Counting a prompt

POST /v1/tokenize renders a chat body exactly as /v1/chat/completions would — chat template, tool block, thinking controls, system block — and answers how many tokens it prefills, without submitting it: no lane, no GPU, no KV page. The count is the usage.prompt_tokens the same body reports when served, and max_model_len is the context the load runs (--model-max-context, or what the 27B’s load adapted it to, Adaptation), so one call says whether a body fits.

curl http://127.0.0.1:8000/v1/tokenize \
  -H 'Content-Type: application/json' \
  -d '{"messages":[{"role":"user","content":"Hello"}],"tools":[...],"return_token_ids":true}'
# {"count":1234,"max_model_len":262144,"token_ids":[...]}

Send messages or a raw prompt (tokenized with no template), never both. return_token_ids and return_text add the ids and the rendered prompt. Sampling fields are ignored, stream is refused, and a body with an image_url part is a 400 media_not_countable: an image’s cost is its grid after preparation, which only fetching and decoding it would tell — send the request. POST /v1/detokenize takes {"token_ids":[...]} and answers {"text":"..."}; an id outside the vocabulary is a 400 naming its index. The tokenizer normalizes text to NFC, so a decomposed e + accent comes back as the composed é (the text a call returns is the rendered prompt as sent, before that). Neither route enters the scheduler; the count of calls is ignis_tokenize_requests_total{route} on the metrics listener.