chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[Reasoning + Structured Output] make reasoning compatible with structured output (#12551)

Browse Source 
Signed-off-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
This commit is contained in:
Muqi Li
2025-12-08 01:28:39 -08:00
committed by GitHub
co-authored by Xinyuan Tong
parent f72a77038f
commit 06836ad02a
8 changed files with 72 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/entrypoints/openai/serving_chat.py
+23 -24
View File
@@ -213,6 +213,7 @@ class OpenAIServingChat(OpenAIServingBase):
return_hidden_states=request.return_hidden_states,
rid=request.rid,
extra_key=self._compute_extra_key(request),
reasoning=self._get_reasoning_from_request(request),
priority=request.priority,
custom_labels=custom_labels,
custom_logit_processor=request.custom_logit_processor,
@@ -443,7 +444,10 @@ class OpenAIServingChat(OpenAIServingBase):
prompt = prompt[: -len(conv.sep2)]
else:
prompt = conv.get_prompt()
if self._get_enable_thinking_from_request(request):
if self._get_reasoning_from_request(
request
) and self.reasoning_parser not in ["qwen3", "qwen3-thinking", "glm4"]:
# qwen3 and glm4 think internally without a leading <think> token
prompt += "<think>" # Note(Xinyuan): hard code thinking token
image_data = conv.image_data if conv.image_data else None
@@ -775,7 +779,7 @@ class OpenAIServingChat(OpenAIServingBase):
if reasoning_parser and request.separate_reasoning:
is_force_reasoning = (
self.template_manager.force_reasoning
or self._get_enable_thinking_from_request(request)
or self._get_reasoning_from_request(request)
)
try:
parser = ReasoningParser(
@@ -1022,7 +1026,7 @@ class OpenAIServingChat(OpenAIServingBase):
if index not in reasoning_parser_dict:
is_force_reasoning = (
self.template_manager.force_reasoning
or self._get_enable_thinking_from_request(request)
or self._get_reasoning_from_request(request)
)
reasoning_parser_dict[index] = ReasoningParser(
self.reasoning_parser,
@@ -1052,27 +1056,22 @@ class OpenAIServingChat(OpenAIServingBase):
idx += len(list(tool_calls)) if tool_calls is not None else 0 # noqa
return idx
def _get_enable_thinking_from_request(self, request: ChatCompletionRequest) -> bool:
"""Extracts the 'enable_thinking' flag from request chat_template_kwargs.
NOTE: This parameter is only useful for models that support enable_thinking
flag, such as Qwen3.
Args:
request_obj: The request object (or an item from a list of requests).
Returns:
The boolean value of 'enable_thinking' if found, otherwise False.
"""
if hasattr(request, "chat_template_kwargs") and request.chat_template_kwargs:
# For Qwen3 models, `enable_thinking` is supported.
if self.reasoning_parser in ["qwen3", "glm45"]:
return request.chat_template_kwargs.get("enable_thinking", False)
# For DeepSeek-V3.1 models, `thinking` is supported.
elif self.reasoning_parser in ["deepseek-v3"]:
return request.chat_template_kwargs.get("thinking", False)
else:
return False
return False
def _get_reasoning_from_request(self, request: ChatCompletionRequest) -> bool:
"""Judge whether the request needs reasoning"""
if not self.reasoning_parser:
return False
if self.reasoning_parser in ["deepseek-v3"]:
return (
request.chat_template_kwargs is not None
and request.chat_template_kwargs.get("thinking") is True
)
if self.reasoning_parser in ["qwen3", "glm45"]:
# qwen3 and glm45 are reasoning by default
return (
not request.chat_template_kwargs
or request.chat_template_kwargs.get("enable_thinking", True) is True
)
return True # default
async def _process_tool_call_stream(
self,