Inspiration Notes

These repositories are architecture references only. Kipp must not vendor or copy their code. Before adapting anything beyond a general idea, inspect the source file and its repository license, preserve required notices, and record the provenance explicitly.

antirez/ds4

Study the narrow, model-specific native-engine structure; readable C core; isolated Metal and CUDA paths; mmap-oriented loading; KV-cache design; and model-aware quantization decisions.

ggml-org/llama.cpp and ggml-org/ggml

Study GGUF metadata and tensor layout, quantization formats and kernels, backend boundaries, and portable low-level tensor representations. Kipp will support only the subset required by its chosen model.

ggml-org/whisper.cpp

Study how a deliberately narrow C engine keeps model loading, inference, and platform acceleration understandable without becoming a generic runtime.

karpathy/llm.c

Study naming, file organization, direct C/CUDA implementations, and the use of small reference paths to make low-level numerical code approachable.

GeeeekExplorer/nano-vllm

Study the concepts behind paged attention, request lifecycle, continuous batching, and KV-cache ownership. Reimplement any selected ideas natively.

vllm-project/vllm and sgl-project/sglang

Study scheduler behavior, memory management, prefix reuse, and serving semantics. Treat their breadth as research material, not Kipp's initial scope.

Research workflow

Once Nia is configured manually, index every repository listed in .cursor/commands/nia.md. Notes derived from source research should name the repository, file, revision, and applicable license.