Avoid growing prefill cost in long conversations?

[wjakob]

In longer converstations with a multimodel, the compute time goes up with each additional message. In the non-VLM mlx-llm project, it is possible to use resume a conversation from the last state without having an increasing prefill cost. Does functionality like this exist in mlx-vlm?