№ 0917GitHub
The Muser book: how to write an inference engine
A 40-chapter book on Muser, an engine that runs Muse Glimmer on Apple Silicon Metal. It covers kquant and DFlash speculative lanes, exact KV-cache replay with kvpack, and GB10 NVFP4 prefill handed off to a Mac, with every number tied to an evidence receipt.
muser-book
How to Write an Inference Engine — a zero-to-hero book about Muser: how
one token of Muse Glimmer is generated on Apple Silicon, how the kquant and
DFlash speculative lanes work, how kvpack saves and replays KV-cache state
exactly, and how a GB10 node prefills NVFP4 and hands the KV to a Mac over
an authenticated transport.
40 chapters, a 424-term glossary, a 322-entry bibliography, and two
appendices (kernel table, environment flags) — every measured number cited
to a retained evidence receipt.
Read it
The rendered book publishes to GitHub Pages from this repository's main
branch: https://highperformanceailab.com/muser-book/.
Build it locally
cargo install mdbook mdbook-mermaid # or brew install mdbook mdbook-mermaid
mdbook build # output in book/
mdbook serve # http://localhost:3000
python3 scripts/check_social_meta.py # verify rendered social metadata
How the book keeps itself honest
- Chapters are written against pinned revisions of the source trees
(src/PINNED.md); a re-pin is an explicit event, never a
silent follow. - Every measured number carries a
[receipt …]citation to the append-only
evidence volume; nothing is rounded into existence. - Style rules and citation notation live in
src/STYLE.md.
The book is part of the Muser program:
- muser — the engine
- kvpack — exact KV-cache replay
- muser-console — the dashboard
License
MIT OR Apache-2.0, at your option (see LICENSE-MIT and LICENSE-APACHE).
The bundled src/mermaid.min.js is MIT-licensed
(mermaid-js) and carries its own
license header.
