One admission, one spelling

An admission is the kernel's hold, written the same way in a session and in a server:

hold reqs (1), kv (min(known, hit + budget_left(engine)))
     at admission (known = computed < prompt ? prompt : computed + 1,
                   hit = min(cachedin(kv), reusable(known, blocksize(kv)))) {
  …
} cache (prompt + o);

(lib/vllm.sq.) There were three spellings of this one CStmt::Hold: hold … at admission (…) … cache, enter … at admission (…) … keep in a session, and admit if … fit where … keep in a server. #136 kept the first. The IR did not move: every program's IR is the one it had.

The serving vocabulary a reader from vLLM looks for is now written by the program, as a name: vllm_request is a def in a library, and the call is what the server block says. The language keeps the mechanism (hold, at admission, reserve, reuse, cache, lease) and the program the policy (criterion 2).

The words enter, keep, fit, where and admit as a statement are refused with the spelling that replaced them, and none of them may name a definition, a parameter, an attribute or a binding, so that no name means the old statement in one place and the program's thing in another. admit via is the pool option and is unchanged.

Self-critique

The subject of the sentence. The philosophy argued that a keyword's subject should be its block's: the scheduler admits, so admit if in a server; the session enters, so enter in a session. That argument chose a word per side for one construct, which is two spellings of one meaning, and it put a serving word in the language where a def puts it in the program. The subject survives where it separates constructs (turn, end and request are the session's and are refused in a server); it no longer names one construct twice.

hold is a resource word. A serving engineer reads "hold a slot and the blocks" less readily than "admit". The answer is the definition's name, vllm_request, not a keyword: the name can say vLLM, a keyword cannot say every engine.