TTL settings don't seem to have any effect

[stevelittlefish]

I have the following in my docker compose file:

    environment:
      - CHAT_COMPLETION_BASE_URL=http://192.168.11.27:11434/v1
      - CHAT_COMPLETION_API_KEY=xxx
      # Keep models in memory forever to stop Ollama from hogging all of the VRAM
      - STT_MODEL_TTL=-1
      - TTS_MODEL_TTL=-1

If I make a request to the API, I see the GPU VRAM usage go back to zero a few minutes after the request.