Translation in progress Some entries still include text in the original language.
BackIndexed by this site 2026-10-27 21:30 3 days ago
Originally published Oct 26
Selected 2026-10-27 21:31
Category Event LMSYS Org · Indexed by this site 2026-10-27 21:30 · Originally publishedOct 26 SelectedAttention63
SGLang security update: improved request isolation SGLang security update: improving request isolation
AI summary
SGLang 维护方与 Cloudflare 协调披露 SGLang PD 分离服务中的跨请求 KV 缓存隔离缺陷。Cloudflare 的调查发现,其 Workers AI 上的 GLM-5.3 部署在 2026 年 8 月 27 日至 9 月 9 日间,超出 prefill 缓存容量的请求可能返回其他用户会话的内容,起因是配置的上下文长度 1,048,576 token 超过 prefill 缓存池的 681,280 token,被拒请求仍进入分离交接,decode 读到未为本请求填充的缓存页。相关的一类竞态是请求中止后已发出的 RDMA 写入落入被重新分配的槽位,造成无报错的单向跨请求泄漏,长上下文会扩大窗口;SGLang 自 v0.4.x 至 v0.5.21(MORI 后端至 v0.5.22)的 PD 默认配置受影响。维护方建议 PD 部署升级并运行能配对中止的路由器。
Why it was selected
The original source explains the race condition mechanism and affected versions of cross-request KV cache leakage under PD-disaggregated serving; teams running SGLang multi-tenant long-context services can check their own deployments against it.
Reported by another 4 sources Read original lmsys.org
Why it was selected SelectedAttention63 5 new sources in 24 hours
Attention 0–100: content value contributes up to 50 and circulation up to 50 (accounts weighted by likes); circulation decays over time. How it was rated
The original source explains the race condition mechanism and affected versions of cross-request KV cache leakage under PD-disaggregated serving; teams running SGLang multi-tenant long-context services can check their own deployments against it.