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

Add sglang:decode_sum_seq_lens metric (#15066)

Browse Source
This commit is contained in:
fzyzcjy
2025-12-14 08:42:45 +08:00
committed by GitHub
parent fd37cc5d38
commit fdc93b019c
2 changed files with 9 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
python/sglang/srt/managers/scheduler_metrics_mixin.py
View File

@@ -317,6 +317,7 @@ class SchedulerMetricsMixin:
self.stats.swa_token_usage = swa_token_usage
if self.is_ssm_model:
self.stats.mamba_usage = mamba_usage
self.stats.decode_sum_seq_lens = batch.seq_lens_cpu.sum().item()
self.stats.gen_throughput = self.last_gen_throughput
self.stats.num_queue_reqs = len(self.waiting_queue)
self.stats.num_grammar_queue_reqs = len(self.grammar_queue)

8
python/sglang/srt/metrics/collector.py
View File

@@ -178,6 +178,7 @@ class SchedulerStats:
pending_prealloc_token_usage: float = 0.0
swa_token_usage: float = 0.0
mamba_usage: float = 0.0
decode_sum_seq_lens: int = 0
gen_throughput: float = 0.0
num_queue_reqs: int = 0
num_grammar_queue_reqs: int = 0
@@ -265,6 +266,12 @@ class SchedulerMetricsCollector:
labelnames=labels.keys(),
multiprocess_mode="mostrecent",
)
self.decode_sum_seq_lens = Gauge(
name="sglang:decode_sum_seq_lens",
documentation="The sum of all sequence lengths in decode.",
labelnames=labels.keys(),
multiprocess_mode="mostrecent",
)
self.gen_throughput = Gauge(
name="sglang:gen_throughput",
documentation="The generation throughput (token/s).",
@@ -641,6 +648,7 @@ class SchedulerMetricsCollector:
)
self._log_gauge(self.swa_token_usage, stats.swa_token_usage)
self._log_gauge(self.mamba_usage, stats.mamba_usage)
self._log_gauge(self.decode_sum_seq_lens, stats.decode_sum_seq_lens)
self._log_gauge(self.gen_throughput, stats.gen_throughput)
self._log_gauge(self.num_queue_reqs, stats.num_queue_reqs)
self._log_gauge(self.num_grammar_queue_reqs, stats.num_grammar_queue_reqs)