services:
  diffusiongemma-26b-a4b-nvfp4:
    build:
      context: .
      dockerfile: docker/vllm/Dockerfile
      args:
        VLLM_IMAGE: "${VLLM_IMAGE:-vllm/vllm-openai:gemma}"
    image: "${IMAGE_NAME:-diffusiongemma-26b-a4b-nvfp4-vllm:latest}"
    container_name: "${CONTAINER_NAME:-diffusiongemma-26b-a4b-nvfp4}"
    restart: "${RESTART_POLICY:-unless-stopped}"
    network_mode: host
    ipc: host
    shm_size: "32gb"
    mem_limit: "${MEM_LIMIT:-112g}"
    memswap_limit: "${MEMSWAP_LIMIT:-112g}"
    env_file:
      - .env.example
      - path: .env
        required: false
    environment:
      NVIDIA_VISIBLE_DEVICES: all
      NVIDIA_DRIVER_CAPABILITIES: compute,utility
      VLLM_USE_V2_MODEL_RUNNER: "${VLLM_USE_V2_MODEL_RUNNER:-1}"
    volumes:
      - ./models/hf:/root/.cache/huggingface
      - ./logs:/logs
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]