Reproduce “DFlare: Scaling Up Draft Capacity for Block Diffusion Speculative Decoding” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiebin Zhang, Zhenghan Yu, Song Liu, Eugene J.Yu, Zheng Li, Dawei Zhu, Jiangshan Duo, Weimin Xiong and 4 more
Researchers from Peking University and Tencent developed DFLARE, a framework for block diffusion speculative decoding that scales draft model capacity using adaptive layer-wise fusion and heterogeneous KV projections. The method demonstrated wall-clock speedups up to 5.52x and average acceptance lengths of 7.47 for Qwen3-4B, surpassing DFlash across multiple LLMs and benchmarks.
Key insights
- Providing distinct, adaptively learned target hidden state combinations to each draft layer enables effective depth scaling and specialization, mitigating diminishing returns observed in prior methods.
- Separating Key and Value projections for target context versus draft tokens allows the attention mechanism to process different information types more effectively, enhancing overall model performance.
- A curriculum-style training loss with dynamic weighting facilitates comprehensive optimization across all token positions within a block, leading to more robust and accurate draft predictions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.