Skip to main content
← Lab

latencytail

confirmedC++ · llama.cppInference & servingJuly 6, 2026

the question

What does each concurrent request cost a local inference server?

what came out

Three regimes. Below the slot count, a batching tax -- per-token latency grows as the server decodes more active requests per step. Exactly at the slot count (C=4), a jitter spike: p99/p50 jumps to 2.24 (1.5B) and 2.67 (0.5B) versus ~1.2-1.3 elsewhere, the one concurrency where decode isn't smooth. Above it, queueing. Closed-loop C++ libcurl load generator.

method & receipts

  • Result: confirmed
  • Pre-registered the prediction was written down and committed before the run.
  • Reproducible one script re-runs the whole thing from scratch.
  • Independently reviewed an adversarial review pass, kept in the repo.
  • Machine-readable claims every reported number, in a checkable file.
  • Tested — a correctness/benchmark suite ships alongside the code.

→ read the code and re-run it

github.com/v-code01/latencytail