Working around Nvidia’s shortcomings to get a GPU inside Ollama