⚡ LLM Bench

Log in

Admin password — it's in config.json on the server (also printed in the service journal).

Jobs — concurrency sweep & schedule

levels = concurrency sweep, one level after another (e.g. 1,2,5,10,20). Each level runs stage seconds. interval = seconds from the end of one sweep to the next (345600 = 4 days, the old k6 loop). Sweeps never overlap; due jobs queue.

Targets & defaults

Results — TTFT & TPS

Time to first token (TTFT p50, dashed = p95)

Per-stream tokens/sec (after first token)