Is there any way to simulate interactive mode behavior?

[Bibarius]

Dear community!
I have just recently started to get acquainted with llama.cpp and it is fantastic!
But there is one question that I have no way of understanding.
I am running llama.cpp built for cpu only on my local laptop. Accordingly, it takes a huge amount of time to run the inference with a given prompt. This behavior is quite understandable to me, since it is cpu built and I don't have the strongest laptop. But when I run interactive mode with prompt templates from the repo to simulate gpt-like chat, a kind of "cold start" takes just as long, but when further chatting in interactive mode, responses are generated just on the fly with almost no delay.

Hence I have two questions

1. How is this implemented? What is happening under the hood at this moment?
2. How to achieve this behavior without using interactive mode?

I was trying to figure this out by reading the documentation, and I thought it could be achieved by using "--prompt-cache" and "--prompt-cache-all". But as far as I understand, "--prompt-cache-all" will not be useful in chat simulation mode (the functionality I want to achieve), because after adding a custom prompt to the cached one, it will still re-generate it again, so I won't get as fast as in interactive mode.

I will be glad to get any help!