A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 7, 07:45 UTC
[Bugfix] Fix coord_socket assertion in DPEngineCoreProc for offline DP mode (#35916)
Signed-off-by: Jaewon Lee <jaewon@meta.com>
Jaewon committed
Mar 3, 2026 at 16:10 UTC
f22ff2958c398ae0950598cdbb9c677c027fa5db
1 file changed
+5
-1
vllm/v1/engine/core.py
+5
-1
@@ -1571,7 +1571,11 @@ class DPEngineCoreProc(EngineCoreProc):
1571
1572
def resume_scheduler(self):
1573
super().resume_scheduler()
1574
- if not self.engines_running and self.scheduler.has_unfinished_requests():
1574
+ if (
1575
+ self.has_coordinator
1576
+ and not self.engines_running
1577
+ and self.scheduler.has_unfinished_requests()
1578
+ ):
1579
# Wake up other DP engines.
1580
self.output_queue.put_nowait(
1581
(-1, EngineCoreOutputs(start_wave=self.current_wave))