Skip to content

spec(SAMPLE-REASONING): W3 runs before W2, on measured recipe demand (#605) #616

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
localai-bot merged 1 commit into from
Aug 13, 2026
Merged

spec(SAMPLE-REASONING): W3 runs before W2, on measured recipe demand (#605) #616

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
39 changes: 39 additions & 0 deletions .agents/specs/reasoning-parsers.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -176,6 +176,45 @@ RED-first.
- **W3** — the engine-backed reasoning adapters (`qwen3`/`mimo`, `gemma4`,
`glm45/47`, `seed_oss`, `deepseek_v4`, `nemotron_v3`, `inkling`) as reasoning
faces over the already-landed `TOOLS-STREAMING-PARSER` engine.

### Ordering amendment 2026-08-13 — W3 runs BEFORE W2 (#605)

The waves above were numbered without demand data. The recipe-surface sweep
supplies it, and it inverts the order.

Measured over `vllm-project/recipes` @ `86c7777aa699482ef1ebd0c5da9fc540ccc00a40`
(157 official model recipes), `--reasoning-parser` is passed **76 times across 20
distinct values**. We resolve 15 of those uses; the other **61 abort startup** with
`unknown reasoning parser`.

| Wave | Recipe uses it covers | Notable |
|---|---:|---|
| **W3** (engine-backed adapters) | **43 / 76** | `qwen3` 18, `glm45` 11, `gemma4` 6, `deepseek_v4` 2, `nemotron_v3` 3, `inkling` 2 |
| W2 (remaining text families) | 18 / 76 | `kimi_k2` 4, `poolside_v1` 4, `hy_v3` 2, `step3p5` 2, `minimax_m3` 1 |
| — | 0 | `ernie45`, `granite`, `cohere_command3/4`, `openai_gptoss` — **zero recipe demand**, all in W2 |

`qwen3` alone (18 uses) outweighs every W2 name combined except `kimi_k2` and
`poolside_v1`. It is also the parser the published Qwen3.5 and Qwen3.6 recipes pass
to models **we already ship token-exact and gated** — so today the engine serves
the model and rejects its own recipe's flag.

**Therefore: W3 runs first, and within W3 the first brick is
`Qwen3ParserReasoningAdapter`** (upstream `vllm/reasoning/qwen3_engine_reasoning_parser.py`,
re-exported from `vllm/parser/engine/registered_adapters.py`), which serves BOTH
`qwen3` and `mimo` — `vllm/reasoning/__init__.py:87` registers `mimo` onto the same
class, so two names land for one port. `glm45` and `glm47` share
`Glm47MoeParserReasoningAdapter` the same way (`__init__.py:55,59`).

Three names belong in W3 and were missing from its list:

- `nano_v3` (1 recipe use) — add to W3.
- `kimi_k3` (1) and `ling3` (1) — **post-pin**; they are not in the registry at
`5559679229bc961848b121ccdeaa8fa5d79bec98`. Recorded here so they are not
rediscovered; they land with the next pin advance, not before.

W2 is not cancelled — it is resequenced behind W3. The four zero-demand names stay
in scope because upstream registers them and we mirror upstream; they are simply
not what a user hits first.
- **W4** — request-time `chat_template_kwargs` threading (`adjust_request`) +
reasoning-gated grammar FSM hold (cross-ref structured-output).

Expand Down
Loading