Tiny add command line args for prefill delayer and unify names (#16830)

This commit is contained in:
fzyzcjy
2026-01-10 07:54:25 +08:00
committed by GitHub
parent ef35d8fe4e
commit cec19b56c6
6 changed files with 81 additions and 35 deletions

View File

@@ -271,6 +271,7 @@ class SchedulerMetricsCollector:
self,
labels: Dict[str, str],
enable_lora: bool = False,
prefill_delayer_max_delay_passes: int = 30,
) -> None:
# We need to import prometheus_client after setting the env variable `PROMETHEUS_MULTIPROC_DIR`
from prometheus_client import Counter, Gauge, Histogram, Summary
@@ -761,13 +762,12 @@ class SchedulerMetricsCollector:
labelnames=list(labels.keys()) + ["category", "num_prefill_ranks"],
)
max_delay_passes = envs.SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES.get()
self.prefill_delayer_wait_forward_passes = Histogram(
name="sglang:prefill_delayer_wait_forward_passes",
documentation="Histogram of forward passes waited by prefill delayer.",
labelnames=labels.keys(),
# Need bucket "<=0" for zero-delay cases
buckets=[0, 5, 20, max_delay_passes - 1],
buckets=[0, 5, 20, prefill_delayer_max_delay_passes - 1],
)
self.prefill_delayer_wait_seconds = Histogram(
name="sglang:prefill_delayer_wait_seconds",