01
P/D 分离：从 KV 交接到路由、背压与 SLO
ai-systems / llm-inference
llm-inferenceprefilldecodekv-cache
+2
02
分层 KV Cache：从 GPU 命中到外部存储复用
ai-systems / llm-inference
llm-inferencekv-cacheprefillnetworks
+1
03
KV Cache Hit Ratio 修正模型：从直觉到统一公式
ai-systems / llm-inference
llm-inferencekv-cachesimulatorprefill
+1
04
量化：从 Scale 到 W4A8 的完整坐标
ai-systems / llm-inference
llm-inferencequantizationprefilldecode
+1
05
Prefill Trace：Worker 供给、DSA/MLA 与 Chunked Prefill
ai-systems / llm-inference
llm-inferenceprefilltracemla
+2
06
GDN 与 Chunked Prefill：为什么 prepare_chunk_indices 会出现在 trace 里
ai-systems / llm-inference
llm-inferencegdnqwen3nextchunked-prefill
+3
07
Chunked Prefill 深入分析：调度、Chunk Size 与 Attention 形状
ai-systems / llm-inference
llm-inferencechunked-prefillschedulingprefill
+3
08
Causal Attention：为什么 KV hit 后 Attention 按 1 - h² 缩放
ai-systems / llm-inference
llm-inferenceattentionkv-cachesimulator
+1
09
Compute-bound vs Memory-bound：推理的两大瓶颈
ai-systems / llm-inference
LLMInferencePerformanceGPU
+3