01
大模型系统的 18 个第一性原理锚点
ai-systems / llm-inference
llm-inferenceattentionkv-cachemoe
+1
02
DeepSeek V4.1 Flash：从架构到推理成本
ai-systems / llm-inference
llm-inferenceattentionkv-cachemoe
+1
03
P/D 分离：从 KV 交接到路由、背压与 SLO
ai-systems / llm-inference
llm-inferenceprefilldecodekv-cache
+2
04
分层 KV Cache：从 GPU 命中到外部存储复用
ai-systems / llm-inference
llm-inferencekv-cacheprefillnetworks
+1
05
vLLM Async Scheduling：三态配置、投机解码与状态提交
ai-systems / llm-inference
llm-inferencevllmdecodekv-cache
+1
06
KV Cache Hit Ratio 修正模型：从直觉到统一公式
ai-systems / llm-inference
llm-inferencekv-cachesimulatorprefill
+1
07
KV Cache：推理性能的命根子
ai-systems / llm-inference
LLMInferenceKV CachePagedAttention
+2
08
Attention 架构演化：从多头注意力（MHA）到 GQA、MLA
ai-systems / llm-inference
llm-inferenceattentionkv-cache
09
DeepSeek MLA：低秩 KV Cache 与推理效率
ai-systems / llm-inference
llm-inferenceattentionkv-cachedeepseek
+1
10
推理框架对比 2026：从 Engine 到 Serving Stack
ai-systems / llm-inference
vllmsglangtensorrt-llminference-framework
+2
11
Causal Attention：为什么 KV hit 后 Attention 按 1 - h² 缩放
ai-systems / llm-inference
llm-inferenceattentionkv-cachesimulator
+1
12
CSA/HCA 注意力：DeepSeek-V4 的混合压缩稀疏机制
ai-systems / llm-inference
attentionsparse-attentionkv-cachedeepseek