[Bug/Perf]: hybrid-SWA prefix caching collapses to zero for ALL requests in multi-session round-robin at ~25% pool occupancy (Gemma-4-31B; e
![[Bug/Perf]: hybrid-SWA prefix caching collapses to zero for ALL requests in multi-session round-robin at ~25% pool occupancy (Gemma-4-31B; e](https://www.chat-gpts.plus/wp-content/uploads/2026/09/48435-ed9f024c-768x403.jpg)
该问题发生在 vLLM 混合注意力(hybrid-SWA)模型的多会话轮询负载中,当 KV 缓存池占用率约达 25% 时,跨请求前缀缓存复用会突然降为零,表现为“全部命中”或“全部未命中”的断崖式切换。优先排查 SWA 组的尾部块是否被提前回收,可借助 CPU offloading(48GiB+)验
![[Bug][DCP] GLM-5.3 dense prefill consumes uninitialized K rows on non-owner ranks](https://www.chat-gpts.plus/wp-content/uploads/2026/09/54907-d7bce83d-768x403.jpg)







