[ghm2199]

And Incidentally prefill would also be how caching,say, a system prompt saves you some $ for API usage with LLM providers. They only compute the kv cache for the new tokens after the system prompt.