zen5-engine/LLM.md

zen5 — engine fork of antirez/ds4

Path: /Users/a/work/zen/zen5 Purpose: zenlm-branded inference engine for the zen5 model line. Fork of antirez/ds4 (DwarfStar 4 — DeepSeek V4 Flash native engine).

State

• MIT-licensed fork. Upstream copyrights preserved in LICENSE and NOTICE.md.
• Git remote: upstream → antirez/ds4 (no origin yet — set when the zenlm fork repo exists).
• Build: make on macOS Metal produces ds4, ds4-server, ds4-bench plus zen5* symlinks.
• Model weights are NOT in this repo. Use ./download_model.sh q2-imatrix to fetch ~81 GB of GGUF into ./gguf/.

Quick orient

File	What it is
README.md	zen5-branded readme
README-UPSTREAM.md	original antirez/ds4 README, verbatim
NOTICE.md	attribution, fork policy, how to track upstream
LICENSE	MIT — ds4.c authors + ggml authors (unchanged)
Makefile	Metal / CUDA / CPU builds
ds4.c (800 KB)	core inference loop
ds4_metal.m (640 KB)	Metal backend
ds4_cuda.cu (465 KB)	CUDA backend
ds4_server.c (550 KB)	OpenAI/Anthropic/Responses HTTP server
ds4_cli.c (52 KB)	interactive CLI with linenoise
gguf-tools/	offline GGUF gen, imatrix collection, quality testing
dir-steering/	activation-direction steering data
speed-bench/	benchmark CSVs and plots
tests/test-vectors/	official V4 Flash continuation vectors

Sibling context

• ~/work/zen/LLM.md — zen ecosystem index.
• ~/work/zen/models/ — model weight directories (zen4, zen4-nano, etc.).
• Future zenlm origin for this fork: TBD (likely zenlm/zen5).