# Backendy LLM uruchamiane kontenerowo. Sterowane profilami, żeby jeden plik
# obsłużył kilka wariantów i żeby `docker compose up` bez profilu nic nie robił.
#
#   docker compose --profile vllm-gpu up -d      # Linux + NVIDIA
#   docker compose --profile ollama   up -d      # awaryjnie, obie platformy
#
# Zwykle nie wywołujesz tego wprost - robi to `task llm:up`.

name: agentic-codemod-llm

services:

  vllm:
    profiles: ["vllm-gpu"]
    image: vllm/vllm-openai:${VLLM_IMAGE_TAG:-latest}
    command: >
      --model ${LLM_SERVE_MODEL:-Qwen/Qwen3-4B-Instruct-2507}
      --max-model-len=16384
      --enable-auto-tool-choice
      --tool-call-parser=hermes
      --gpu-memory-utilization=${VLLM_GPU_UTIL:-0.90}
    ports:
      - "${LLM_PORT:-8000}:8000"
    volumes:
      # Cache modeli poza kontenerem - restart nie oznacza ponownego pobierania.
      - ${HF_CACHE:-${HOME}/.cache/huggingface}:/root/.cache/huggingface
    environment:
      HUGGING_FACE_HUB_TOKEN: ${HUGGING_FACE_HUB_TOKEN:-}
    # vLLM używa pamięci dzielonej do komunikacji między workerami
    ipc: host
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    healthcheck:
      test: ["CMD-SHELL", "python3 -c \"import urllib.request;urllib.request.urlopen('http://localhost:8000/health')\""]
      interval: 15s
      timeout: 5s
      retries: 40
      start_period: 300s
    restart: unless-stopped

  ollama:
    profiles: ["ollama"]
    image: ollama/ollama:${OLLAMA_IMAGE_TAG:-latest}
    ports:
      - "${OLLAMA_PORT:-11434}:11434"
    volumes:
      - ollama-models:/root/.ollama
    healthcheck:
      test: ["CMD-SHELL", "ollama list >/dev/null 2>&1"]
      interval: 10s
      timeout: 5s
      retries: 30
      start_period: 30s
    restart: unless-stopped

volumes:
  ollama-models: