services:
  vllm:
    image: vllm/vllm-openai:latest
    container_name: cortex-vllm
    restart: unless-stopped
    runtime: nvidia
    ipc: host
    ports:
      - "8005:8005"
    environment:
      - NVIDIA_VISIBLE_DEVICES=0,1
      - HUGGING_FACE_HUB_TOKEN=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
      - PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
    volumes:
      - /opt/vllm/huggingface:/root/.cache/huggingface
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 2
              capabilities: [gpu]
    command:
      - --model=Qwen/Qwen3-14B-AWQ
      - --port=8005
      - --quantization=awq
      - --tensor-parallel-size=2
      - --gpu-memory-utilization=0.82
      - --max-model-len=65536
      - --hf-overrides={"rope_scaling":{"rope_type":"yarn","factor":2.0,"original_max_position_embeddings":32768}}
      - --max-num-seqs=16
      - --kv-cache-dtype=fp8
      - --served-model-name=qwen3-14b-awq
      - --enable-auto-tool-choice
      - --tool-call-parser=hermes
      - --reasoning-parser=deepseek_r1
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:8005/health"]
      interval: 30s
      timeout: 10s
      retries: 5
      start_period: 180s
    networks:
      - cortex-net

  hermes-agent:
    image: nousresearch/hermes-agent:latest
    container_name: cortex-hermes-agent
    restart: unless-stopped
    depends_on:
      vllm:
        condition: service_healthy
    shm_size: 1g
    ports:
      - "9119:9119"
      - "8642:8642"
    volumes:
      - /opt/hermes:/opt/data
    environment:
      - HERMES_UID=${HERMES_UID:-10000}
      - HERMES_GID=${HERMES_GID:-10000}
      - OPENAI_BASE_URL=http://vllm:8005/v1
      - OPENAI_API_KEY=sk-dummy-key-1234567890
      - GATEWAY_HEALTH_URL=http://localhost:8642
    entrypoint: ["/bin/bash", "-c"]
    command:
      - |
        set -e
        hermes-agent gateway run &
        GATEWAY_PID=$$!
        until curl -sf http://localhost:8642 >/dev/null 2>&1; do
          if ! kill -0 $$GATEWAY_PID 2>/dev/null; then
            echo "FEHLER: Gateway-Prozess ist abgestürzt, breche ab."
            exit 1
          fi
          sleep 1
        done
        hermes-agent dashboard --host 0.0.0.0 --no-open &
        DASHBOARD_PID=$$!
        wait -n $$GATEWAY_PID $$DASHBOARD_PID
        EXIT_CODE=$$?
        kill $$GATEWAY_PID $$DASHBOARD_PID 2>/dev/null || true
        exit $$EXIT_CODE
    healthcheck:
      test: ["CMD-SHELL", "curl -f http://localhost:8642 && curl -f http://localhost:9119"]
      interval: 30s
      timeout: 10s
      retries: 5
      start_period: 60s
    networks:
      - cortex-net

networks:
  cortex-net:
    driver: bridge