Skip to content

Latest commit

 

History

History
98 lines (74 loc) · 4.67 KB

File metadata and controls

98 lines (74 loc) · 4.67 KB

Acknowledgments

Hebrus is the result of an open-source lineage, not a clean-room rewrite. This document records the principal projects and implementations that materially shaped the current repository. It describes technical provenance; it does not claim endorsement by any upstream author, project, company, or model provider.

Project origin: antirez/ds4

Hebrus began as a fork of antirez/ds4 and retains substantial core-engine implementation, architecture, utilities, and Git history from that project. Salvatore Sanfilippo (antirez) and the upstream contributors created the foundation on which this fork continues to build.

The fork has since concentrated on Apple Metal execution, embedded ExpertMajor stores, adaptive residency, SSD streaming, additional qualified model families, and the evidence and tooling required to maintain those paths. The repositories have diverged, but the contribution policy still asks whether general, reproducible improvements should also be proposed upstream. See FORK_NOTES.md for the current boundary.

GGML and llama.cpp

llama.cpp and GGML are essential engineering references for this repository. Their work informs GGUF handling, quantization formats and tables, CPU quantized dot products, Metal kernels, converters, test methods, and many implementation decisions.

Hebrus does not link against GGML or llama.cpp as a runtime dependency. Some source-level pieces are retained or adapted under the MIT license, including quant layouts and tables, CPU quant/dot logic, and selected kernel techniques. The GGML authors' copyright notice remains in LICENSE, and source-specific attribution remains in the relevant files.

Pinned llama.cpp converter implementations are also used as independent Qwen validation references. A reference implementation or fixture oracle is not the same thing as a linked dependency.

MLX and MLX-LM

MLX and MLX-LM informed the current Qwen MLX affine4/group-64 storage layout, scheduling comparisons, and validation work. The published Qwen release repacks routed weights from an MLX-format source into the existing checksummed ExpertMajor v2 GGUF container.

Hebrus does not link the MLX libraries at runtime. This credit does not imply Apple authorship, sponsorship, or endorsement of Hebrus.

YaRN

The Metal RoPE implementation in metal/dsv4_rope.metal contains an algorithm based on YaRN, with the source header's MIT attribution to Jeffrey Quesnelle and Bowen Peng preserved.

Bundled utilities

The repository includes:

  • linenoise, by Salvatore Sanfilippo and Pieter Noordhuis, for terminal line editing;
  • Rax, by Salvatore Sanfilippo, for radix-tree storage used by the server path.

These files carry BSD-style terms, including the applicable no-endorsement language. Their exact bundled notices are collected in THIRD_PARTY_NOTICES.md and remain in the source headers.

Validation references and model ecosystems

Qwen numerical fixtures use a pinned Hugging Face Transformers implementation and a pinned llama.cpp Qwen converter as independent validation references. The collectors use local deterministic fixtures and do not make those projects runtime dependencies.

The project also depends intellectually on the researchers and open-weight model teams behind the DeepSeek, Qwen, and GLM families, and on the wider GGUF, Metal, and local-inference communities. Exact model, artifact, revision, and benchmark provenance belongs in the corresponding model documentation and dated evidence records rather than in a general credit list.

Development process

The long-form engine reference discloses substantial AI assistance during development, with humans directing design, testing, benchmarking, and release decisions. That assistance does not replace the authorship, copyright, or license notices attached to upstream and third-party work.

Licenses and model terms

The project-level software license is LICENSE. Bundled third-party files retain their own notices and conditions. Model weights, datasets, model cards, and external repositories have their own terms; this repository's software license does not grant rights to those assets.