DFlash: Block Diffusion for Flash Speculative Decoding

Jian Chen, Yesheng Liang, Zhijian Liu
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:15343-15355, 2026.

Abstract

Autoregressive large language models (LLMs) deliver strong performance but require inherently sequential decoding, leading to high inference latency and poor GPU utilization. Speculative decoding mitigates this bottleneck by using a fast draft model whose outputs are verified in parallel by the target LLM. However, existing methods still rely on autoregressive drafting, which remains sequential and constrains practical speedups. Diffusion LLMs offer a promising alternative by enabling parallel generation, but current diffusion models typically underperform compared with autoregressive models. In this paper, we introduce DFlash, a speculative decoding framework that employs a lightweight block diffusion model for parallel drafting. We show that speculative decoding provides a natural and effective setting for diffusion models. By generating draft tokens in a single forward pass, DFlash enables efficient drafting, and by conditioning the draft model on context features extracted from the target model, it achieves high-quality drafts with higher acceptance rates. Experiments show that DFlash achieves over 6$\times$ lossless acceleration across a range of models and tasks, delivering up to 2.5$\times$ higher speedup than the state-of-the-art speculative decoding method EAGLE-3.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-chen26bv, title = {{DF}lash: Block Diffusion for Flash Speculative Decoding}, author = {Chen, Jian and Liang, Yesheng and Liu, Zhijian}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {15343--15355}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/chen26bv/chen26bv.pdf}, url = {https://proceedings.mlr.press/v306/chen26bv.html}, abstract = {Autoregressive large language models (LLMs) deliver strong performance but require inherently sequential decoding, leading to high inference latency and poor GPU utilization. Speculative decoding mitigates this bottleneck by using a fast draft model whose outputs are verified in parallel by the target LLM. However, existing methods still rely on autoregressive drafting, which remains sequential and constrains practical speedups. Diffusion LLMs offer a promising alternative by enabling parallel generation, but current diffusion models typically underperform compared with autoregressive models. In this paper, we introduce DFlash, a speculative decoding framework that employs a lightweight block diffusion model for parallel drafting. We show that speculative decoding provides a natural and effective setting for diffusion models. By generating draft tokens in a single forward pass, DFlash enables efficient drafting, and by conditioning the draft model on context features extracted from the target model, it achieves high-quality drafts with higher acceptance rates. Experiments show that DFlash achieves over 6$\times$ lossless acceleration across a range of models and tasks, delivering up to 2.5$\times$ higher speedup than the state-of-the-art speculative decoding method EAGLE-3.} }
Endnote
%0 Conference Paper %T DFlash: Block Diffusion for Flash Speculative Decoding %A Jian Chen %A Yesheng Liang %A Zhijian Liu %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-chen26bv %I PMLR %P 15343--15355 %U https://proceedings.mlr.press/v306/chen26bv.html %V 306 %X Autoregressive large language models (LLMs) deliver strong performance but require inherently sequential decoding, leading to high inference latency and poor GPU utilization. Speculative decoding mitigates this bottleneck by using a fast draft model whose outputs are verified in parallel by the target LLM. However, existing methods still rely on autoregressive drafting, which remains sequential and constrains practical speedups. Diffusion LLMs offer a promising alternative by enabling parallel generation, but current diffusion models typically underperform compared with autoregressive models. In this paper, we introduce DFlash, a speculative decoding framework that employs a lightweight block diffusion model for parallel drafting. We show that speculative decoding provides a natural and effective setting for diffusion models. By generating draft tokens in a single forward pass, DFlash enables efficient drafting, and by conditioning the draft model on context features extracted from the target model, it achieves high-quality drafts with higher acceptance rates. Experiments show that DFlash achieves over 6$\times$ lossless acceleration across a range of models and tasks, delivering up to 2.5$\times$ higher speedup than the state-of-the-art speculative decoding method EAGLE-3.
APA
Chen, J., Liang, Y. & Liu, Z.. (2026). DFlash: Block Diffusion for Flash Speculative Decoding. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:15343-15355 Available from https://proceedings.mlr.press/v306/chen26bv.html.

Related Material