xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-05-10 04:05:53 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Misc] Optimize the Qwen3_ReasoningParser extract_reasoning_content (#17515)

Browse Source 
Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
This commit is contained in:
Chauncey 2025-05-01 18:29:01 +08:00 committed by GitHub
parent fbefc8a78d
commit 015069b017
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 25 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

49
vllm/reasoning/qwen3_reasoning_parser.py
View File

@ -1,6 +1,5 @@
# SPDX-License-Identifier: Apache-2.0 # SPDX-License-Identifier: Apache-2.0
import re
from collections.abc import Sequence from collections.abc import Sequence
from typing import Optional, Union from typing import Optional, Union
@ -31,9 +30,6 @@ class Qwen3ReasoningParser(ReasoningParser):
self.think_start_token = "<think>" self.think_start_token = "<think>"
self.think_end_token = "</think>" self.think_end_token = "</think>"
self.reasoning_regex = re.compile(
rf"{self.think_start_token}(.*?){self.think_end_token}", re.DOTALL)
if not self.model_tokenizer: if not self.model_tokenizer:
raise ValueError( raise ValueError(
"The model tokenizer must be passed to the ReasoningParser " "The model tokenizer must be passed to the ReasoningParser "
@ -121,29 +117,34 @@ class Qwen3ReasoningParser(ReasoningParser):
def extract_reasoning_content( def extract_reasoning_content(
self, model_output: str, request: ChatCompletionRequest self, model_output: str, request: ChatCompletionRequest
) -> tuple[Optional[str], Optional[str]]: ) -> tuple[Optional[str], Optional[str]]:
"""
Extract reasoning content from the model output.
# Check if the model output contains the <think> tokens. For text <think>abc</think>xyz:
- 'abc' goes to reasoning_content
- 'xyz' goes to content
Returns:
tuple[Optional[str], Optional[str]]: reasoning content and content
"""
# Check if the model output contains the <think> and </think> tokens.
if (self.think_start_token not in model_output if (self.think_start_token not in model_output
or self.think_end_token not in model_output): or self.think_end_token not in model_output):
return None, model_output return None, model_output
else: # Check if the <think> is present in the model output, remove it
# Use a regex to find the reasoning content # if it is present.
reasoning_content = self.reasoning_regex.findall(model_output)[0] model_output_parts = model_output.partition(self.think_start_token)
model_output = model_output_parts[2] if model_output_parts[
1] else model_output_parts[0]
# Check if the model output contains the </think> tokens.
# If the end token is not found, return the model output as is.
if self.think_end_token not in model_output:
return None, model_output
# Remove the reasoning content from the model output # Extract reasoning content from the model output.
# Although <think> token is always at the reasoning_content, _, content = model_output.partition(
# beginning of the line, we cannot guarantee that the self.think_end_token)
# other models will follow this convention.
# Therefore, we need to add :start_index.
start_index = model_output.find(self.think_start_token)
if start_index != -1:
end_index = start_index + len(
f"{self.think_start_token}{reasoning_content}{self.think_end_token}"
)
model_output = model_output[:start_index] + \
model_output[end_index:]
if len(model_output) == 0: final_content = content or None
return reasoning_content, None return reasoning_content, final_content
return reasoning_content, model_output