Event ·Multi-source

SGLang 推出低延迟模式,与 NVIDIA 和 AMD 合作开发

LMSYS 发布 SGLang 低延迟模式,与 NVIDIA 和 AMD 联合优化。在 SemiAnalysis 的 AgentX trace 回放中,DeepSeek-V4.1-Flash 配 DSpark 在 NVIDIA GB300 上 P90 单用户 token 速度提升 1.38 倍(671 到 928 tok/s),在 AMD MI355X 上提升 1.56 倍(332 到 516 tok/s)。技术细节上,NVIDIA 侧通过 kernel 重叠把 decoder 层时间从 96.8 µs 降到 68.4 µs,AMD 侧用每层 2 个 persistent kernel 替代几十个小 kernel。新插件系统支持外部库的自定义 kernel 接入并自动回退,GLM-5.3、K3 和 MiniMax M3 的低延迟 kernel 正在开发中。

Current read

SGLang 出了低延迟模式,在 GB300 和 MI355X 上跑 DeepSeek-V4.1-Flash 快了 1.38 到 1.56 倍,做推理部署的可以看看。

11 sources68° AI heatLast updated 2026/10/20 00:39:06

Evidence

11 sources
Primary sourceLMSYS Org (SGLang)
View source
Related source 3SuperTechFans
View source
Related source 4shao__meng
View source
Related source 5Teortaxes
View source
Related source 6@atomic_chat_hq
View source
Related source 7marktechpost
View source
Related source 8Yuchen Jin
View source
Related source 9TestingCatalog
View source
Related source 10kimmonismus
View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.