A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 10, 06:30 UTC
[Frontend] Add `reasoning_effort` to `OpenAIServing._preprocess_chat()` (#31956)
Signed-off-by: Sanghoon Yoon <seanyoon@kakao.com>
Sanghoon Yoon committed
Jan 13, 2026 at 12:21 UTC
60b77e146315f4f0f23b9aa83c07d51f38927a1e
2 files changed
+13
-1
vllm/entrypoints/openai/serving_chat.py
+5
-1
@@ -306,6 +306,10 @@ class OpenAIServingChat(OpenAIServing):
306
)
307
if error_check_ret is not None:
308
return error_check_ret
309
+
310
+ chat_template_kwargs = request.chat_template_kwargs or {}
311
+ chat_template_kwargs.update(reasoning_effort=request.reasoning_effort)
312
+
313
conversation, engine_prompts = await self._preprocess_chat(
314
request,
315
tokenizer,
@@ -316,7 +320,7 @@ class OpenAIServingChat(OpenAIServing):
320
continue_final_message=request.continue_final_message,
321
tool_dicts=tool_dicts,
322
documents=request.documents,
319
- chat_template_kwargs=request.chat_template_kwargs,
323
+ chat_template_kwargs=chat_template_kwargs,
324
default_chat_template_kwargs=self.default_chat_template_kwargs,
325
tool_parser=tool_parser,
326
add_special_tokens=request.add_special_tokens,
vllm/entrypoints/openai/serving_responses.py
+8
@@ -589,6 +589,13 @@ class OpenAIServingResponses(OpenAIServing):
589
prev_msg=self.msg_store.get(prev_response.id) if prev_response else None,
590
prev_response_output=prev_response.output if prev_response else None,
591
)
592
+
593
+ chat_template_kwargs = dict(
594
+ reasoning_effort=None
595
+ if request.reasoning is None
596
+ else request.reasoning.effort
597
+ )
598
+
599
_, engine_prompts = await self._preprocess_chat(
600
request,
601
tokenizer,
@@ -597,6 +604,7 @@ class OpenAIServingResponses(OpenAIServing):
604
tool_parser=self.tool_parser,
605
chat_template=self.chat_template,
606
chat_template_content_format=self.chat_template_content_format,
607
+ chat_template_kwargs=chat_template_kwargs,
608
)
609
return messages, engine_prompts
610