Support Thinking Budget (via custom_logit_processor for OpenAI API) [Fix #6572] (#11416)

Signed-off-by: ybyang <ybyang7@iflytek.com>
Co-authored-by: YorkSu <york_su@qq.com>

python/sglang/srt/entrypoints/openai/protocol.py

@@ -243,6 +243,8 @@ class CompletionRequest(BaseModel):
     lora_path: Optional[Union[List[Optional[str]], Optional[str]]] = None
     session_params: Optional[Dict] = None
     response_format: Optional[Union[ResponseFormat, StructuralTagResponseFormat]] = None
+    custom_params: Optional[Dict] = None
+    custom_logit_processor: Optional[str] = None
 
     # For PD disaggregation
     bootstrap_host: Optional[Union[List[str], str]] = None
@@ -504,6 +506,10 @@ class ChatCompletionRequest(BaseModel):
     stream_reasoning: bool = True
     chat_template_kwargs: Optional[Dict] = None
 
+    # Custom logit processor for advanced sampling control
+    custom_logit_processor: Optional[Union[List[Optional[str]], str]] = None
+    custom_params: Optional[Dict] = None
+
     # For request id
     rid: Optional[Union[List[str], str]] = None
     # Extra key for classifying the request (e.g. cache_salt)
@@ -636,6 +642,7 @@ class ChatCompletionRequest(BaseModel):
             "ignore_eos": self.ignore_eos,
             "skip_special_tokens": self.skip_special_tokens,
             "logit_bias": self.logit_bias,
+            "custom_params": self.custom_params,
         }
 
         if self.response_format and self.response_format.type == "json_schema":

python/sglang/srt/entrypoints/openai/serving_chat.py

@@ -196,6 +196,7 @@ class OpenAIServingChat(OpenAIServingBase):
             extra_key=self._compute_extra_key(request),
             priority=request.priority,
             custom_labels=custom_labels,
+            custom_logit_processor=request.custom_logit_processor,
         )
 
         return adapted_request, request

python/sglang/srt/entrypoints/openai/serving_completions.py

@@ -121,6 +121,7 @@ class OpenAIServingCompletion(OpenAIServingBase):
             extra_key=self._compute_extra_key(request),
             priority=request.priority,
             custom_labels=custom_labels,
+            custom_logit_processor=request.custom_logit_processor,
         )
 
         return adapted_request, request
@@ -149,6 +150,7 @@ class OpenAIServingCompletion(OpenAIServingBase):
             "ignore_eos": request.ignore_eos,
             "skip_special_tokens": request.skip_special_tokens,
             "logit_bias": request.logit_bias,
+            "custom_params": request.custom_params,
         }
 
         # Handle response_format constraints