Publish elastic-EP active ranks from a dedicated step (#25721)

python/sglang/srt/managers/scheduler.py

@@ -2923,10 +2923,16 @@ class Scheduler(
                     pooled_hidden_states=pooler_output.pooled_hidden_states,
                 )
 
-        if (
+        self._maybe_report_active_ranks()
+
+        return ret
+
+    def _maybe_report_active_ranks(self) -> None:
+        if not (
             self.server_args.enable_dp_attention
             and self.server_args.elastic_ep_backend is not None
         ):
+            return
         # Get the tensors indicating rank activeness
         tp_active_ranks = self.tp_group.active_ranks.detach().cpu().numpy()
         tp_active_ranks_cpu = self.tp_group.active_ranks_cpu.detach().numpy()
@@ -2936,8 +2942,6 @@ class Scheduler(
             ActiveRanksOutput(status=dp_active_ranks.tolist())
         )
 
-        return ret
-
     def launch_batch_sample_if_needed(
         self, batch_result: GenerationBatchResult
     ) -> Union[GenerationBatchResult]: