Reproduce “D3D: Decoupled Diffusion Drafting for Asynchronous 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
Baoyou Chen, Ruiqiao Mei, Guancong Lin, Tao Song, Weihao Yuan, Qun Liu, Siyu Zhu
D3D enables asynchronous speculative decoding with a standalone diffusion drafter that prepares successor blocks while the target verifies current tokens. It uses its own prefix representations and reuses a prepared branch only when verification matches; otherwise drafting waits. On Qwen3-32B, it reports 5.71× average decoding speedup across seven tasks, versus DSpark’s 3.92×, under the evaluated setups—not per-GPU efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.