Move _get_request_router to PR 4 and make private

This method is consumed by the ingress routing logic in this PR, not by
the base substrate PR. Renamed to private with TODO to migrate to
DeploymentHandle.choose_replica() once ray-project#60865 lands.

Signed-off-by: Seiji Eicher <seiji@anyscale.com>

Author: eicherseiji

python/ray/llm/_internal/serve/core/ingress/ingress.py

@@ -617,7 +617,7 @@ async def _pick_routed_direct_ingress_endpoint(
         if handle is None:
             raise RuntimeError(f"No handle for model {model_id}")
 
-        request_router = handle.get_request_router()
+        request_router = handle._get_request_router()
         if request_router is None:
             raise RuntimeError(f"Request router not initialized for {model_id}")
 

python/ray/serve/handle.py

@@ -1079,6 +1079,22 @@ def options(
             response_serialization=response_serialization,
         )
 
+    def _get_request_router(
+        self,
+    ) -> Optional["ray.serve._private.request_router.request_router.RequestRouter"]:
+        """Temporary: expose the request router for direct-ingress routing.
+
+        TODO: Replace with DeploymentHandle.choose_replica() once #60865 lands.
+        """
+        if self._router is None:
+            return None
+
+        asyncio_router = getattr(self._router, "_asyncio_router", None)
+        if asyncio_router is not None:
+            return asyncio_router.request_router
+
+        return getattr(self._router, "request_router", None)
+
     def remote(
         self, *args, **kwargs
     ) -> Union[DeploymentResponse[Any], DeploymentResponseGenerator[Any]]: