o
Û¾i
ã @ s8 d dl mZ d dlmZ d dlmZ G dd„ deƒZdS )é )ÚChatCompletionRequest)ÚResponsesRequest)ÚBaseThinkingReasoningParserc @ s^ e Zd ZdZedefdd„ƒZedefdd„ƒZdedee B de
ed B ed B f fd
d „Z d S ) ÚQwen3ReasoningParserau
Reasoning parser for the Qwen3 model.
The Qwen3 model uses ... tokens to denote reasoning text
within its output. The model provides a strict switch to disable reasoning
output via the 'enable_thinking=False' parameter. This parser extracts the
reasoning content enclosed by and tokens from the model's
output.
Úreturnc C ó dS )z(The token that starts reasoning content.z© ©Úselfr r úY/home/ubuntu/.local/lib/python3.10/site-packages/vllm/reasoning/qwen3_reasoning_parser.pyÚ start_token ó z Qwen3ReasoningParser.start_tokenc C r )z&The token that ends reasoning content.zr r r r r Ú end_token r
z Qwen3ReasoningParser.end_tokenÚ model_outputÚrequestNc C st | j |vs
| j|vrd|fS | | j ¡}|d r |d n|d }| j|vr)d|fS | | j¡\}}}|p5d}||fS )aª
Extract reasoning content from the model output.
Qwen3 has stricter requirements - it needs both start and end tokens
to be present, unlike other models that work with just the end token.
For text abcxyz:
- 'abc' goes to reasoning
- 'xyz' goes to content
Returns:
tuple[Optional[str], Optional[str]]: reasoning content and content
Né é r )r r Ú partition)r
r r Úmodel_output_partsÚ reasoningÚ_ÚcontentÚ
final_contentr r r Úextract_reasoning" s ÿ
z&Qwen3ReasoningParser.extract_reasoning) Ú__name__Ú
__module__Ú __qualname__Ú__doc__ÚpropertyÚstrr r r r Útupler r r r r r
s
ÿÿþr N)Ú0vllm.entrypoints.openai.chat_completion.protocolr Ú*vllm.entrypoints.openai.responses.protocolr Ú vllm.reasoning.basic_parsersr r r r r r Ú s