Personal project自己的项目
mini-DeepInfer
The question这个项目在研究什么
A readable DeepSeek-V4.1 runtime可阅读的 DeepSeek-V4.1 运行时
Paged KV blocks and state slots, prefill/decode scheduling with preemption, zero-overhead scheduling, deterministic sampling, EP × DP, and paired prefill measurements against official DeepInfer.分页 KV block 与 state slot、带抢占的 prefill/decode 调度、零开销调度、确定性采样、EP × DP，以及与官方 DeepInfer 的配对 prefill 测量。
New to the topic? Start with a small example. Already familiar? Go straight to the articles.第一次接触，先跑一个小例子；已有基础，直接进入文章。
Scope & current evidence研究边界与当前证据Read before applying a result应用结论前请留意
A learning implementation, not a faster DeepInfer. Real-model prefill reaches 11.9% of the official engine in one paired run on 4 × B200; decode CUDA graphs were off for both and KV budgets were not forced equal.这是学习型实现，不是更快的 DeepInfer。4 × B200 单轮配对测量中，真实模型 prefill 达到官方的 11.9%；两边均未开启 decode CUDA Graph，KV 预算也未强制相等。
Personal learning implementation个人学习实现 · 7 working outlines篇研究提纲