services: ollama: image: ollama/ollama:latest container_name: kcg-ollama restart: unless-stopped ports: - "11434:11434" volumes: - /home/kcg-ollama/data:/root/.ollama deploy: resources: limits: memory: 64G reservations: memory: 40G environment: - OLLAMA_NUM_PARALLEL=4 - OLLAMA_MAX_LOADED_MODELS=1 - OLLAMA_KEEP_ALIVE=24h - OLLAMA_FLASH_ATTENTION=1 - OLLAMA_NUM_THREADS=48