chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Add cuda_graph_forward_passes_total and num_retracted_reqs_total (#15189)

Browse Source
This commit is contained in:
fzyzcjy
2025-12-17 21:19:42 +08:00
committed by GitHub
parent 6c4bf8a0be
commit b12b40dea4
3 changed files with 25 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/metrics/collector.py
+21
View File
@@ -325,11 +325,18 @@ class SchedulerMetricsCollector:
)
# Retract
# TODO maybe remove this old gauge in favor of the new counter
self.num_retracted_reqs = Gauge(
name="sglang:num_retracted_reqs",
documentation="The number of retracted requests.",
labelnames=labels.keys(),
)
self.num_retracted_reqs_total = Counter(
# The name is `requests` instead of `reqs` to avoid dup name error
name="sglang:num_retracted_requests_total",
documentation="Total number of retracted requests.",
labelnames=labels.keys(),
)
self.num_paused_reqs = Gauge(
name="sglang:num_paused_reqs",
documentation="The number of paused requests by async weight sync.",
@@ -605,12 +612,18 @@ class SchedulerMetricsCollector:
labelnames=list(labels.keys()) + ["stage"],
)
# TODO maybe remove this old gauge in favor of the new counter
self.is_cuda_graph = Gauge(
name="sglang:is_cuda_graph",
documentation="Whether the batch is using CUDA graph.",
labelnames=labels.keys(),
multiprocess_mode="mostrecent",
)
self.cuda_graph_passes_total = Counter(
name="sglang:cuda_graph_passes_total",
documentation="Total number of forward passes categorized by CUDA graph.",
labelnames=list(labels.keys()) + ["mode"],
)
self.new_token_ratio = Gauge(
name="sglang:new_token_ratio",
@@ -639,6 +652,14 @@ class SchedulerMetricsCollector:
def observe_queue_time(self, latency: float) -> None:
self._log_histogram(self.queue_time, latency)
def increment_num_retracted_reqs(self, num: int) -> None:
self.num_retracted_reqs_total.labels(**self.labels).inc(num)
def increment_cuda_graph_pass(self, value: bool) -> None:
# leave room for piecewise cuda graph, etc
mode = "decode_cuda_graph" if value else "decode_none"
self.cuda_graph_passes_total.labels(**self.labels, mode=mode).inc(1)
def log_stats(self, stats: SchedulerStats) -> None:
self._log_gauge(self.num_running_reqs, stats.num_running_reqs)
self._log_gauge(self.num_used_tokens, stats.num_used_tokens)