# yaml-language-server: $schema=https://cubeship.dev/schema/template/v1.json
version: 1
name: 'Ollama'
# The first release that keeps a volume's data across deploys.
minCubeship: "0.7.0"
project: ollama

apps:
  - key: server
    name: ollama
    image: ollama/ollama
    tag: "0.34.0"
    port: 11434
    health: /
    # No domain: Ollama's API has no authentication, and a public one would
    # let anyone run models on this machine and pull new ones onto its disk.
    # Other apps reach it at its internal address.
    volumes:
      - path: /root/.ollama
    # Ollama reads this memory limit, not the machine's, when it loads a model.
    limits: { cpu: 4, memory: 8Gi }
    env:
      # Without a GPU the default is three, and two models in memory at
      # once leave room for neither.
      OLLAMA_MAX_LOADED_MODELS: "1"
      # Fixed rather than guessed: every doubling costs memory, and on a
      # CPU, speed.
      OLLAMA_CONTEXT_LENGTH: "4096"