---
id: 20261021-T0-04
title: "PolyGraph：用计划缓冲的CUDA Graph加速自适应投机解码"
title_en: "PolyGraph Uses Plan-Buffered CUDA Graphs to Speed Adaptive Speculative Decoding"
url: https://ai.daily.yangsir.net/daily/20261021-T0-04
issue_date: 2026-10-21
publish_date: 2026-10-20T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2610.23861
---

# PolyGraph：用计划缓冲的CUDA Graph加速自适应投机解码

自适应投机解码会在每个解码步改变活跃批大小B和草稿长度d，而CUDA Graph只能重放固定的启动配置，验证宽度M=B×d的波动迫使服务系统预先捕获大量图配置。论文提出PolyGraph，通过计划缓冲的方式让CUDA Graph适配动态变化的解码参数，减少预捕获开销。这对LLM推理服务的意义是降低显存与启动成本，让自适应投机解码更容易落地到生产环境。

## English Version

**PolyGraph Uses Plan-Buffered CUDA Graphs to Speed Adaptive Speculative Decoding**

Adaptive speculative decoding changes the active batch size B and draft length d at every decode step, but CUDA Graphs replay a fixed launch configuration, so the fluctuating verification width M = B×d forces serving systems to pre-capture many graph variants. PolyGraph addresses this with plan-buffered CUDA Graphs that accommodate dynamic decoding parameters while cutting pre-capture overhead. For LLM serving, that means lower memory and startup costs, making adaptive speculative decoding more practical to deploy in production.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2610.23861)

**详情页**：https://ai.daily.yangsir.net/daily/20261021-T0-04

---

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