Skip to content

strix-halo: lead with the searcher's query, and add Start here / What to expect - #24

Merged
bong-water-water-bong merged 1 commit into
mainfrom
seo/strix-halo
Nov 6, 2026
Merged

bong-water-water-bong merged 1 commit into
mainfrom
seo/strix-halo

Conversation

@bong-water-water-bong

Copy link
Copy Markdown
Contributor

Target query: "run local LLMs on AMD Strix Halo (Ryzen AI MAX+ 395) on Linux" — the highest-intent search a Strix Halo owner makes, and strix-halo.html had 0 impressions in 28 days (Search Console, to 2026-11-04) while the home page did 75.

What the page already had, and what it did not

It had the hardware, the two devices and the measured table. It did not have how to start (two bullets at the very end) or what to expect of memory, which is the first question a 128 GB unified-memory machine raises. Its search title was Strix Halo LLM guide: Ryzen AI MAX+ 395, Linux — a title for people who already know what Strix Halo is.

Changes

where change
title: → Local LLMs on Strix Halo: Ryzen AI MAX+ 395 (43 chars; the build fails the build at 46 + the · 1bit engine suffix, so the brief's full "Run local LLMs on AMD Strix Halo (Ryzen AI MAX+ 395) on Linux" is the h1, where it already is)
description: names the query verbatim, then iGPU / NPU / Lemonade / where to start — 150 of the 160 characters the build enforces
new ## Start here the three routes (boot the stick → os-v3, add to the Lemonade you run → lemonade, 1bit serve → serve + models), each with its page; carries the positioning line (one binary, model agnostic, the stick is the appliance)
new ## What to expect unified 128 GB: fits by file size, not per-device copies, context on top; the two already published small-row sizes (27B ternary 6.7 GiB, 2B ternary 1.11 GiB); MoE streaming from NVMe for a model bigger than memory; and that the NPU route is in progress, a second device rather than a GPU replacement
## Ways to run it removed — Start here now carries it; the closing paragraph becomes ## Where it ships with the release/FAQ pointer and no release text

No new numbers: 6.7 GiB / 1.11 GiB are docs/models.md's own rows, and the token rates stay where they were.

Inside the brief's rules

  • one h1 — the build's check passes (# Local LLMs on AMD Strix Halo (Ryzen AI MAX+ 395) on Linux).
  • internal links to models / serve / lemonade / npu / faq and from every page: the landing pages sit in the docs sidebar of every doc page, so models.html and serve.html link back to this one.
  • no /install.html link (docs/install.md is engine #859, held): the stick route links os-v3.html instead.
  • no release text, no kernel internals, no unpublished numbers.

Verified at a263962

ONEBIT_ENGINE_DIR at engine main: 57 pages, rc 0; strix-halo.html has one <h1>, the new title + description, the three new sections and 0 Ways to run it; every .html href on the page resolves in _site (42 checked, 0 missing). Site main is eb5b721, branch is not behind.

…it lacked

Owner brief TDOCS-SEO: the Strix Halo page must answer "run local LLMs on AMD
Strix Halo (Ryzen AI MAX+ 395) on Linux", with the hardware, what runs where,
what to expect and how to start. The page had the hardware and the two devices,
and left "how to start" to two bullets at the very end.

- title: "Strix Halo LLM guide: Ryzen AI MAX+ 395, Linux" -> "Local LLMs on
  Strix Halo: Ryzen AI MAX+ 395" (the site caps a search title at 46 characters
  before its " · 1bit engine" suffix, so the brief's longer wording is the h1,
  which it already is)
- description names the query verbatim and stays at 150 of the 160 characters
  the build enforces
- new "Start here": the stick, the Lemonade you already run, or 1bit serve,
  each with its page, plus the positioning line
- new "What to expect": unified 128 GB means a model fits by file size, the
  two published small-row sizes, MoE streaming for the rest, and that the NPU
  route is in progress rather than a GPU replacement
- the old "Ways to run it" bullets are gone (Start here carries them); the
  release/FAQ pointer stays as "Where it ships", with no release text

Built at this head: 57 pages, rc 0, one h1, every .html link on the page resolves
in _site.
@bong-water-water-bong

Copy link
Copy Markdown
Contributor Author

@pi-prs ready for review at a2639621. TDOCS-SEO (owner: findable before the release). Built green here: 58 pages rc 0, one

, meta within limits. strix-halo:

@bong-water-water-bong bong-water-water-bong left a comment

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

VERDICT: READY at a2639621.

Built with the CI's own pin and the heading guard (markdown==3.10.3, engine at prs1243): 57 pages -> _site, rc 0 —
so the title/description gates, the one-<h1> rule and #23's heading-count guard all pass. Rendered meta:

strix-halo   title 57 | desc 150   Local LLMs on Strix Halo: Ryzen AI MAX+ 395 · 1bit engine     (max 60/160)

Every new internal link resolves, checked against the built tree rather than by eye: 0 broken .html targets across
all 57 pages, including the two this section adds (blog-boot-a-stick.html, moe-streaming.html).

The two figures in the new copy are the site's own, from models.html: "PrismML Ternary-Bonsai-2-27B — PQ2_0,
6.7 GiB, HRX" and "Microsoft BitNet b1.58 2B-4T — TQ2_0 ternary GGUF, 1.11 GiB". So "a 27B ternary model at 6.7 GiB
and a 2B ternary model at 1.11 GiB" is quoted, not invented.

And the claim I went looking to flag turns out to be sourced: "every ggml weight type loads" is docs/models.md:22's
own sentence ("Every ggml weight type runs on HRX, including NVFP4, Q2_0, MXFP4 and the ternary TQ, PTQ and PQ…"), so the
page is not outrunning the docs. The "NPU route is in progress — correctness first, speed next" framing matches
docs/npu.md and PORTING.md.

One observation, not a request: npu.html and lemonade.html render descriptions of 170 and 165 characters, over
the tool's own 160 limit — but that is not this PR (it does not touch those PAGE_META entries); those two descriptions
come with the page content. Worth a separate one-line fix, because a search result truncates around 160.

@bong-water-water-bong
bong-water-water-bong merged commit 3013e6d into main Nov 6, 2026
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant