sgl-project/sglang · error · ValueError

think_end_token '{reasoning_parser.detector.think_end_token}

Error message

think_end_token '{reasoning_parser.detector.think_end_token}' could not be encoded by the tokenizer.

What it means

ReasonerGrammarBackend encodes the reasoning parser's think-end marker (e.g. '</think>') with the model tokenizer; if encoding yields no token IDs, it raises ValueError because strict reasoning enforcement cannot detect the end of thinking.

Source

Thrown at python/sglang/srt/constrained/reasoner_grammar_backend.py:257

                old_output_ids, new_output_ids, next_state
            )


class ReasonerGrammarBackend(BaseGrammarBackend):
    def __init__(
        self,
        grammar_backend: BaseGrammarBackend,
        reasoning_parser: ReasoningParser,
        tokenizer: Union[PreTrainedTokenizer, PreTrainedTokenizerFast],
        enable_strict_thinking: bool = False,
    ):
        super().__init__()
        self.grammar_backend = grammar_backend
        think_end_ids = tokenizer.encode(
            reasoning_parser.detector.think_end_token, add_special_tokens=False
        )
        if not think_end_ids:
            raise ValueError(
                f"think_end_token '{reasoning_parser.detector.think_end_token}' "
                f"could not be encoded by the tokenizer."
            )
        self.think_end_ids = think_end_ids
        self._enable_strict_thinking = enable_strict_thinking
        self.think_excluded_token_ids = self._get_think_excluded_token_ids(
            reasoning_parser, tokenizer
        )
        self.max_think_tokens = envs.SGLANG_MAX_THINK_TOKENS.get()
        self.enable_token_filter = self.enable_strict_thinking and (
            self.think_excluded_token_ids is not None or self.max_think_tokens >= 0
        )
        if (
            self.enable_token_filter
            and not self.grammar_backend.is_support_token_filter
        ):
            raise ValueError(
                "Strict reasoning format requested but the grammar backend does not "

View on GitHub (pinned to 0132848349)

Solutions

Verify reasoning_parser.detector.think_end_token is set to the correct literal (e.g. '</think>') for the model
Test tokenizer.encode(think_end_token, add_special_tokens=False) manually; if empty, fix the tokenizer/parser config
Use a different --reasoning-parser value matching the model

Defensive patterns

Strategy: validation

Validate before calling

ids = tokenizer.encode(parser.detector.think_end_token, add_special_tokens=False)
assert ids, 'think_end_token not encodable by this tokenizer'

Prevention

Add a startup assertion for reasoning-parser markers on new model integrations
Unit-test detector marker tokens against the model tokenizer in CI

When it happens

Trigger: Constructing ReasonerGrammarBackend where reasoning_parser.detector.think_end_token is empty or not representable by the tokenizer (yields zero tokens with add_special_tokens=False).

Common situations: Custom reasoning parsers with a misconfigured think_end_token; tokenizers whose vocab genuinely lacks pieces for the marker; empty-string detector fields on new model integrations.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/5aaf461dcd5170db. Report an issue: GitHub.