined
    cap_add:
      - SYS_PTRACE
    restart: unless-stopped

volumes:
  huggingface_cache: