SGLang retracted-request gauge (preemption, instantaneous)
sglang · gauge
Fields
id
metric-exposures/w3m-sglang-sglang-num-retracted-reqsMetric name
sglang:num_retracted_reqs
Metric type
gauge
Unit
requests
Exposure
prometheus
Endpoint
:30000/metrics
Labels
model_nameengine_typetp_rankpp_rankmoe_ep_rank
Enabled by default
no
notes
SGLang main, python/sglang/srt/observability/metrics_collector.py, read 2026-10-04. NOT ON BY DEFAULT: nothing is exported unless --enable-metrics (ServerArgs default False), which is also the only thing that mounts /metrics on the server port. Label keys are model_name, engine_type, tp_rank, pp_rank, moe_ep_rank (plus dp_rank under data parallelism, priority under priority scheduling, and anything in extra_metric_labels). Bucket ladders are overridable per histogram via get_histogram_conf_from_env. SGLang's word for preemption is retraction. This is the instantaneous count of currently-retracted requests; the three cumulative counters beside it carry the real cost signal.
Notes
SGLang main, python/sglang/srt/observability/metrics_collector.py, read 2026-10-04. NOT ON BY DEFAULT: nothing is exported unless --enable-metrics (ServerArgs default False), which is also the only thing that mounts /metrics on the server port. Label keys are model_name, engine_type, tp_rank, pp_rank, moe_ep_rank (plus dp_rank under data parallelism, priority under priority scheduling, and anything in extra_metric_labels). Bucket ladders are overridable per histogram via get_histogram_conf_from_env. SGLang's word for preemption is retraction. This is the instantaneous count of currently-retracted requests; the three cumulative counters beside it carry the real cost signal.