文章背景与核心概要
大语言模型(LLM)的推理服务经常受到键值(KV)缓存不断增长所带来的内存瓶颈限制。尽管 PagedAttention 通过使用固定大小的内存块成功缓解了分配器层面的内存碎片,但由于现代 KV 淘汰算法运行在更精细的 Token 级别,它依然面临严重的块内碎片(intra-block fragmentation)问题。
为了解决这一痛点,vToken 引入了一种轻量级的虚拟化层,将逻辑 Token 的活跃状态与物理块的放置位置进行解耦。通过利用 Token 表间接寻址和异步物理重打包(repacking),vToken 在保持与现有 PagedAttention 内核以及 CUDA Graphs 兼容的同时,消除了块内部的内存浪费。
vToken: Token-Level Virtualization for Reclaimable KV Caches
arXiv: 2608.13263
Submitted: 13 August 2026
Authors: Yuanhang Gao, Xiangrui Yang, Yuanfeng Chen, Hongjia Chen, Qianru Lv, Wenfei Wu, Dongsheng Li
Summary
大语言模型(LLM)服务经常受到由键值(KV)缓存增长引起的内存瓶颈的限制。虽然 PagedAttention 成功地使用固定大小的内存块缓解了分配器层面的碎片,但由于现代 KV 淘汰算法运行在更精细的 Token 级别粒度上,它遭遇了块内碎片的问题。
Large Language Model (LLM) serving is frequently constrained by memory bottlenecks caused by the growth of Key-Value (KV) caches. While PagedAttention successfully mitigates allocator-level fragmentation using fixed-size memory blocks, it suffers from intra-block fragmentation because modern KV eviction algorithms operate at a finer, token-level granularity.
vToken 引入了一个轻量级的虚拟化层,将逻辑 Token 的活跃性与物理块的放置位置解耦。通过利用 Token 表间接寻址和异步物理重打包,vToken 消除了块内浪费的内存,同时保持了与现有 PagedAttention 内核和 CUDA Graphs 的兼容性。
vToken introduces a lightweight virtualization layer that decouples logical token liveness from physical block placement. By utilizing token-table indirection and asynchronous physical repacking, vToken eliminates wasted memory within blocks while maintaining compatibility with existing PagedAttention kernels and CUDA Graphs.
Key Contributions
- Token-Level Virtualization: Decouples the logical view of tokens from their physical storage, allowing for efficient, granular reclamation.
- Asynchronous Repacking: Enables physical memory reclamation without disrupting the logical sequence flow.
- Performance Gains:
- Reduces retained KV blocks per request by 27.2% – 72.3%.
- Improves SLA-constrained throughput by up to 1.37×.
- Extends maximum feasible concurrency by up to 2× under constrained budgets.
- Developer Efficiency: Reduces the integration footprint for new eviction policies from 500+ lines of code to under 50.
主要贡献
- Token 级虚拟化: 将 Token 的逻辑视图与其物理存储解耦,从而实现高效、精细化的内存回收。
- 异步重打包: 能够在不中断逻辑序列流的情况下实现物理内存回收。
- 性能提升:
- 将每个请求保留的 KV 块减少了 27.2% – 72.3%。
- 在满足 SLA 约束的吞吐量上提升高达 1.37 倍。
- 在受限预算下,将最大可行并发度扩展高达 2 倍。
- 开发效率: 将新淘汰策略的集成代码量从 500 多行代码缩减至 50 行以下。
Technical Details
- Compatibility: Fully preserves PagedAttention kernels and CUDA Graph support.
- Implementation: Developed and evaluated within the vLLM framework.
- Evaluation Benchmarks: Tested against H2O, Random, and Scissorhands eviction policies.
技术细节
- 兼容性: 完全保留了 PagedAttention 内核以及对 CUDA Graph 的支持。
- 实现: 在 vLLM 框架内开发并进行了评估。
- 评估基准: 针对 H2O、Random 和 Scissorhands 淘汰策略进行了测试。
Access & Resources
- Full Paper: View PDF
- Experimental HTML: View HTML
- Source Code: TeX Source
- DOI: 10.48550/arXiv.2608.13263
获取与资源
- 完整论文: 查看 PDF
- 实验性 HTML: 查看 HTML
- 源代码: TeX 源码
- DOI: 10.48550/arXiv.2608.13263
Subjects: Artificial Intelligence (cs.AI); Distributed, Parallel, and Cluster Computing (cs.DC); Operating Systems (cs.OS)