diff --git a/python/sglang/srt/managers/data_parallel_controller.py b/python/sglang/srt/managers/data_parallel_controller.py
index 57632e146d94..ff7c15b805cd 100644
--- a/python/sglang/srt/managers/data_parallel_controller.py
+++ b/python/sglang/srt/managers/data_parallel_controller.py
@@ -559,6 +559,10 @@ def launch_tensor_parallel_group(
 
         self.max_total_num_tokens = scheduler_info[0]["max_total_num_tokens"]
         self.max_req_input_len = scheduler_info[0]["max_req_input_len"]
+        # Retain the full first child init dict so callers that build
+        # the upstream handshake (run_data_parallel_controller_process,
+        # below) can forward every field — not just the two limits.
+        self.scheduler_init_info = scheduler_info[0]
 
     def maybe_external_dp_rank_routing(self, req: Req):
         if req.routed_dp_rank is not None:
@@ -649,10 +653,12 @@ def run_data_parallel_controller_process(
             proc.pid for proc in controller.scheduler_procs if proc is not None
         ]
         pipe_writer.send(
+            # Spread the first child scheduler's full init dict (status +
+            # both max_* limits + the multimodal/tokenizer metadata added
+            # by the get_init_info patch above) so dp_size > 1 doesn't
+            # strip them before they reach the gRPC servicer.
             {
-                "status": "ready",
-                "max_total_num_tokens": controller.max_total_num_tokens,
-                "max_req_input_len": controller.max_req_input_len,
+                **controller.scheduler_init_info,
                 SCHEDULER_PIDS_ARG: scheduler_pids,
             }
         )
diff --git a/python/sglang/srt/managers/scheduler.py b/python/sglang/srt/managers/scheduler.py
index 689cf9837bcc..6b9fdcc4cea6 100644
--- a/python/sglang/srt/managers/scheduler.py
+++ b/python/sglang/srt/managers/scheduler.py
@@ -1494,10 +1494,32 @@ def get_init_info(self) -> Dict[str, Any]:
         This method provides the initialization info needed by the tokenizer manager
         and other components to verify the scheduler is ready.
         """
+        # Fields below `max_req_input_len` are read by smg-grpc-servicer's
+        # GetModelInfo bridge (sglang/server.py:81-97, servicer.py:325-329)
+        # out of scheduler_info via .get(..., default). Without them the
+        # gateway sees defaults — supports_vision in particular is always
+        # False, which makes multimodal workers invisible to image/audio
+        # routing. They are no-ops for HTTP-mode consumers.
+        hf_cfg = self.model_config.hf_config
+        pad_id = getattr(hf_cfg, "pad_token_id", None)
+        bos_id = getattr(hf_cfg, "bos_token_id", None)
+        # `hf_eos_token_id` is annotated Optional[Set[int]] but the
+        # producer accepts int/list/None inputs; coerce defensively.
+        eos_ids = self.model_config.hf_eos_token_id
+        eos_token_ids = sorted([eos_ids] if isinstance(eos_ids, int) else eos_ids or [])
         result_dict = {
             "status": "ready",
             "max_total_num_tokens": self.max_total_num_tokens,
             "max_req_input_len": self.max_req_input_len,
+            "is_generation": self.is_generation,
+            "supports_vision": self.model_config.is_multimodal,
+            "vocab_size": self.model_config.vocab_size,
+            "eos_token_ids": eos_token_ids,
+            # int32 proto field; coerce missing IDs to the smg-grpc-servicer
+            # int defaults (0 for pad, 1 for bos) so encoding doesn't crash
+            # on `None`.
+            "pad_token_id": pad_id if pad_id is not None else 0,
+            "bos_token_id": bos_id if bos_id is not None else 1,
         }
 
         return result_dict