Google engineers: LLM benchmark harnesses silently drop requests — 200 QPS in, 38 out

AI Engineer · youtube · 2026-09-20

Google engineers Ashok Chandrasekar and Jason Kramberger show why many published LLM inference benchmarks don't reproduce: Python GIL-bound harnesses asked for 200 QPS quietly delivered 38; busy clients inflated measured latency by up to 58 seconds; a "20% throughput gain" came from temperature 0; and two harnesses tokenized the same dataset differently. Their fix is Inference Perf, a CNCF multiprocess load generator with client-side telemetry vs server metrics, verified at 5,000 QPS, plus a declarative workload catalog and the Prism UI from llm-d.

Original post →

More from Infra

Infra channel →