arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2611.06425cs.AIcs.LG
收藏

D3D:用于异步推测解码的解耦扩散草稿生成

D3D: Decoupled Diffusion Drafting for Asynchronous Speculative Decoding

Shanghai Innovation Institute(上海创新研究院);Shanghai Academy of AI for Science(上海科学人工智能研究院);Fudan University(复旦大学);Nanjing University(南京大学);Huawei Technologies Co., Ltd.(华为技术有限公司)

Baoyou Chen, Ruiqiao Mei, Guancong Lin, Tao Song, Weihao Yuan, Qun Liu, Siyu Zhu

AI总结:

该研究针对异步推测解码提出 D3D 解耦扩散草稿生成方法,通过分块去噪与辅助自回归监督消除轮次依赖,剪枝后在多基准测试中实现了优于 DFlash 的解码加速比。

摘要:

目标条件扩散草稿生成器会并行预测候选 token,但每个新块必须等待更新后的目标模型隐藏状态,这迫使草稿生成与验证串行运行,即便每个块仅需一次前向传播即可生成,也限制了加速效果。我们提出 D3D,一种独立的扩散草稿生成器,它消除了这种轮次间依赖关系。D3D 将分块去噪与辅助自回归监督相结合,无需目标隐藏状态即可生成被长期接受的前缀。这种独立性使 D3D 能在验证阶段以一次打包前向传播生成后续块,从而实现异步推测解码。我们根据层对接受长度的影响逐步剪枝层,以降低打包草稿生成的成本:在以 Qwen3-8B 为目标模型的三个基准测试中,剪去 36 层中的 5 层后,仍保留了超过 99% 的平均接受长度。在七个基准测试中,D3D 以 Qwen3-8B 和 Qwen3-32B 为目标模型时,解码加速比平均分别为 5.00× 和 5.71×,而 DFlash 的对应值为 4.03× 和 3.67×;在使用 Qwen3-32B 的 GSM8K 基准测试中,加速比达到 7.48×。代码已公开提供。

英文摘要:

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 https://github.com/fudan-generative-vision/D3D.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑