Skip to content

Benchmark

Reproducible UCM benchmark methodology, tooling, and reference results. This page is a placeholder — populate it per the spec below.

What to add

Reader goal: how much faster is UCM, and how do I reproduce the number myself?

Required content:

  • Method: test scenario design (prefix cache hit rate, concurrency, input/output length), hardware, model, framework version. Reference the test design already in GLM-5.1 4-node PD.
  • Usage: how to run — script/command, dataset, parameters, reproduction steps.
  • Data: standard result tables (throughput token/s, TTFT, TPOT, speedup), including comparison vs the "no-UCM full compute" baseline.
  • Comparison: cross-comparison tables across stores / engines / models to aid selection.

Don't:

  • Don't duplicate all raw data from detail pages here; this is an aggregation entry that links back to them.
  • Don't mark a result "verified" unless there is a reproducible script + hardware record (review line 275).

Acceptance:

  • A reader can reproduce one result by following the page.
  • Result tables label their source and whether they are verified.

Owner: (to be assigned)

Reference