最新博客文章
3 / 200AI 开发实战教程，从零到一构建你的 AI 应用。
All200
Beginner32
Intermediate92
Advanced76
All
Prompt Engineering12
RAG Retrieval Augmented17
Agent Development34
Model Fine-tuning21
Deployment and Operations25
AI Application Development40
DeepSeek API16
Multimodal Development6
AI Safety6
Model Inference3
Retrieval Systems1
RAG Retrieval1
Model Architecture3
API Development1
Agent Development10
Model Deployment1
Multimodal1
Model Evaluation2
Tutorial
Date
Long-Context Architecture: RoPE Extrapolation, Sparse Attention, and KV Cache
Model Architecture
Advanced
A context window looks like one number, but it is really two: what you can feed in and what the model can effectively use. Three levers decide whether long context holds — RoPE extrapolation, attention sparsity, and the KV Cache lifecycle from memory accounting to quantization and eviction — plus how Prefill differs from Decode, and where each trick stops working.
DeepSeek-V4.1-Flash: CED Asymmetric Architecture & KV Cache Compression
Model Architecture
Advanced
DeepSeek-V4.1-Flash redefines Agent-era inference with a 552B MoE and CED asymmetric architecture. We break down KV Cache compression and its cost model.
MoE Architectures in Practice: Routing, Load Balancing, and Expert Parallelism
Model Architecture
Advanced
MoE is sold as cheaper compute per token, but its real effect is a change in communication topology. What separates dense from sparse activation, top-k softmax gating, the capacity factor and the token-dropping tax, why the load-balancing loss conflicts with the main objective, gradient-free bias balancing, shared experts, and the all-to-all traffic behind expert parallelism.
