arXiv cs.AI· Baoyou Chen, Ruiqiao Mei, Guancong Lin, Tao Song, Weihao Yuan, Qun Liu, Siyu Zhu·· 11 小时前AI 评分39
D3D：面向异步投机解码的解耦扩散草稿模型
D3D: Decoupled Diffusion Drafting for Asynchronous Speculative Decoding
AI 导读
D3D 是一种独立的扩散草稿模型，通过块级去噪加辅助自回归监督，摆脱对目标模型隐藏状态的依赖，可在验证阶段并行草拟后续块，实现异步投机解码。以 Qwen3-8B 和 Qwen3-32B 为目标时，D3D 在七个基准上平均解码加速 5.00× 和 5.71×，DFlash 为 4.03× 和 3.67×；Qwen3-32B 在 GSM8K 上加速达 7.48×。
正文 · 原文
Abstract:Target-conditioned diffusion drafters predict candidate tokens in parallel, yet each new block must wait for updated target-model hidden states. This forces drafting and verification to run sequentially, limiting speedup even though each block is generated in one pass. We introduce D3D, a standalone diffusion drafter that removes this inter-round dependency. D3D combines blockwise denoising with auxiliary autoregressive supervision to achieve long accepted prefixes without target hidden states. This independence lets D3D draft successor blocks during verification in one packed pass, enabling asynchronous speculative decoding. We progressively prune layers according to their effect on acceptance length to reduce the cost of packed drafting. Removing five of 36 layers retains over 99\% of mean acceptance length across three benchmarks with Qwen3-8B as the target. Across seven benchmarks, D3D averages 5.00$\times$ and 5.71$\times$ decoding speedup with Qwen3-8B and Qwen3-32B targets, compared with 4.03$\times$ and 3.67$\times$ for DFlash. On GSM8K with Qwen3-32B, speedup reaches 7.48$\times$. Code is publicly available at this https URL.
| Subjects: | Artificial Intelligence (cs.AI); Machine Learning (cs.LG) |
| Cite as: | arXiv:2611.06425 [cs.AI] |
| (or arXiv:2611.06425v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2611.06425 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Baoyou Chen [view email]
[v1]
Wed, 4 Nov 2026 12:27:58 UTC (1,882 KB)
来源：arXiv cs.AI · arxiv.org