How to preallocate KV cache for parallel requests in llama-server?

[lostmsu]

Right now I use gpt-oss and

--ctx-size 0
--kv-unified
--parallel 8

That results in llama-server allocating exactly 128K KV slots (the context size of gpt-oss). But these are going to be shared between the parallel requests. How can I instead preallocate 256K KV slots to be shared by up to 8 parallel requests, but limit each request to 128K?