← back · updates every 4 s, no reload

Live

waiting for feed…

vLLM throughput

gen tok/s

—

prompt tok/s

—

running reqs

—

kv cache %

—

prefix hit %

—

Active run — task streams

click a row to expand

completed this run

Polls /queue_state.json (orch scheduler writes it ~2 s) and /live_feed.json (bench/live_commands.sh writes it ~4 s for the active model). If the orchestrator isn't running the queue goes empty, but the per-task feed continues if a model is still up.