Fix dp-aware incompatible with completions and chat completions APIs (#14647)

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

@@ -257,6 +257,9 @@ class CompletionRequest(BaseModel):
     bootstrap_port: Optional[Union[List[Optional[int]], int]] = None
     bootstrap_room: Optional[Union[List[int], int]] = None
 
+    # For data parallel rank routing
+    data_parallel_rank: Optional[int] = None
+
     # For request id
     rid: Optional[Union[List[str], str]] = None
     # Extra key for classifying the request (e.g. cache_salt)
@@ -530,6 +533,9 @@ class ChatCompletionRequest(BaseModel):
     bootstrap_port: Optional[Union[List[Optional[int]], int]] = None
     bootstrap_room: Optional[Union[List[int], int]] = None
 
+    # For data parallel rank routing
+    data_parallel_rank: Optional[int] = None
+
     # OpenAI/SGLang default sampling parameters
     _DEFAULT_SAMPLING_PARAMS = {
         "temperature": 1.0,

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

@@ -210,6 +210,7 @@ class OpenAIServingChat(OpenAIServingBase):
             bootstrap_host=request.bootstrap_host,
             bootstrap_port=request.bootstrap_port,
             bootstrap_room=request.bootstrap_room,
+            data_parallel_rank=request.data_parallel_rank,
             return_hidden_states=request.return_hidden_states,
             rid=request.rid,
             extra_key=self._compute_extra_key(request),

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

@@ -116,6 +116,7 @@ class OpenAIServingCompletion(OpenAIServingBase):
             bootstrap_host=request.bootstrap_host,
             bootstrap_port=request.bootstrap_port,
             bootstrap_room=request.bootstrap_room,
+            data_parallel_rank=request.data_parallel_rank,
             return_hidden_states=request.return_hidden_states,
             rid=request.rid,
             extra_key=self._compute_extra_key(request),