chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

Tiny unify realtime_tokens_total metric (#15747)

Browse Source
This commit is contained in:
fzyzcjy
2025-12-24 21:29:43 +08:00
committed by GitHub
parent d5fbbfd988
commit b3b818fd86
1 changed files with 10 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
python/sglang/srt/metrics/collector.py
View File

@@ -669,22 +669,11 @@ class SchedulerMetricsCollector:
multiprocess_mode="mostrecent",
)
self.realtime_prefill_compute_tokens_total = Counter(
name="sglang:realtime_prefill_compute_tokens_total",
documentation="Total number of prefill compute tokens processed (updated on each log interval).",
labelnames=labels.keys(),
self.realtime_tokens_total = Counter(
name="sglang:realtime_tokens_total",
documentation="Total number of tokens processed (updated on each log interval).",
labelnames=list(labels.keys()) + ["mode"],
)
self.realtime_prefill_cache_tokens_total = Counter(
name="sglang:realtime_prefill_cache_tokens_total",
documentation="Total number of prefill cache tokens processed (updated on each log interval).",
labelnames=labels.keys(),
)
self.realtime_decode_tokens_total = Counter(
name="sglang:realtime_decode_tokens_total",
documentation="Total number of decode tokens processed (updated on each log interval).",
labelnames=labels.keys(),
)
self.gpu_execution_seconds_total = Counter(
name="sglang:gpu_execution_seconds_total",
documentation="Total time that GPU is busy executing a workload.",
@@ -729,13 +718,12 @@ class SchedulerMetricsCollector:
def increment_realtime_tokens(
self, prefill_compute_tokens=0, prefill_cache_tokens=0, decode_tokens=0
):
self.realtime_prefill_compute_tokens_total.labels(**self.labels).inc(
prefill_compute_tokens
)
self.realtime_prefill_cache_tokens_total.labels(**self.labels).inc(
prefill_cache_tokens
)
self.realtime_decode_tokens_total.labels(**self.labels).inc(decode_tokens)
for mode, delta in [
("prefill_compute", prefill_compute_tokens),
("prefill_cache", prefill_cache_tokens),
("decode", decode_tokens),
]:
self.realtime_tokens_total.labels(**self.labels, mode=mode).inc(delta)
def increment_gpu_execution_seconds(self, category: str, t: float):
logger.debug(f"GPU execution seconds: {category=} {t=:.3f}")