arxiv:2611.06425

D3D: Decoupled Diffusion Drafting for Asynchronous Speculative Decoding

Published on Nov 4

· Submitted by

Baoyou Chen on Nov 5

Fudan University

Upvote

Authors:

Baoyou Chen ,

Abstract

A standalone diffusion drafter using blockwise denoising and auxiliary autoregressive supervision enables asynchronous speculative decoding by drafting successor blocks during verification, yielding large decoding speedups.

Generated by deepseek-ai/DeepSeek-V4.1-Flash

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 5.00times and 5.71times decoding speedup with Qwen3-8B and Qwen3-32B targets, compared with 4.03times and 3.67times for DFlash. On GSM8K with Qwen3-32B, speedup reaches 7.48times. Code is publicly available at https://github.com/fudan-generative-vision/D3D.

View arXiv page View PDF GitHub 20 Add to collection

Community

cbyzju

Paper author Paper submitter 5 days ago

🚀 Introducing D3D: Decoupled Diffusion Drafting for Asynchronous Speculative Decoding!

Existing diffusion-based speculative drafters rely on target-model hidden states, forcing drafting and verification to execute sequentially.

D3D breaks this dependency with a standalone diffusion drafter, enabling asynchronous drafting and verification through packed speculative decoding.

Key highlights:

Asynchronous speculation: Overlaps diffusion drafting with target verification, eliminating inter-round dependencies.
Effective drafting: Combines blockwise denoising with auxiliary autoregressive supervision to achieve long accepted prefixes.
Significant speedups: Achieves 5.00× and 5.71× average decoding speedups with Qwen3-8B and Qwen3-32B across seven benchmarks, outperforming DFlash (4.03× and 3.67×). Up to 7.48× on GSM8K!

📄 Paper: https://arxiv.org/abs/2611.06425

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

🤗 Models: https://huggingface.co/collections/fudan-generative-ai/d3d

We've released our code, model checkpoints, and benchmarking scripts.

We welcome feedback, questions, and discussions!

polyarti

1 day ago

Thank you, that's a lovely trade-off: keeping over 99% of the acceptance length and still getting faster, 4.44x to 4.69x, is the kind of result I'd happily take. Did the drop past 31 layers look gradual, or more like a cliff?

librarian-bot

4 days ago

This is an automated message from the Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend

polyarti

3 days ago

This stopped my scroll: drafting the next block while the target model is still verifying sounds like free speed. How much does acceptance length suffer once the layers are pruned?

cbyzju

Paper author 2 days ago

Thanks for your interest! In our experiments with Qwen3-8B as the target, pruning the D3D-8B drafter from 36 to 31 layers retains over 99% of the original average acceptance length across MATH-500, HumanEval, and MT-Bench. Meanwhile, the average decoding speedup improves from 4.44× to 4.69×. Further pruning leads to a more noticeable drop in acceptance length, making 31 layers a good trade-off in our setting. More details are in Figure 2.