bi_150-vllm/vllm/reasoning/qwen3_reasoning_parser.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

from collections.abc import Sequence

from vllm.entrypoints.openai.chat_completion.protocol import (
    ChatCompletionRequest,
)
from vllm.entrypoints.openai.engine.protocol import DeltaMessage
from vllm.entrypoints.openai.responses.protocol import (
    ResponsesRequest,
)
from vllm.reasoning.basic_parsers import BaseThinkingReasoningParser
from vllm.tokenizers import TokenizerLike


class Qwen3ReasoningParser(BaseThinkingReasoningParser):
    """
    Reasoning parser for the Qwen3/Qwen3.5 model family.

    The Qwen3 model family uses <think>...</think> tokens to denote reasoning
    text. Starting with Qwen3.5, the chat template places <think> in the
    prompt so only </think> appears in the generated output. The model
    provides a strict switch to disable reasoning output via the
    'enable_thinking=False' parameter.

    When thinking is disabled, the template places <think>\\n\\n</think>\\n\\n
    in the prompt. The serving layer detects this via prompt_is_reasoning_end
    and routes deltas as content without calling the streaming parser.

    NOTE: Models up to the 2507 release (e.g., Qwen/Qwen3-235B-A22B-Instruct-2507)
    use an older chat template where the model generates <think> itself.
    This parser handles both styles: if <think> appears in the generated output
    it is stripped before extraction (non-streaming) or skipped (streaming).
    """

    def __init__(self, tokenizer: TokenizerLike, *args, **kwargs):
        super().__init__(tokenizer, *args, **kwargs)

        chat_kwargs = kwargs.get("chat_template_kwargs", {}) or {}
        # Qwen3 defaults to thinking enabled; only treat output as
        # pure content when the user explicitly disables it.
        self.thinking_enabled = chat_kwargs.get("enable_thinking", True)

    @property
    def start_token(self) -> str:
        """The token that starts reasoning content."""
        return "<think>"

    @property
    def end_token(self) -> str:
        """The token that ends reasoning content."""
        return "</think>"

    def extract_reasoning(
        self, model_output: str, request: ChatCompletionRequest | ResponsesRequest
    ) -> tuple[str | None, str | None]:
        """
        Extract reasoning content from the model output.

        The <think> token is placed in the prompt by the chat template,
        so typically only </think> appears in the generated output.
        If <think> is present (e.g. from a different template), it is
        stripped before extraction.

        When thinking is explicitly disabled and no </think> appears,
        returns (None, model_output) — all output is content.
        Otherwise (thinking enabled, default), a missing </think> means
        the output was truncated and everything is reasoning:
        returns (model_output, None).

        Returns:
            tuple[Optional[str], Optional[str]]: reasoning content and content
        """

        # Strip <think> if present in the generated output.
        model_output_parts = model_output.partition(self.start_token)
        model_output = (
            model_output_parts[2] if model_output_parts[1] else model_output_parts[0]
        )

        if self.end_token not in model_output:
            if not self.thinking_enabled:
                # Thinking explicitly disabled — treat everything as content.
                return None, model_output
            # Thinking enabled but no </think>: output was truncated.
            # Everything generated so far is reasoning.
            return model_output, None

        # Extract reasoning content from the model output.
        reasoning, _, content = model_output.partition(self.end_token)

        final_content = content or None
        return reasoning, final_content

    def extract_reasoning_streaming(
        self,
        previous_text: str,
        current_text: str,
        delta_text: str,
        previous_token_ids: Sequence[int],
        current_token_ids: Sequence[int],
        delta_token_ids: Sequence[int],
    ) -> DeltaMessage | None:
        """
        Extract reasoning content from a streaming delta.

        Since <think> is placed in the prompt by the chat template, all
        generated tokens before </think> are reasoning and tokens after
        are content.

        NOTE: When thinking is disabled, no think tokens appear in the
        generated output. The serving layer detects this via
        prompt_is_reasoning_end and routes deltas as content without
        calling this method.
        """
        # Strip <think> from delta if present (old template / edge case
        # where the model generates <think> itself).
        if self.start_token_id in delta_token_ids:
            start_idx = delta_text.find(self.start_token)
            if start_idx >= 0:
                delta_text = delta_text[start_idx + len(self.start_token) :]

        if self.end_token_id in delta_token_ids:
            # End token in this delta: split reasoning from content.
            end_index = delta_text.find(self.end_token)
            if end_index >= 0:
                reasoning = delta_text[:end_index]
                content = delta_text[end_index + len(self.end_token) :]
                if not reasoning and not content:
                    return None
                return DeltaMessage(
                    reasoning=reasoning if reasoning else None,
                    content=content if content else None,
                )
            # end_token_id in IDs but not in text (already stripped)
            return None

        # No end token in this delta.
        if not delta_text:
            # Nothing left after stripping start token.
            return None
        elif self.end_token_id in previous_token_ids:
            # End token already passed: everything is content now.
            return DeltaMessage(content=delta_text)
        else:
            # No end token yet: still in reasoning phase.
            return DeltaMessage(reasoning=delta_text)