services:
  # Embeddings run on a second vLLM instance rather than a separate toolchain
  # (e.g. text-embeddings-inference): whatever vLLM build works on SM120 then
  # covers embeddings too, instead of having to solve Blackwell support twice.
  embeddings:
    image: "{{ vllm_image }}"
    container_name: embeddings
    networks:
      - llmnet
    ports:
      - "8001:8000"
    volumes:
      - "{{ 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
      - "{{ embedding_model }}"
      - --served-model-name
      - "{{ embedding_model_name }}"
      - --task
      - embed
      # small fixed slice — the chat model gets the rest
      - --gpu-memory-utilization
      - "{{ embeddings_gpu_memory_utilization }}"
    ipc: host
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    runtime: nvidia
    restart: unless-stopped