SGLang generation throughput gauge (token/s)

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. 'The generation throughput (token/s).' This is the ONE engine in this set that ships a ready-made throughput GAUGE rather than cumulative token counters - which means it is a sampled level at the logging interval, not a rate() over a counter, and the two are not numerically identical.