Fix num_running_reqs gauge on disagg prefill servers (#31495)

Co-authored-by: Zhiqiang Xie <xiezhq@stanford.edu>
This commit is contained in:
cctry
2026-07-16 15:43:06 -07:00
committed by GitHub
co-authored by Zhiqiang Xie
parent 640101da13
commit d28e35b1a1
@@ -635,7 +635,17 @@ class SchedulerMetricsReporter:
)
# Basics
self.stats.num_running_reqs = prefill_stats.num_running_reqs
if (
self.scheduler.disaggregation_mode == DisaggregationMode.PREFILL
and batch is not None
):
# running_batch is never populated in PD prefill, so count the
# current forward batch instead.
self.stats.num_running_reqs = QueueCount.from_reqs(
batch.reqs, priority_enabled
)
else:
self.stats.num_running_reqs = prefill_stats.num_running_reqs
self.stats.num_queue_reqs = QueueCount.from_reqs(
self.scheduler.waiting_queue, priority_enabled
)