Vllm Reasoning Parsers

vLLM reasoning parsers

Target: operators wiring up --reasoning-parser NAME on a chat-completion endpoint, or developers authoring a parser for a new thinking model. Source of truth: vllm/reasoning/ on main.

What a reasoning parser actually does

When a reasoning-trained model emits a single token stream like

<think>user asked X, let me check Y...</think>The answer is 42.

vLLM splits this into two fields on the chat-completion response: reasoning (the CoT) and content (the final answer). --reasoning-parser NAME selects the class that does the split. Without it, the whole stream lands in content.

Field-name note. On current main the field is reasoning (see ChatMessage.reasoning / DeltaMessage.reasoning in vllm/entrypoints/openai/chat_completion/protocol.py). Pre-v0.19 code and many third-party docs / clients call it reasoning_content. If a client is reading reasoning_content against a current-main server it will see null every time even when the parser ran correctly.

The parser is also the gating authority for xgrammar / structured output: by default, grammar enforcement is held off until is_reasoning_end(input_ids) flips true, so the model thinks freely before being constrained to JSON. Flip that default with --structured-outputs-config.enable_in_reasoning=true — then the grammar applies from token 0 regardless of reasoning state (useful for structured CoT).

The contract (ReasoningParser ABC)

vllm/reasoning/abs_reasoning_parsers.py. Every parser implements:

Optional:

• count_reasoning_tokens(token_ids) — for usage accounting. Default returns 0. BaseThinkingReasoningParser implements it via depth counter so nested <think>…<think>…</think>…</think> are handled.
• adjust_request(request) — mutate request (rare).
• prepare_structured_tag(original, tool_server) — emit a structural_tag JSON (GPT-OSS uses this for harmony).
• reasoning_start_str / reasoning_end_str — expose delimiter strings so ReasoningConfig.initialize_token_ids can derive the multi-token ID sequences automatically.

Streaming vs. non-streaming are two independent code paths. A parser that's correct on extract_reasoning can be buggy on extract_reasoning_streaming and vice versa. Every parser must be tested on both.

The CLI path (what --reasoning-parser qwen3 actually triggers)

vllm/engine/arg_utils.py:985 (v0.27.1; grep for --reasoning-parser — line moves between releases) declares reasoning_parser: str = StructuredOutputsConfig.reasoning_parser. api_server.py validates it against ReasoningParserManager.list_registered() at startup (invalid name = fast-fail with the list of registered names).

On request, OpenAIServingChat instantiates a fresh parser per request via self.reasoning_parser_cls(tokenizer, chat_template_kwargs=chat_template_kwargs) (vllm/entrypoints/openai/chat_completion/serving.py:240). "Fresh per request" is load-bearing for stateful parsers — see Hunyuan in the matrix.

Registry: vllm/reasoning/__init__.py has a _REASONING_PARSERS_TO_REGISTER dict of name → (filename, ClassName) that feeds ReasoningParserManager.register_lazy_module. Lazy import means a broken parser file won't crash vLLM startup until somebody selects it.

Plugin path: --reasoning-parser-plugin /path/to/my_parser.py calls ReasoningParserManager.import_reasoning_parser(path), which importlib-loads the file. The file registers itself via @ReasoningParserManager.register_module(["my-name"]) at import time. Then --reasoning-parser my-name selects it.

The 15 things that go wrong

See references/pitfalls.md for each with repros and fixes. Quick index:

1. reasoning_content is null on DeepSeek-R1 — chat template injected <think> into the prompt, so the model never emitted a start token. Parser must tolerate missing start (base BaseThinkingReasoningParser does, via the .partition(start_token) pattern).

2. content is empty, CoT in reasoning_content with enable_thinking=False — parser didn't branch on chat_template_kwargs. Qwen3 / DeepSeek-V3 / Kimi K2 route to IdentityReasoningParser (or internal flag) when thinking is off. DeepSeek-V3 has two thinking-default variants: deepseek_v3 (default off) and holo2 = DeepSeekV3ReasoningWithThinkingParser (default on). glm45 used to be in that second group and no longer is — as of v0.25.1 it moved onto the new adapter path alongside glm47, so don't assume glm45 and holo2 still behave alike.

3. Gibberish JSON when guided_json + enable_thinking=False — the reasoning parser's is_reasoning_end on the prompt must return True so xgrammar enforces from token 0. Serving layer caches this result as prompt_is_reasoning_end_arr[i]. If the parser only checks for </think> in input_ids and the thinking-disabled chat template emits <think>\n\n</think>\n\n in the prompt, this works; if it does something else, xgrammar silently stays gated.

4. Truncated output = wrong field — when the model hits max_tokens mid-reasoning, is the whole output "reasoning" or "content"? Qwen3 with enable_thinking=True (default) → (model_output, None) = all reasoning. Qwen3 with enable_thinking=False → (None, model_output) = all content. DeepSeek-R1 base → (model_output, None). Know the parser's policy before shipping.

5. Nested <think> tags — base parser's partition stops at the first </think>; if a CoT contains a literal </think> substring (sometimes seen in distillation artifacts), content gets split wrong. count_reasoning_tokens uses a depth counter so counts are right, but the split is not.

6. Tool calls break with reasoning enabled — tool parser only sees the content half of extract_reasoning's return. If the reasoning parser returns (everything, None), tool parser sees nothing. Kimi K2 handles this by treating <|tool_calls_section_begin|> as an implicit reasoning-end marker.

7. Stateful parser reused across requests — Hunyuan A13B's extract_reasoning_streaming is a token-ID state machine with self.current_state / self.token_buffer. Second concurrent request on the same instance = interleaved garbage. vLLM already instantiates per-request, but custom plugins must not hoist state to class-level.

8. Multi-token delimiter — <think> may encode as a single token (DeepSeek-R1 vocab) or as multiple (<think, > in some tokenizers). Single-token path: vocab.get("<think>"). Multi-token: tokenizer.encode("<think>") and do sequence match — the current in-tree worked example is kimi_k3_reasoning_parser.py, whose <|open|>think<|sep|> marker is a 3-token sequence, so its token-id helpers search for the subsequence rather than one id. (The old example, GPT-OSS reasoning_end_token_ids_prefix, was deleted at v0.27.0 — see pitfall 10.) BaseThinkingReasoningParser raises at init if vocab.get returns None — don't inherit from it unless the delimiter is a single vocab entry.

9. Single-token delta spam — if delta is exactly one token that is <think> or </think>, return None from streaming so the client doesn't see an empty delta. Almost every parser has this skip.

10. Harmony / GPT-OSS is different — no <think> tag; the channel markers are multi-token. At v0.27.0 (PR #45560) GptOssReasoningParser was gutted: is_reasoning_end / is_reasoning_end_streaming now return True unconditionally, and extract_reasoning, extract_reasoning_streaming and extract_content_ids all raise NotImplementedError pointing at HarmonyParser. The old backward-scan machinery (reasoning_end_token_ids_prefix, reasoning_max_num_between_tokens = 20, eom_token_id) is deleted — don't look for it, and don't cite it as the multi-token-delimiter example. Boundary detection plus all parsing now live in HarmonyParser (vllm/parser/harmony.py). Don't copy DeepSeek-R1 for a harmony model.

11. Mistral moved onto the unified engine at v0.27.0 (PR #48947). --reasoning-parser mistral is unchanged, but the registry class is now MistralParserReasoningAdapter and vllm/reasoning/mistral_reasoning_parser.py is a three-line re-export — the logic is in vllm/parser/mistral.py, shared with --tool-call-parser mistral. The hard ValueError: The tokenizer must be an instance of MistralTokenizer. is gone (it existed at v0.25.1; git grep finds it nowhere at v0.27.0): is_mistral_tokenizer() now selects a path, falling back to the generic adjust_request when the tokenizer isn't Mistral or can't do grammar. Still prefer --tokenizer-mode mistral — grammar-enforced tool_choice needs the mistral-common path.

12. Granite is regex-on-text — "Here is my thought process:" / "Here is my response:" are phrases, not special tokens. Streaming parser has to buffer partial matches across deltas (Here is my thou…) which makes it the most complex parser in the tree. Read granite_reasoning_parser.py:140+ before modifying.

13. Implicit reasoning-end — Kimi K2 ends at </think> or <|tool_calls_section_begin|>. MiniMax M2 never emits <think>, only </think>. Custom is_reasoning_end must encode these facts or xgrammar gates at the wrong moment.

14. --enable-reasoning is gone — older docs / Stack Overflow answers still reference it. Since roughly v0.8 the only flag is --reasoning-parser NAME; the enable/disable is implicit in whether one is passed.

15. reasoning_content is always null — but parser worked fine. Current-main response field is reasoning, not reasoning_content (renamed in protocol.py). Client-side name mismatch that looks exactly like a parser failure. Before debugging parsers, jq '.choices[0].message | keys' to see what fields actually exist — if reasoning is there, it's just a client rename.

The per-model matrix

references/parser-matrix.md — one row per registered name (29 at v0.27.0: deepseek_r1, deepseek_v3, deepseek_v4, poolside_v1, cohere_command3, cohere_command4, ernie45, gemma4, glm45, glm47, openai_gptoss, granite, holo2, hunyuan_a13b, hy_v3, inkling, kimi_k2, kimi_k3, mimo, minimax_m2, minimax_m2_append_think, minimax_m3, mistral, nemotron_v3, olmo3, qwen3, seed_oss, step3, step3p5) with: delimiter style, start-token-in-prompt-or-output, thinking-disable mechanism, truncation policy, structured-output gating peculiarities.

Routing (which family each name belongs to — <think> two-token, delegating wrapper, stateful, harmony, phrase-regex, tokenizer-gated) lives in the matrix Family column. The non-obvious cases worth knowing before reading it: openai_gptoss is harmony (extract_reasoning raises NotImplementedError), mistral requires MistralTokenizer, hunyuan_a13b is a token-ID state machine, granite is phrase-regex on text, and nemotron_v3 swaps reasoning↔content on enable_thinking=False.

Writing a custom parser

references/writing-custom-parser.md for the step-by-step. This ReasoningParser subclass shape remains the supported path for out-of-tree parsers — it is what --reasoning-parser-plugin loads, and 17 of the 29 in-tree names still use it.

Be aware there is now a second, in-tree-only shape: vllm/parser/<model>.py defines one parser class per model, and make_adapters(XParser) in vllm/parser/engine/registered_adapters.py derives both XParserReasoningAdapter and XParserToolAdapter from it — the unified reasoning+tool design of RFC #32713. 12 of the 29 names are on it at v0.27.0: deepseek_v4, gemma4, glm45, glm47, inkling, kimi_k2, mimo, minimax_m2, mistral, nemotron_v3, qwen3, seed_oss. If you open one of those as a worked example you will find a three-line re-export and the logic elsewhere.

Do not infer membership from the filename. Only some are named *_engine_reasoning_parser.py; kimi_k2_reasoning_parser.py, minimax_m2_reasoning_parser.py, glm47_moe_reasoning_parser.py and mistral_reasoning_parser.py are ordinary names that are nonetheless shims. The test is the import: grep -l "registered_adapters import" vllm/reasoning/*.py.

Copy a legacy-path parser when writing your own. Shape:

from vllm.reasoning import ReasoningParser, ReasoningParserManager
from vllm.entrypoints.openai.engine.protocol import DeltaMessage

@ReasoningParserManager.register_module(["my_model"])
class MyReasoningParser(ReasoningParser):
    def __init__(self, tokenizer, *args, **kwargs):
        super().__init__(tokenizer, *args, **kwargs)
        # resolve delimiter token IDs via self.vocab or tokenizer.encode(...)

    def is_reasoning_end(self, input_ids): ...
    def extract_content_ids(self, input_ids): ...
    def extract_reasoning(self, model_output, request): ...
    def extract_reasoning_streaming(self, prev_text, cur_text, delta_text,
                                    prev_ids, cur_ids, delta_ids): ...

Register decorator → lazy entry in ReasoningParserManager.lazy_parsers. Pass the file to vllm serve ... --reasoning-parser-plugin /path/to/my_parser.py --reasoning-parser my_model.

Shortcuts when delimiter is a single vocab token: inherit BaseThinkingReasoningParser (vllm/reasoning/basic_parsers.py), override start_token / end_token properties only. That covers 90% of the surface — see deepseek_r1_reasoning_parser.py (32 lines) for the minimal concrete subclass.

Testing a parser

Unit tests live in tests/reasoning/ (DeepSeek R1, Qwen3, Granite, Hunyuan, GPT-OSS, Kimi K2 all have coverage). Run a single file with:

.venv/bin/python -m pytest tests/reasoning/test_qwen3_reasoning_parser.py -v

Every new parser needs tests for:

• Non-streaming: start-in-prompt, start-in-output, truncated (no end), nested, empty.
• Streaming: delta that is only the start token, only the end token, end token spanning two deltas, content-before-end-token in same delta, content-after-end-token in same delta.
• is_reasoning_end on both pre-reasoning, mid-reasoning, and post-reasoning token ID sequences.
• Thinking-disabled path if applicable.

Companion skills

• vllm-chat-templates — how the chat template injects <think> into the prompt (or not); why enable_thinking=False can change what the parser sees.
• vllm-tool-parsers — tool-call extraction runs on the post-reasoning content only. A wrong reasoning split silently breaks tool calling.
• vllm-configuration — chat_template_kwargs flows through ChatCompletionRequest.chat_template_kwargs into the parser __init__.
• vllm-performance-tuning — structured output tuning; reasoning parser choice affects xgrammar gating latency.
• vllm-input-modalities — reasoning parsers don't apply to embedding / rerank / ASR endpoints.

Upstream issue/PR anchors for each pitfall → references/sources.md.