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 6, 10:45 UTC
[bug] Fix deadlock with pause resume and collective_rpc (#37024)
Signed-off-by: hao-aaron <ahao@anyscale.com>
Aaron Hao committed
Mar 18, 2026 at 18:49 UTC
6accb21f2a9aa2411a2eb23f96ac90828554f3f4
1 file changed
+5
-1
vllm/v1/engine/core.py
+5
-1
@@ -1632,7 +1632,11 @@ class DPEngineCoreProc(EngineCoreProc):
1632
if self.has_coordinator and request_wave != self.current_wave:
1633
if request_wave > self.current_wave:
1634
self.current_wave = request_wave
1635
- elif not self.engines_running:
1635
+ elif (
1636
+ not self.engines_running
1637
+ and self.scheduler.pause_state == PauseState.UNPAUSED
1638
+ ):
1639
+ self.engines_running = True
1640
# Request received for an already-completed wave, notify
1641
# front-end that we need to start the next one.
1642
self.output_queue.put_nowait(