[Frontend] Add /collective_rpc API endpoint (#23075)

Signed-off-by: 22quinn <33176974+22quinn@users.noreply.github.com>

vllm/entrypoints/openai/api_server.py

@@ -1044,6 +1044,34 @@ if envs.VLLM_SERVER_DEV_MODE:
         is_sleeping = await engine_client(raw_request).is_sleeping()
         return JSONResponse(content={"is_sleeping": is_sleeping})
 
+    @router.post("/collective_rpc")
+    async def collective_rpc(raw_request: Request):
+        try:
+            body = await raw_request.json()
+        except json.JSONDecodeError as e:
+            raise HTTPException(status_code=HTTPStatus.BAD_REQUEST.value,
+                                detail=f"JSON decode error: {e}") from e
+        method = body.get("method")
+        if method is None:
+            raise HTTPException(status_code=HTTPStatus.BAD_REQUEST.value,
+                                detail="Missing 'method' in request body")
+        # For security reason, only serialized string args/kwargs are passed.
+        # User-defined `method` is responsible for deseralization if needed.
+        args: list[str] = body.get("args", [])
+        kwargs: dict[str, str] = body.get("kwargs", {})
+        timeout: Optional[float] = body.get("timeout")
+        results = await engine_client(raw_request).collective_rpc(
+            method=method, timeout=timeout, args=tuple(args), kwargs=kwargs)
+        if results is None:
+            return Response(status_code=200)
+        response: list[Any] = []
+        for result in results:
+            if result is None or isinstance(result, (dict, list)):
+                response.append(result)
+            else:
+                response.append(str(result))
+        return JSONResponse(content={"results": response})
+
 
 @router.post("/scale_elastic_ep",
              dependencies=[Depends(validate_json_request)],