serve: APIs serving GLM-5.3-Flash

CHANGELOG.md

@@ -5,6 +5,58 @@ and versions follow [Semantic Versioning](https://semver.org/). Rebuild the comp
 
 ## [Unreleased]
 
+### Added
+
+- `/tokenize` and `/detokenize` (also under `/v1`), as SGLang and vLLM serve them. A chat request is counted exactly as
+  `/v1/chat/completions` would send it to the engine. The router sends them to any live engine without touching its
+  routing state ([docs/API.md](docs/API.md)).
+- SGLang's native `/generate` (also PUT): `text` or `input_ids`, one prompt or a batch, SGLang's `sampling_params`,
+  streaming, output logprobs and request ids. The router routes it by its prompt or session
+  ([docs/API.md](docs/API.md#generate)).
+- SGLang's `/model_info`, `/server_info`, `/get_load`, `/health_generate`, `/abort_request` and `/flush_cache` (and
+  their `get_` names), on each engine and on the router, which asks one engine or all of them
+  ([docs/API.md](docs/API.md#info-and-admin-endpoints)).
+- `n` choices on `/v1/chat/completions` and `/v1/completions`, sharing the prompt through the prefix cache; a seed gives
+  choice i `seed + i`, as in vLLM ([docs/API.md](docs/API.md#sampling-extras)).
+- SGLang's `stop_regex`, and keeping the stop text with `include_stop_str_in_output` (vLLM) or `no_stop_trim` (SGLang).
+- `skip_special_tokens: false`, and SGLang's `matched_stop` on every finished choice.
+- `stop_regex` refuses a pattern whose search can take exponential or polynomial time, and matches within the last
+  2048 characters of output.
+- SGLang's `no_logs` on `/generate`: the request's record stays out of the production log.
+- The OpenAI Responses API, `POST /v1/responses`, stateless: reasoning, message and function-call items, and the
+  Responses stream events ([docs/API.md](docs/API.md#responses-api)).
+- The Anthropic Messages API, `POST /v1/messages` and `/v1/messages/count_tokens`: thinking, text and tool-use blocks,
+  Anthropic's stream events and error shape ([docs/API.md](docs/API.md#anthropic-messages-api)).
+- The router routes `/v1/responses` and `/v1/messages` by session and prefix as it routes chat, and sends
+  `/v1/messages/count_tokens` to any live engine.
+- Scheduling flags with SGLang's names, host only and per engine: `--schedule-policy` (`fcfs`, `lpm`, `lof`, `spf`),
+  priority scheduling with the request field `priority` and priority preemption, `--max-queued-requests` and
+  `--max-queued-tokens` (a full queue answers 503), and `--radix-eviction-policy` (`lru`, `lfu`, `slru`, `priority`,
+  `tlru`). The defaults schedule as before ([docs/CONFIGURATION.md](docs/CONFIGURATION.md#scheduling)).
+- `GET /v1/stats` adds the engine's live state and request rates, TTFT, decode-speed and end-to-end percentiles over
+  recent windows, and `GET /dashboard` draws them; the router serves both for each engine
+  ([docs/API.md](docs/API.md#statistics-and-the-dashboard)). The fields the router reads are unchanged.
+- `TRNRT_PROD_DIR`: an opt-in on-disk log of request records (counts and times only) and gauges.
+- `TRNRT_PROMPT_CACHE_MB`: an opt-in prompt token cache that builds a chat turn's prompt ids from the previous turn's,
+  with the same ids as a full encode or none.
+- Structured outputs through xgrammar, an optional dependency: `response_format` `json_object` and `json_schema`, vLLM's
+  `guided_*` and `structured_outputs`, and SGLang's `json_schema`, `regex` and `ebnf`. Also `min_tokens` and
+  `thinking_token_budget`. They need `--structured-outputs`, which adds graphs that take a token mask
+  ([docs/API.md](docs/API.md#structured-outputs-penalties-and-thinking-budgets)). Tested on CPU; not yet run on
+  Trainium2, and the check that the default graphs are unchanged is pending.
+- Repetition, frequency and presence penalties, with `--penalties` and `--penalty-buckets`, which add graphs that take
+  penalty inputs ([docs/CONFIGURATION.md](docs/CONFIGURATION.md#structured-outputs-penalties-and-thinking-budgets)).
+  Tested on CPU; not yet run on Trainium2.
+
+### Changed
+
+- The stop token that ends a request (EOS or a `stop_token_ids` entry) is no longer part of its text when the tokenizer
+  does not mark it special, as in vLLM and SGLang.
+- `min_new_tokens` (SGLang's name for `min_tokens`) and `spaces_between_special_tokens: false` are refused with 400;
+  they were ignored.
+- `service_tier` other than `"auto"` or `"default"` is refused with 400 on `/v1/chat/completions`, `/v1/completions`
+  and `/v1/responses`.
+
 ## [0.1.0] - 2026-11-08
 
 The first public release.