[PD] Skip singleton transfer-status all-reduces (#40003)

Co-authored-by: Pranjal Shankhdhar <pranjalssh@users.noreply.github.com>
Co-authored-by: Jialin Ouyang <Jialin.Ouyang@gmail.com>

python/sglang/srt/disaggregation/utils.py

@@ -185,6 +185,9 @@ def _apply_metadata_gate(polls, decode_reqs, metadata_buffers) -> None:
 
 def _all_reduce_polls(polls: List[int], group: dist.ProcessGroup) -> List[int]:
     """MIN-reduce poll states so no rank commits ahead of its peers."""
+    if dist.get_world_size(group) == 1:
+        return polls
+
     tensor_to_reduce = torch.tensor(polls, dtype=torch.uint8, device="cpu")
     dist.all_reduce(tensor_to_reduce, op=dist.ReduceOp.MIN, group=group)
     return tensor_to_reduce.tolist()

python/sglang/srt/managers/scheduler.py

@@ -4888,9 +4888,12 @@ class Scheduler(
                 self.load_publisher.publish_load_stat(
                     self.load_inquirer.get_loads, force=True, snapshot=snapshot
                 )
-            if self.enable_hicache_storage:
-                # Storage workers need the GIL between I/O calls. Yield while
-                # there is no GPU batch so polling cannot starve their acks.
+            if (
+                self.enable_hicache_storage
+                or self.disaggregation_mode != DisaggregationMode.NULL
+            ):
+                # Storage and transfer workers need the GIL between I/O calls.
+                # Singleton PD polls no longer yield through a collective.
                 time.sleep(0)
             return
         self.metrics_reporter.record_scheduler_idle()