Skip to content

Versioning & How to Refresh

Baseline: vLLM 0.26.0 · model Qwen2.5-7B-Instruct · single RTX 4090 (24 GB)

This page is the site's maintenance stance, not a lesson. It explains what version everything is pinned to, why the numbers are illustrative, and the exact recipe for refreshing content when vLLM moves on.


vLLM iterates fast — APIs, flags, CLI parameters, and even source-code line numbers change between releases. To keep code that runs 1:1 on your machine, this site pins a single baseline version and annotates it at the top of every page (ADR-0004). This note tells you what that means as a reader, and gives a maintainer a checklist for keeping the site honest as vLLM advances.

What is pinned

Dimension Baseline Where it shows
vLLM v0.26.0 callout at the top of every page (ADR-0004)
Model Qwen2.5-7B-Instruct (quantized) baseline callouts; a few English-ecosystem examples cross-reference Llama-3.1-8B-Instruct (ADR-0001)
GPU single RTX 4090 (24 GB) baseline callouts (ADR-0001)
Source read-along file / symbol / line anchors align to the v0.26.0 tag Part 2 / 3 / 5 source-reading sections

A handful of sections that discuss a newer feature may annotate a higher version and call out the difference explicitly (ADR-0004). Always trust the callout at the top of the page you are on.

The docs toolchain is pinned too

The site is built with MkDocs Material + mkdocs-static-i18n + pymdown-extensions, all capped to their current major version in requirements.txt (e.g. mkdocs-material>=9.5,<10) so a fresh pip install reproduces today's rendering instead of silently pulling a breaking release. This matters right now: the Material team has announced that MkDocs 2.0 will be backward-incompatible — the plugin and theming systems are being rewritten with no migration path — so an uncapped upgrade would break the build. Treat a toolchain bump like a vLLM bump: change one dependency, run mkdocs build --strict, and eyeball a rendered page (math, admonitions, the language switch) before trusting it.

Why the numbers are illustrative

Following ADR-0004, verification here is static: the author does not execute Labs or reproduce the figures. Every performance number in the site is therefore an illustrative / order-of-magnitude reference — a shape to reason about, not a measurement to quote. You reproduce the real numbers yourself, on your own AutoDL box, for your own model and traffic.

Numbers that are pure arithmetic (KV-cache size, VRAM budgets, roofline arithmetic intensity) are exact — they are just multiplication, not benchmarks — and are labelled as such where they appear.

How facts are verified (dual-channel)

Every API/flag/CLI claim and every conceptual claim is checked against a primary source before it lands — never written from memory:

  • Context7 → vLLM (and related libraries') API, flags, CLI — the concrete signatures you type.
  • Sonarconcepts, papers, and specifications — algorithm provenance, number-format specs, hardware parameters.

This verification stays static: no code is executed and no figures are reproduced (ADR-0004). It catches hallucinated flags and misattributed algorithms at the source; it does not turn illustrative numbers into measured ones.

When vLLM upgrades — the refresh recipe

A major vLLM release will gradually make the baseline examples stale. When you decide to bump the baseline (a deliberate, whole-site action — ADR-0004 keeps re-verification deferred, not automated), work through this checklist:

  1. Bump the version annotation — the page-top baseline callout on affected pages, and the Baselines table on the Home page.
  2. Re-verify every touched flag / API / CLI via Context7 against the new version's official docs. Assume nothing carried over.
  3. Re-verify concept / paper / spec claims via Sonar for any section you deepen or change.
  4. Re-align source read-along anchors — the file / symbol / line references in Part 2 / 3 / 5 must point at the new release tag, not the old line numbers.
  5. Re-run the Labs to recalibrate numbers you care about — on your AutoDL box, following the Capstone measurement discipline (greedy, fixed seed, stated eval set). Update figures only where you re-measured; leave the rest labelled illustrative.
  6. Keep the build greenmkdocs build --strict is the one structural gate; broken links, orphan pages, and i18n gaps all fail it.

The stance is honest but lightweight: annotate the version, verify what you touch, and defer a full re-verification sweep rather than standing up CI too early.

Found drift? Open a refresh issue

If you hit a flag that no longer exists, code that won't run, or a number that feels off for the current release, open a "Content refresh / vLLM version upgrade" issue from the repository's issue templates. The template walks through the same checklist above so nothing is missed.

See also

  • Home · Baselines — the one-glance baseline table and the "on the numbers" stance.
  • Capstone — the measurement discipline to follow when you recalibrate numbers.
  • Glossary — the bilingual terminology every page reuses.