---
id: 20260314-T0-08
title: "SDSL论文提出推测解码吞吐量优化新方法"
title_en: "SDSL Paper Simplifies Speculative Decoding Throughput"
url: https://ai.daily.yangsir.net/daily/20260314-T0-08
issue_date: 2026-03-14
publish_date: 2026-03-13T04:00:00.000Z
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2603.11053
---

# SDSL论文提出推测解码吞吐量优化新方法

arXiv论文《Speculative Decoding Scaling Laws》提出推测解码吞吐量优化定律。该研究建立了模型规模与推理速度的数学关系，通过调整推测策略将推理速度提升40%。方法无需重新训练模型，降低了大语言模型部署成本，为工业应用提供实用指南。

## English Version

**SDSL Paper Simplifies Speculative Decoding Throughput**

The arXiv paper 'SDSL' introduces scaling laws for speculative decoding throughput. It establishes mathematical relationships between model size and inference speed, boosting speed by 40% through strategy adjustments. The method requires no retraining, reducing LLM deployment costs for industrial applications.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2603.11053)

**详情页**：https://ai.daily.yangsir.net/daily/20260314-T0-08

---

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