services:
  sd-server:
    image: sdcpp-cuda:5090
    build:
      context: ./sd-cpp
      dockerfile: docker/Dockerfile.cuda
      args:
        CUDA_VERSION: "12.8.0"
        UBUNTU_VERSION: "24.04"
        CUDA_ARCHITECTURES: "120"
        GGML_CUDA_FA_ALL_QUANTS: "ON"
    entrypoint: ""
    command: >
      /sd-server
      --diffusion-model /models/Qwen-Image-2.1-Turbo-Q8_0.gguf
      --llm /models/Qwen-Image-2.1-Turbo-Abliterated-Uncensored-Q8_0.gguf
      --vae /models/qwen_image_2.1_vae_bf16.safetensors
      --llm_vision /models/mmproj-Qwen-Image-2.1-Turbo-Abliterated-Uncensored-F16.gguf
      --diffusion-fa
      --listen-ip 0.0.0.0
      --listen-port 8080
    ports:
      - "8080:8080"
    volumes:
      - ./models:/models
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    restart: unless-stopped

  api-gateway:
    build:
      context: .
      dockerfile: gateway/Dockerfile
    ports:
      - "3000:3000"
    environment:
      SD_SERVER_URL: http://sd-server:8080
      GEMMA_URL: http://host.docker.internal:1234/v1
      OFFLOAD_WORKERS: ${OFFLOAD_WORKERS:-}
    volumes:
      - ./output:/app/output
    depends_on:
      - sd-server
    restart: unless-stopped

  # Optional encoder offload worker — uncomment to enable distributed text encoding.
  # Runs the same abliterated encoder GGUF via llama.cpp on a remote 8GB box,
  # returning hidden-state tensors over LAN. The gateway picks it up via OFFLOAD_WORKERS.
  # encoder-worker:
  #   image: ghcr.io/ggml-org/llama.cpp:latest
  #   command: >
  #     llama-server
  #     -m /models/Qwen-Image-2.1-Turbo-Abliterated-Uncensored-Q8_0.gguf
  #     --port 8081
  #     --no-webui
  #   ports:
  #     - "8081:8081"
  #   volumes:
  #     - ./models:/models
  #   deploy:
  #     resources:
  #       reservations:
  #         devices:
  #           - driver: nvidia
  #             count: 1
  #             capabilities: [gpu]
  #   restart: unless-stopped