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 15, 06:00 UTC
[BugFix] Fix DP Coordinator incorrect debug log message (#19624)
Signed-off-by: Nick Hill <nhill@redhat.com>
Nick Hill committed
Jun 13, 2025 at 17:18 UTC
bd517eb9fef75625d6f11fbda384b6d88eac7a6f
1 file changed
+13
-10
vllm/v1/engine/coordinator.py
+13
-10
@@ -183,11 +183,12 @@ class CoordinatorProc:
183
# engines are paused, so that we can wake the other
184
# engines.
185
engine_to_exclude, wave = msgspec.msgpack.decode(buffer)
186
- if wave < self.current_wave:
187
- # If the wave number is stale, ensure the message is
188
- # handled by all the engines.
189
- engine_to_exclude = None
186
if not self.engines_running:
187
+ if wave < self.current_wave:
188
+ # If the wave number is stale, ensure the message
189
+ # is handled by all the engines.
190
+ engine_to_exclude = None
191
+
192
self.engines_running = True
193
self.stats_changed = True
194
self._send_start_wave(publish_back, self.current_wave,
@@ -203,22 +204,24 @@ class CoordinatorProc:
204
assert outputs.utility_output is None
205
206
eng_index = outputs.engine_index
206
- if outputs.scheduler_stats:
207
+ scheduler_stats = outputs.scheduler_stats
208
+ if scheduler_stats:
209
# 1. Updated request load stats - update our local
210
# state with these.
211
stats = self.engines[eng_index].request_counts
210
- stats[0] = outputs.scheduler_stats.num_waiting_reqs
211
- stats[1] = outputs.scheduler_stats.num_running_reqs
212
+ stats[0] = scheduler_stats.num_waiting_reqs
213
+ stats[1] = scheduler_stats.num_running_reqs
214
self.stats_changed = True
215
216
if (wave := outputs.wave_complete) is not None:
217
# 2. Notification from rank 0 engine that we've
218
# moved into the global paused state
217
- # (engines_running==False)
219
+ # (engines_running==False).
220
if self.current_wave <= wave:
221
+ new_wave = wave + 1
222
logger.debug("Moving DP wave from %d to %d.",
220
- self.current_wave, wave)
221
- self.current_wave = wave + 1
223
+ self.current_wave, new_wave)
224
+ self.current_wave = new_wave
225
self.engines_running = False
226
self.stats_changed = True
227
elif (wave := outputs.start_wave) is not None and (