[OpenAI] Propagate PD routing metadata through /v1/responses (#35503)

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

@@ -1652,6 +1652,18 @@ class ResponsesRequest(BaseModel):
         default=None, description="Cache salt for request caching"
     )
 
+    # For PD disaggregation
+    bootstrap_host: Optional[Union[List[str], str]] = None
+    bootstrap_port: Optional[Union[List[Optional[int]], int]] = None
+    bootstrap_room: Optional[Union[List[int], int]] = None
+
+    # For DP routing — external router assigns a specific DP worker
+    routed_dp_rank: Optional[int] = None
+    # For PD disagg — hint telling decode which prefill DP worker has the KV cache
+    disagg_prefill_dp_rank: Optional[int] = None
+    # Deprecated: use routed_dp_rank instead
+    data_parallel_rank: Optional[int] = None
+
     # SGLang sampling extras. ``None`` defers to ``--preferred-sampling-params``.
     frequency_penalty: float = 0.0
     presence_penalty: float = 0.0
@@ -1669,6 +1681,11 @@ class ResponsesRequest(BaseModel):
         "repetition_penalty": 1.0,
     }
 
+    @model_validator(mode="before")
+    @classmethod
+    def _handle_deprecated_dp_rank(cls, values):
+        return _migrate_deprecated_dp_rank(values)
+
     @model_validator(mode="before")
     @classmethod
     def normalize_reasoning_to_thinking(cls, values):

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

@@ -435,6 +435,10 @@ class OpenAIServingResponses(OpenAIServingChat):
                         else {}
                     )
 
+                    effective_routed_dp_rank = self.extract_routed_dp_rank_from_header(
+                        raw_request, request.routed_dp_rank
+                    )
+
                     adapted_request = GenerateReqInput(
                         **prompt_kwargs,
                         **logprob_kwargs,
@@ -464,6 +468,11 @@ class OpenAIServingResponses(OpenAIServingChat):
                         session_id=request.session_id,
                         extra_key=request.extra_key,
                         cache_salt=request.cache_salt,
+                        bootstrap_host=request.bootstrap_host,
+                        bootstrap_port=request.bootstrap_port,
+                        bootstrap_room=request.bootstrap_room,
+                        routed_dp_rank=effective_routed_dp_rank,
+                        disagg_prefill_dp_rank=request.disagg_prefill_dp_rank,
                         # background+stream streams on this connection, so don't detach.
                         background=request.background and not request.stream,
                         require_reasoning=require_reasoning,