services:
  vllm:
    image: "{{ vllm_image }}"
    container_name: vllm
    networks:
      - llmnet
    ports:
      - "8000:8000"
    volumes:
      # model cache on local disk — tens of GB per model
      - "{{ appdata_path }}/models/huggingface:/root/.cache/huggingface"
    environment:
      - NVIDIA_VISIBLE_DEVICES=all
      - NVIDIA_DRIVER_CAPABILITIES=compute,utility
      - "HUGGING_FACE_HUB_TOKEN={{ hf_token | default('') }}"
    command:
      - --model
      - "{{ vllm_model }}"
      - --served-model-name
      - "{{ vllm_served_model_name }}"
      # fixed VRAM share so the embedding server keeps its slice (§0: "it just works")
      - --gpu-memory-utilization
      - "{{ vllm_gpu_memory_utilization }}"
      - --max-model-len
      - "{{ vllm_max_model_len }}"
    # vLLM needs a large shared-memory segment; without this it dies on startup
    ipc: host
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    runtime: nvidia
    restart: unless-stopped