Files
engine/NOTICE
zeekayandHanzo Dev b4f543c223 legal: state the mistral.rs provenance accurately, and in LLM.md
NOTICE claimed "the initial import was squashed into a single commit" and that
the tree "does not pin the exact upstream revision it diverged from". Both are
false, and false against the repo's own interest: nothing was squashed. The
root commit is upstream's own 1ecfb7084 (EricLBuehler, 2024-02-26), 3,697 of
5,495 commits on main are Eric Buehler's, and the divergence IS pinned --
7483b396f (2024-02-27), the same SHA tools/upstream-map.sh already names. The
rename rewrote hashes past that point, which is why merge-base collapses there
and why the SHAs stop matching; that is a rewrite, not a squash. Upstream is
merged through mistral.rs ba0fedae (PR #2188, 2026-06-02).

Also in NOTICE:
- CUTLASS was credited under BSD-3-Clause. No CUTLASS source is vendored and
  no BSD-3-Clause source exists in this tree (control: 39 files carry Apache
  headers). The NVIDIA-copyright code is FasterTransformer, Apache-2.0,
  reached via the vLLM kernels. Credited as what it is.
- The candle crate list omitted hanzo-transformers and hanzo-flash-attn-v3,
  both real workspace deps.

README credited "Candle" but linked it to hanzoai/ml -- our own fork standing
in for upstream, a leftover from the 2026-06-01 brand sweep. Points at
huggingface/candle now, with our fork named as the fork.

LLM.md had no provenance section at all: it documented `git merge upstream/master`
without ever saying what upstream is. Adds one, and records why LICENSE's
"Copyright (c) 2024 Eric Buehler" line is load-bearing -- c6917647b deleted it
on 2026-06-01 and 4b52c3416 restored it 45 days later, and MIT S2 is satisfied
by retaining exactly that line. Also records that Apache-2.0 S4(d) makes NOTICE
a distribution obligation for the vendored vLLM/FlashInfer/Marlin kernels.

No LICENSE, LICENSE-APACHE, COPYING or AUTHORS file is touched.

Co-authored-by: Hanzo Dev <[email protected]>
2026-08-03 21:32:58 -07:00

92 lines
5.2 KiB
Plaintext

Hanzo Engine
Copyright (c) 2024 Hanzo AI, Inc.
This product is licensed under the MIT License (see LICENSE).
-------------------------------------------------------------------------------
Third-party attribution
-------------------------------------------------------------------------------
Hanzo Engine is derived from mistral.rs (https://github.com/EricLBuehler/mistral.rs),
Copyright (c) 2024 Eric Buehler, used under the MIT License:
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
Hanzo Engine is an independent fork: all crates, binaries, and APIs are
Hanzo-branded (hanzo-engine), and Hanzo AI is not affiliated with Mistral AI.
Forked from https://github.com/EricLBuehler/mistral.rs. Upstream history is
retained in full, not squashed: this tree's root commit is upstream's own
1ecfb70842640b12c8321d69be58bec5331ddffc (EricLBuehler, 2024-02-26), and 3,697
of the 5,495 commits reachable from main are authored by Eric Buehler. The
crate rename (mistralrs-* -> hanzo-*) rewrote commit hashes from
7483b396fa0513362977a6e033e5259d1a7319e7 (2024-02-27) onward, so beyond that
point our SHAs no longer equal upstream's even where the content is theirs;
tools/upstream-map.sh maps the paths. Upstream work is merged through
mistral.rs 0a14af4a43e6f61afbed71f72f22c162ab944c0d (PR #2188, 2026-06-02).
Hanzo Engine tracks its own release line (1.7.x) rather than upstream tags.
Hanzo Engine is built on candle (https://github.com/huggingface/candle),
Copyright (c) 2023 HuggingFace Inc., dual-licensed under the MIT License OR the
Apache License, Version 2.0; used here under the MIT License (full text above).
The candle crates are consumed as Hanzo's fork (https://github.com/hanzoai/ml),
published as hanzo-ml (candle-core), hanzo-nn (candle-nn), hanzo-transformers
(candle-transformers), hanzo-flash-attn, hanzo-flash-attn-v3, and
hanzo-metal-kernels. That fork carries candle's own LICENSE-MIT and
LICENSE-APACHE and its own NOTICE; this file does not restate them.
Hanzo Engine bundles third-party GPU kernels, each under its own permissive
license (per-file copyright headers retained in-source):
- vLLM (Apache-2.0), Copyright (c) 2023 The vLLM team
- FlashInfer (Apache-2.0), Copyright (c) 2023-2025 FlashInfer team
- Apple MLX (MIT), Copyright (c) 2023-2024 Apple Inc.
- NVIDIA FasterTransformer (Apache-2.0), Copyright (c) 2020-2023 NVIDIA
CORPORATION (reached via the vLLM kernels; no CUTLASS source is vendored,
and no BSD-3-Clause-licensed source is present in this tree)
- Marlin (Apache-2.0), Copyright (c) 2024 Elias Frantar
- ggml / llama.cpp (MIT), Copyright (c) ggml authors (Georgi Gerganov et al.)
-------------------------------------------------------------------------------
DEVIATIONS
-------------------------------------------------------------------------------
Hanzo Engine has diverged substantially from upstream mistral.rs:
- Full Hanzo rebrand: every workspace crate is renamed hanzo-* (hanzo-engine,
hanzo-server-core, hanzo-router, hanzo-quant, hanzo-vision, hanzo-audio,
hanzo-paged-attn, hanzo-llm-mcp, hanzo-sandbox, hanzo-code-exec, hanzo-cli,
hanzo-pyo3, hanzo-film), and the CLI/binary is `hanzo`.
- Candle is consumed as Hanzo's fork (hanzo-ml, hanzo-nn, hanzo-flash-attn,
hanzo-metal-kernels) rather than upstream candle-core / candle-nn.
- OpenAI-compatible `/v1` server plus an Anthropic-compatible
`POST /v1/messages` endpoint (streaming + tool use), served from one router.
- All-modality serving: text/chat, vision (hanzo-vision), audio in and 24 kHz
speech out (hanzo-audio), video and image/diffusion models, long-form film
orchestration (hanzo-film), and 3D (hanzo-3d).
- GPU backend coverage across CUDA, Apple Metal, and AMD ROCm/HIP
(Dockerfile.cuda, Dockerfile.rocm, and the Metal kernels).
- Broad quantization support: GGUF, ISQ, UQFF, Q8_0/Q4_K, IQ4, MXFP4, HQQ,
AFQ, and bitsandbytes FP4.
- Zen model serving (native/zen3-serving and the zen model paths).
- Adapter primitives: AnyMoE plus LoRA and QLoRA adapters.
- Hanzo ZAP transport integration (hanzo-engine/src/zap) and a disk-first KV
cache for cross-restart sessions and agent prefix reuse.
- Sandboxed local code execution (hanzo-sandbox, hanzo-code-exec) for the
agentic runtime.