Add a way to clear the whole prompt cache (slots + host RAM)

[yuya0408]

I found that with the default settings, erasing all slots does not clear the host RAM cache. The next request loads the prompt back from it.

I want to clear the whole prompt cache of the server between benchmark runs, so that a run does not hit the cache of the previous run. Right now, the only ways to do this are: (1) restart the server, or (2) use --cache-ram 0 and erase all slots.

Both have a problem. With (1), the server reloads the model on every restart, so the runs take longer. With (2), I can only measure with the host RAM cache turned off, but it is on by default. So I measure a setup that is different from what users normally run.

vLLM (/reset_prefix_cache) and SGLang (/flush_cache) have an endpoint to clear the cache, and their benchmark tools use it between runs.

So I propose one of these: (a) a new endpoint, or (b) an option for the existing erase action that also clears the host RAM cache.

My measurements:

Version: 0.7.0-dev (build 11716, commit 797de63b2), official Windows CPU build
Model: Qwen/Qwen3-0.6B-GGUF, Q8_0
Server: llama-server -m Qwen3-0.6B-Q8_0.gguf -c 8192 --slot-save-path ./slots

P, Q: two different prompts, about 1,300 tokens each
Steps: P -> Q -> erase all slots (POST /slots/{id}?action=erase) -> P

timings.cache_n of the last P (prompt = 1318 tokens):

server options	cache_n
(default)	1317
--parallel 2	1317
--cache-ram 0	0
--parallel 2 --cache-ram 0	0

Without Q in between (P -> erase all slots -> P), cache_n = 0 in all four cases.