---
id: 20260619-T0-06
title: "JetFlow：利用并行树起草打破推测解码的扩展瓶颈"
title_en: "JetFlow: Breaking Speculative Decoding Bottlenecks with Parallel Tree Drafting"
url: https://ai.daily.yangsir.net/daily/20260619-T0-06
issue_date: 2026-06-19
publish_date: 2026-06-18T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2606.18394
---

# JetFlow：利用并行树起草打破推测解码的扩展瓶颈

论文《JetFlow》提出了一种新的推测解码方案，通过并行树起草技术解决了传统方法在增加草稿预算时面临的扩展天花板问题。实验显示，该方法能显著提升大语言模型的生成速度，解决了推测解码中高并发下的验证效率难题。

## English Version

**JetFlow: Breaking Speculative Decoding Bottlenecks with Parallel Tree Drafting**

The paper 'JetFlow' proposes a new speculative decoding method using parallel tree drafting to overcome the scaling ceiling faced by traditional methods when increasing draft budgets. Experiments show significant speedups in LLM generation, solving the verification efficiency bottleneck in high-concurrency speculative decoding scenarios.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2606.18394)

**详情页**：https://ai.daily.yangsir.net/daily/20260619-T0-06

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*