Documentation

Two products, fully documented

Everything below is the complete reference: manuals, setup guides, telemetry wiring, the metrics methodology and the honest list of limitations. No frills — the same documents the projects ship with, rendered for reading.

VeloBenchmark

The single-binary benchmarking and live-stats console for OpenAI-compatible endpoints.

Overview & quick start — install, first run, the whole loop
User manual — every screen, every report
Building & installing — source, service, releases
Telemetry setup — OTLP receiver + engine flags
Metrics & methodology — what each number means
Known limitations — the honest edges

veloGB10

The Rust + CUDA inference engine for NVIDIA GB10 systems (DGX Spark & compatible).

Overview — design, models, benchmarks, CLI reference
Qwen3.8-27B setup — single node, TP=2, TP=4
Hy3 setup — two-node TP=2 bring-up
TP cache management — the node blob cache
Endurance report — 8-hour mixed workload
Changelog — release-by-release highlights