Skip to content

faq: memory, 128 GB and Windows answered; NPU and Lemonade get searcher titles - #25

Merged
bong-water-water-bong merged 1 commit into
mainfrom
seo/faq-memory
Nov 6, 2026
Merged

bong-water-water-bong merged 1 commit into
mainfrom
seo/faq-memory

Conversation

@bong-water-water-bong

Copy link
Copy Markdown
Contributor

Brief items 2 and 4: the FAQ's missing purchase questions, and the two pages whose search text lives in PAGE_META rather than in a file of their own.

FAQ: three questions added, each two lines

Directly after Which hardware does it run on?, because that is where a reader starts asking them:

question answer, in short new in 28 days?
How much memory does a model need? its file size plus the context you give it; the CPU, GPU and NPU share one 128 GB pool, so there is no second copy for the second device; the two published small rows are 27B ternary at 6.7 GiB and 2B ternary at 1.11 GiB yes — the page never said this
Which models fit in 128 GB? every row of the measured table, up to 27B-class at 4-bit; a bigger mixture-of-experts model streams experts from NVMe (moe-streaming) yes
Does it run on Windows? no: built and measured on Linux, no Windows package (releases); Lemonade Server has its own Windows backends, so on Windows you get Lemonade without the engine (lemonade) yes

No jab at anyone: the Windows answer is the same sentence docs/lemonade.md already uses ("no Windows package", "Lemonade ships its own backends for … Windows"). The NPU question the brief asks for already existed.

These become FAQPage JSON-LD with no code change — faq_items() already turns every ### Question in a ## Frequently asked questions section into one, so the JSON-LD now carries 15 questions (verified by parsing the rendered script tag).

PAGE_META: two titles that described the engine to itself

page was now
npu.html AMD NPU on Linux: the XDNA 2 NPU engine Ryzen AI NPU on Linux for local LLMs (XDNA 2) — description leads with the hardware and the models it runs today instead of "our own runtime with full-ELF kernel sets"
lemonade.html Lemonade Server backend: the onebit recipe Lemonade Server on Strix Halo with 1bit engine — description names Strix Halo and the iGPU route the searcher is asking about

Both keep the site's limits: title + · 1bit engine ≤ 60 characters, description ≤ 160 (both enforced by the build).

Verified at 24246b2

ONEBIT_ENGINE_DIR at engine main: 57 pages, rc 0; faq.html, npu.html, lemonade.html each exactly one <h1>; the new titles/descriptions are in the rendered <head>; FAQPage JSON-LD parsed back out with 15 mainEntity questions including the three new ones. Site main eb5b721, not behind.

Two files, no behaviour change.

…er titles

Owner brief TDOCS-SEO, items 2 and 4. The FAQ had "does it use the NPU" but
none of the three questions that decide a purchase:

- "How much memory does a model need?" — file size plus context, one 128 GB
  pool shared by CPU, GPU and NPU, with the two already-published small rows
  (27B ternary 6.7 GiB, 2B ternary 1.11 GiB)
- "Which models fit in 128 GB?" — every measured row, up to 27B at 4-bit; a
  bigger MoE model streams experts from NVMe
- "Does it run on Windows?" — no Windows package, and Lemonade's own Windows
  backends are Lemonade's (docs/lemonade.md), stated without a jab

Each answer is two lines and links the page with the detail. All three become
FAQPage JSON-LD automatically — the section already went through faq_items().

Same PR, the two pages whose search text lives in PAGE_META rather than in a
file of their own:

- npu: "AMD NPU on Linux: the XDNA 2 NPU engine" -> "Ryzen AI NPU on Linux for
  local LLMs (XDNA 2)", description leading with the hardware and what it runs
  today instead of "our own runtime with full-ELF kernel sets"
- lemonade: "Lemonade Server backend: the onebit recipe" -> "Lemonade Server on
  Strix Halo with 1bit engine", description naming Strix Halo and the iGPU
  route a searcher is asking about

Built at this head: 57 pages rc 0, one h1 on all three pages, FAQPage JSON-LD
carries 15 questions, every description within the 160-character limit.
@bong-water-water-bong

Copy link
Copy Markdown
Contributor Author

@pi-prs ready for review at 24246b2. TDOCS-SEO (owner: findable before the release). Built green here: 58 pages rc 0, one

, meta within limits. faq

@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 24246b2a.

Built with the CI pin (markdown==3.10.3) plus #23's heading guard: 57 pages -> _site, rc 0. The two PAGE_META
edits are inside the tool's own limits, measured on the rendered pages rather than on the source strings:

npu        title 45 -> 59 rendered | desc 152 -> 157 rendered   (limits 60 / 160)
lemonade   title 46 -> 60 rendered | desc 137                   (limits 60 / 160)

Both re-wordings are honest about what the page now says: npu leads with the query people actually type ("Ryzen AI NPU
on Linux for local LLMs") and drops the "what matches the reference today" phrasing for "what is still in progress", which
is what the page says; lemonade leads with "Lemonade Server on Strix Halo with 1bit engine" instead of "the onebit
recipe", which is the thing being searched for.

The three new FAQ answers are consistent with the site's own data, checked one by one:

  • memory: "one 128 GB pool… no second copy for the second device" — the unified-memory statement the Strix Halo page
    already makes, and the sticky detail (file size plus context) is the one people get wrong.
  • which models fit: "covers every row of the measured table, up to 27B-class models at 4-bit" — models.html's seven
    rows top out at Qwen3.8-27B UD-Q4_K_XL, so "every row" holds; and the MoE caveat is the next sentence, with the
    streaming page linked.
  • Windows: "No… there is no Windows package" — no Windows artifact exists in releases.html, and the answer correctly
    separates Lemonade-on-Windows from the engine.

0 broken internal links across the whole build, including models.html, moe-streaming.html and releases.html from
the new answers. Nothing here needs a device and no file outside site/pages/faq.md and the two PAGE_META entries moves.

@bong-water-water-bong
bong-water-water-bong merged commit 2f847fc 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