arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2512.20573cs.LGcs.AIcs.DC
收藏

快速失败,大获成功:通过扩散大语言模型重新思考推测解码的起草策略

Fail Fast, Win Big: Rethinking the Drafting Strategy in Speculative Decoding via Diffusion LLMs

Princeton University(普林斯顿大学);University of Washington(华盛顿大学);Google(谷歌);Rice University(里士满大学)

Rui Pan, Zhuofu Chen, Hongyi Liu, Arvind Krishnamurthy, Ravi Netravali

AI总结:

FailFast通过动态调整推测长度,利用扩散大语言模型的并行解码优势,实现快速失败和大获成功,提升推测解码效率。

摘要:

扩散大语言模型(dLLMs)提供了快速、并行的标记生成,但其单独使用会受到固有效率-质量权衡的困扰。我们证明,如果仔细应用,dLLMs的属性实际上可以成为推测解码中自回归(AR)验证器的起草者的优势。我们的核心见解是dLLM的并行解码速度显著降低了昂贵拒绝的风险,提供了一种实际机制,有效实现导致大规模加速的长起草。我们提出了FailFast,一种基于dLLM的推测解码框架,通过动态适应其推测长度来实现这一方法。它通过在难以推测的区域花费最小计算来缩小推测延迟以“快速失败”,并通过在更容易的区域积极延长起草长度以减少验证延迟(在许多情况下,每次推测和接受70个标记!)。无需任何微调,FailFast实现了AR LLM的无损失加速,并在各种模型和工作负载上实现了高达4.9倍于普通解码的速度提升,1.7倍于最佳的naive dLLM起草者,以及1.7倍于EAGLE-3。我们开源FailFast在https://github.com/ruipeterpan/failfast。

英文摘要:

Diffusion Large Language Models (dLLMs) offer fast, parallel token generation, but their standalone use is plagued by an inherent efficiency-quality tradeoff. We show that, if carefully applied, the attributes of dLLMs can actually be a strength for drafters in speculative decoding with autoregressive (AR) verifiers. Our core insight is that dLLM's speed from parallel decoding drastically lowers the risk of costly rejections, providing a practical mechanism to effectively realize the (elusive) lengthy drafts that lead to large speedups with speculative decoding. We present FailFast, a dLLM-based speculative decoding framework that realizes this approach by dynamically adapting its speculation length. It "fails fast" by spending minimal compute in hard-to-speculate regions to shrink speculation latency and "wins big" by aggressively extending draft lengths in easier regions to reduce verification latency (in many cases, speculating and accepting 70 tokens at a time!). Without any fine-tuning, FailFast delivers lossless acceleration of AR LLMs and achieves up to 4.9$\times$ speedup over vanilla decoding, 1.7$\times$ over the best naive dLLM drafter, and 1.7$\times$ over EAGLE-3 across diverse models and workloads. We open-source FailFast at https://github.com/ruipeterpan/failfast.

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。

↑