SGLang fixes cross-request KVCache race in PD-disaggregated serving
What happened
Cloudflare reported that GLM-5.3 on Workers AI was affected by a cross-request KVCache problem between August 27 and September 9, 2026, and its investigation found no other Workers AI models affected. An October 26 report says SGLang has disclosed and fixed the race in its PD-disaggregated serving. Two cases are involved: rejected requests still entering the handoff flow, and async writes after an aborted request polluting a reused cache. The report confirms the fix but gives no specific fix date.
Written by AI from the coverage · updated 1 hour ago
Coverage
Follow the reports to see the story from different sides.
- AI HOT · Tips & opinionsSGLang 披露并修复 PD 分离服务中的跨请求 KVCache 竞态问题
SGLang 披露并修复 PD 分离服务中的跨请求 KVCache 问题，涉及拒绝请求仍进入交接流程，以及中断后的异步写入污染已复用缓存。Cloudflare 报告 Workers AI 的 GLM-5.3 在 2026 年 8 月 27 日至 9 月 9 日受到影响，调查未发现其他 Workers AI 模型受影响。
Heat over time
Not enough continuous observations to draw a trend yet.