o
-i
@ s8 d dl mZ d dlmZ d dlmZ G dd deZdS ) )ChatCompletionRequest)ResponsesRequest)BaseThinkingReasoningParserc @ s^ e Zd ZdZedefddZedefddZdedee B de
ed B ed B f fd
d Z d S ) Qwen3ReasoningParserau
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.
returnc C dS )z(The token that starts reasoning content.z selfr r b/home/ubuntu/veenaModal/venv/lib/python3.10/site-packages/vllm/reasoning/qwen3_reasoning_parser.py start_token z Qwen3ReasoningParser.start_tokenc C r )z&The token that ends reasoning content.zr r r r r end_token r
z Qwen3ReasoningParser.end_token model_outputrequestNc C st | j |vs
| j|vrd|fS || j }|d r |d n|d }| j|vr)d|fS || j\}}}|p5d}||fS )a
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
N r )r r partition)r
r r model_output_parts reasoning_content
final_contentr r r extract_reasoning" s
z&Qwen3ReasoningParser.extract_reasoning) __name__
__module__ __qualname____doc__propertystrr r r r tupler r r r r r
s
r N)0vllm.entrypoints.openai.chat_completion.protocolr *vllm.entrypoints.openai.responses.protocolr vllm.reasoning.basic_parsersr r r r r r s