perf(agentx): refresh GLM-5.2 B200 SGLang curve / 性能:刷新 GLM-5.2 B200 SGLang 曲线 (#2808)
* perf(agentx): refresh GLM-5.2 B200 SGLang curve Pin the fixed SGLang nightly, extend the AgentX curve through c20, right-size HiCache for the NScale DRAM budget, and isolate runtime caches per Slurm allocation. 中文:固定包含修复的 SGLang nightly 镜像,将 AgentX 曲线扩展至 c20,按 NScale DRAM 预算调整 HiCache 大小,并为每个 Slurm 分配隔离运行时缓存。 * chore: link InferenceX PR 2808 Backfill the public pull request URL in the appended performance changelog entry. 中文:在追加的性能变更日志条目中回填公开拉取请求链接。 * perf(agentx): drop GLM-5.2 B200 c20 point Keep the validated c1/c4/c8/c12/c16 curve and retain the 169 GB/rank HiCache target for c12/c16. 中文:保留已验证的 c1/c4/c8/c12/c16 曲线,并继续在 c12/c16 使用每个 rank 169 GB 的 HiCache 目标池。
A
Ankur Singh committed
f2be63a98849591c40d80428cef5af6fb185417c
Parent: 5ecf6fd
Committed by GitHub <noreply@github.com>
on 9/2/2026, 11:34:37 PM