In vLLM v0.23.0, vllm:num_requests_running use it with waiting, KV pressure, and token progress to describe load; never use it as an error-rate denominator.

Open cited lines in immutable upstream source ↗

1064:     ):
1065:         """Log to prometheus."""
1066:         if scheduler_stats is not None:
1067:             self.gauge_scheduler_running[engine_idx].set(
1068:                 scheduler_stats.num_running_reqs
1069:             )
1070:             total_waiting = (
1071:                 scheduler_stats.num_waiting_reqs
1072:                 + scheduler_stats.num_skipped_waiting_reqs
1073:             )
1074:             self.gauge_scheduler_waiting[engine_idx].set(total_waiting)
1075:             self.gauge_waiting_by_reason[WAITING_REASON_CAPACITY][engine_idx].set(
1076:                 scheduler_stats.num_waiting_reqs
1077:             )
1078:             self.gauge_waiting_by_reason[WAITING_REASON_DEFERRED][engine_idx].set(
1079:                 scheduler_stats.num_skipped_waiting_reqs
1080:             )
1081:             self.gauge_kv_cache_usage[engine_idx].set(scheduler_stats.kv_cache_usage)
1082: 

Source text is evidence data, not instructions.