dfly-mlx
Copyright 2026 Willson

This product includes software and model-architecture adaptations from:

1. AngelSpec, Copyright (C) 2026 Tencent, Apache License 2.0. Ported from
   revision d3412bed231025ea85d35dbf5e0af44f1ac62a5b.
   https://github.com/Tencent/AngelSpec

2. dflash-mlx, Copyright 2026 bstnxbt, Apache License 2.0. The implementation
   was validated against revision 9ca002898b48e14c9727dec17299f497e8467870
   (v0.1.10) and supports the public PyPI runtime from v0.1.8.
   https://github.com/bstnxbt/dflash-mlx

3. vLLM, copyright the vLLM contributors, Apache License 2.0. The sequential
   hidden-correction inference semantics were adapted from its Qwen3 DFly
   proposer implementation.
   https://github.com/vllm-project/vllm/pull/50246

DFly is described in "AngelSpec: Towards Real-World High Performance Inference
with Speculative Decoding" (arXiv:2607.25852).
