1bit.MONSTER — Documentation

Model-agnostic NPU + GPU + CPU inference engine for AMD Strix Halo. Pure C++23, zero Python at runtime, MIT.

Start here

Doc	What it covers
journey.md	The hero story — reverse-engineering the XDNA 2 NPU in 4 days; full timeline 2026-04-28 → today
CODEBASE.md	Code-level reference — directory map, build targets, backend catalog, engines, kernels, tools
audit-trail.md	1.5 TB of evidence — where the raw data lives, how to verify any claim, repo↔archive mapping
guides/getting-started.md	Install, build, first inference
guides/building.md	Full build guide (CMake, ROCm/TheRock, targets)
guides/architecture.md	Engine internals: loaders, backends, routing
model-families/	Every model family, its own page — full breakdowns
jarvis.md	The flagship voice pipeline (STT → LLM → TTS → voice)
guides/roadmap.md	Where the engine is headed

Model families

1bit auto-detects 552 architecture tokens (1,774 HF arch strings in rcpp_arch_from_string). Each family has a dedicated page with full data (params, 1BP size, backends, measured perf) — start at the model-families index. The Zyphra family is the featured end-to-end ecosystem; wiki/models.md remains the combined support SSOT.

Guides & platform

guides/windows.md — Windows notes
guides/Lemonade-Compat.md — Lemonade compatibility
guides/launch.md — launch/serving

Research & technical deep-dives

research/npu-engine.md · research/npu-dynamic-instr.md — Native XDNA 2 NPU engine
research/npu-ternary-roadmap.md — Ternary/binary NPU kernel roadmap
research/fastflowlm-decode/SUMMARY.md — FastFlowLM reverse-engineering report (historical — fully replaced)
research/fastflowlm-analysis/ — Raw binary analysis, xclbin captures, instruction traces
research/block-scaled-ternary-format.md — 1BP / TQ2 ternary storage
research/kernel-analysis.md · research/hybrid-w4a8-router.md — Kernels & routing
research/vision-module.md — Vision-language support
research/flux-feasibility.md — Flux image-model feasibility notes
research/npu/AMD-XDNA-40COLUMN-UNLOCK.md — NPU column-unlock deep dive

Validation & status

Wiki

Deeper operational docs live in wiki/: models (single source of truth for supported model families), performance (single source of truth for benchmark numbers), Installation, npu-architecture, boot-configuration, Network-Topology.

Goals & vision (not part of the engine)

Aspirational product ideas built on the engine — kept separate so they don't drive the project narrative: goals/.