fix: pause_generation should not populate running_batch on prefill nodes (#20273)

python/sglang/srt/managers/scheduler.py

@@ -3278,7 +3278,14 @@ class Scheduler(
             self.last_batch.filter_batch(
                 chunked_req_to_exclude=list(chunked_req_to_exclude)
             )
-            if not self.last_batch.is_empty():
+            # Skip merge for disagg prefill: completed prefill requests are
+            # already in disagg_prefill_inflight_queue. Merging them into
+            # running_batch leaks them, since the prefill event loop never
+            # calls update_running_batch to clean them up.
+            if (
+                not self.last_batch.is_empty()
+                and self.disaggregation_mode != DisaggregationMode.PREFILL
+            ):
                 if self.running_batch.is_empty():
                     self.running_batch = self.last_batch
                 else:
@@ -3287,7 +3294,7 @@ class Scheduler(
         self.last_batch = None
         self.cur_batch = None
 
-        if recv_req.mode == "retract":
+        if recv_req.mode == "retract" and not self.running_batch.is_empty():
             self.running_batch.filter_batch(v1_spec_info_filtered=True)
             if len(self.running_batch.reqs) != 0:
                 retracted_reqs = self.running_batch.retract_all(self.server_args)