# Rollen "gpu-embed" und "gpu-small" (VMs mit Quadro P2000 / M4000): llama.cpp-Server.
# Maxwell (sm_52) und Pascal (sm_61) brauchen CUDA 12.x – daher eigener Build mit festen Architekturen.
# Profile: embed (Embedding + Rerank), chat (Verifier/Signatur), z. B. `docker compose --profile embed up -d`
name: knaster-gpu-llamacpp

x-llamacpp: &llamacpp
  build:
    context: https://github.com/ggml-org/llama.cpp.git#${LLAMACPP_REF:-b6000}
    dockerfile: .devops/cuda.Dockerfile
    target: server
    args:
      CUDA_VERSION: ${CUDA_VERSION:-12.4.0}
      CUDA_DOCKER_ARCH: ${CUDA_DOCKER_ARCH:-52;61}
  image: knaster/llamacpp-server:${LLAMACPP_REF:-b6000}-cu12
  restart: unless-stopped
  labels: {autoheal: "true"}
  volumes:
    - ${MODELS_DIR:-/srv/models}:/models:ro
  deploy:
    resources:
      reservations:
        devices: [{driver: nvidia, count: 1, capabilities: [gpu]}]

services:
  llamacpp-embed:
    <<: *llamacpp
    container_name: llamacpp-embed
    profiles: ["embed"]
    command: ["-m", "/models/${EMBED_GGUF:-Qwen3-Embedding-0.6B-Q8_0.gguf}", "--embedding",
              "--pooling", "${EMBED_POOLING:-last}", "-c", "8192", "-ub", "8192", "-ngl", "99",
              "--alias", "embed-default", "--metrics", "--host", "0.0.0.0", "--port", "8081"]
    ports: ["8081:8081"]
    healthcheck: &hc
      test: ["CMD", "curl", "-fsS", "http://127.0.0.1:8081/health"]
      interval: 30s
      timeout: 5s
      retries: 3
      start_period: 60s

  llamacpp-rerank:
    <<: *llamacpp
    container_name: llamacpp-rerank
    profiles: ["embed"]
    command: ["-m", "/models/${RERANK_GGUF:-bge-reranker-v2-m3-Q8_0.gguf}", "--reranking", "-c", "8192",
              "-ub", "8192", "-ngl", "99", "--alias", "rerank-default", "--metrics", "--host", "0.0.0.0",
              "--port", "8082"]
    ports: ["8082:8082"]
    healthcheck:
      <<: *hc
      test: ["CMD", "curl", "-fsS", "http://127.0.0.1:8082/health"]

  llamacpp-chat:
    <<: *llamacpp
    container_name: llamacpp-chat
    profiles: ["chat"]
    command: ["-m", "/models/${CHAT_GGUF:-Qwen3-4B-Q4_K_M.gguf}", "-c", "8192", "-ngl", "99",
              "--alias", "verifier-small", "--metrics", "--host", "0.0.0.0", "--port", "8080"]
    ports: ["8080:8080"]
    healthcheck:
      <<: *hc
      test: ["CMD", "curl", "-fsS", "http://127.0.0.1:8080/health"]

  gpu-exporter:
    image: utkuozdemir/nvidia_gpu_exporter:1.2.1
    restart: unless-stopped
    ports: ["9835:9835"]
    deploy:
      resources:
        reservations:
          devices: [{driver: nvidia, count: all, capabilities: [gpu, utility]}]

  node-exporter:
    image: prom/node-exporter:v1.8.1
    restart: unless-stopped
    command: ["--path.rootfs=/host", "--collector.textfile.directory=/textfile"]
    pid: host
    volumes:
      - /:/host:ro,rslave
      - /var/lib/node_exporter/textfile:/textfile:ro
    ports: ["9100:9100"]

  autoheal:
    image: willfarrell/autoheal:1.2.0
    restart: unless-stopped
    environment: {AUTOHEAL_CONTAINER_LABEL: autoheal}
    volumes: ["/var/run/docker.sock:/var/run/docker.sock"]