arXiv cs.AI· Chung-En Ho, Weiyu Sun, Cheng-Jhih Shih, He Li, Yong Liu, Yingyan Celine Lin·· 10 天前AI 评分28
SpecFold：折叠多分支冗余，加速扩散语言模型投机解码
SpecFold: Folding Multi-Branch Redundancy for Faster Speculative Decoding in Diffusion Language Models
AI 导读
SpecFold 是一种算法与系统协同设计方法，通过折叠注意力与 FFN 复用父分支计算，减少扩散语言模型多分支投机验证中的冗余开销。在两类 DLLM、五个模型和五个基准上，其吞吐量最高达 Spiffy 的 1.64 倍、原生解码的 1.99 倍，且任务性能相当。该方法与时间缓存正交，兼容现有 DLLM 投机策略。
正文 · 原文
Abstract:Diffusion large language models (DLLMs) generate text through iterative block denoising, and multi-branch speculative decoding accelerates this process by verifying a main branch together with multiple draft branches in a single forward pass. While prior DLLM acceleration methods primarily exploit temporal redundancy across denoising steps, we identify a complementary redundancy axis within each speculative verification step: multi-branch computational redundancy. During speculative verification, draft branches inherit most tokens from their parents while unmasking a small set of additional positions, causing large portions of hidden states to remain highly similar across branches. We propose SpecFold, an algorithm-system co-design that exploits this multi-branch redundancy to reduce the cost of multi-branch speculative verification. Algorithmically, SpecFold performs token-level residual gating and selectively reuses parent computation through folded attention and FFN while preserving residual hidden states. Systemically, a Triton kernel implementation translates this fine-grained reuse into end-to-end throughput gains through efficient sparse multi-branch execution. SpecFold is orthogonal to temporal caching and compatible with existing DLLM speculation strategies. Across two DLLM families, five models, and five standard benchmarks, SpecFold achieves up to 1.64x throughput over Spiffy and up to 1.99x over vanilla decoding, while maintaining comparable task performance.
| Subjects: | Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.04875 [cs.AI] |
| (or arXiv:2610.04875v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.04875 arXiv-issued DOI via DataCite |
Submission history
From: Chung-En Ho [view email]
[v1]
Sun, 4 Oct 2026 02:25:38 UTC (453 KB)
[v2]
Tue, 6 Oct 2026 03:42:24 UTC (453 KB)
来源：arXiv cs.AI · arxiv.org