Architecture

The model, the words, and the presets

Start with the system map — CONTEXT.md is written to be read end to end and to be enough to hold the system in your head. Then the model walk for the actual dataflow, and the glossary for the vocabulary.

The glossary is not optional reading for this section. “fused”, “arm”, “sidecar”, “iteration”, “group” and “Engram” each have more than one live meaning in this repo, and a reader who does not hold the glossary will misread every other page — including this one. The glossary also ends with a live list of the places where a document and the code disagree about a name or a default, which is the fastest way to find out whether a page you are reading is current.

• dormouse bf16 — all-bf16, no fp32 waste — docs/architecture/bf16-plan.md
• The system map — CONTEXT.md
• design-minimal: the floor of dormouse (2026-09-25) — docs/architecture/design-minimal.md
• f16 plan — docs/architecture/f16-plan.md
• dormouse glossary — docs/glossary.md
• The 28 crates of vendor/dormouse-fused: what each one is for — docs/architecture/library-crate-fate.md
• Mixture arms: the anchors, acted on — docs/architecture/mixture-arms.md
• The model walk — README.md
• Multi-GPU — docs/architecture/multi-gpu.md
• dormouse post-training — self-evolving + EGGROLL — docs/architecture/post-training.md
• TK design on cubecl — docs/architecture/tk-design-on-cubecl-2026-10-04.md