Blog

Engineering notes on SGLang, KV cache management, Mooncake Store, Transfer Engine, and related performance work. Articles are currently written in Chinese.

Browse by series

Choose a topic to follow, or open a runbook when you need a command reference.

Recommended reading

A selection of in-depth code walkthroughs and systems notes.

前缀缓存命中 50%,预填充为什么没有快 2 倍?

从因果注意力的三角计算量出发,逐层推导前缀缓存命中后的剩余注意力计算量、完整预填充的理想加速上限,以及它与真实 TTFT 之间的差距。

Code walkthroughsLLM serving internalsLLM serving

All articles

Every article, newest first.

Rethinking the SGLang HiCache Boundary: From L2 Offload to a Cache Runtime

Starting from an out-of-process HiCache L2 split, this design note follows TreeCore, ReqToTokenPool, the KV allocator, CUDA Graph, and the transfer path to refine the ownership boundary between logical cache control and GPU execution.

Code walkthroughsSGLang runtimeLLM serving

前缀缓存命中 50%,预填充为什么没有快 2 倍?

从因果注意力的三角计算量出发,逐层推导前缀缓存命中后的剩余注意力计算量、完整预填充的理想加速上限,以及它与真实 TTFT 之间的差距。

Code walkthroughsLLM serving internalsLLM serving

如何根据 config.json 估算大模型的注意力缓存

从 config.json 中的 hidden_size、head_dim、KV 头数、sliding_window、kv_lora_rank、Indexer 和 KDA 状态出发,估算常见模型的注意力缓存大小。

Code walkthroughsLLM serving internalsAttention

存储基准测试笔记:不要只看平均延迟

借助 redis-benchmark、memtier_benchmark 和 sysbench,说明测试负载、预热、数据规模、p99、性能拐点和资源隔离为什么比平均值更重要。

Engineering notesStorage systemsPerformance