LLM Handbook
A complete engineering guide to understanding and deploying Large Language Models.
Transformer internals KV cache economics MoE + routing DeepSeek LSA
01
LLM Handbook Part 1: Architecture, Attention Heads & Flash Attention
An engineer's guide to the core mechanics of Large Language Models. Demystifying the Transformer architecture, KV Cache, and how Flash Attention changed the game.
02
LLM Handbook Part 2: Dissecting Modern Architectures (MoE, MTP, GLM)
Exploring the cutting-edge architectural quirks that power modern LLMs: Mixture of Experts (MoE), Multi-Token Prediction (MTP), and GLM's bidirectional attention.
03
LLM Handbook Part 3: DeepSeek-V4, GQA & Lookahead Sparse Attention
How DeepSeek-V4 uses Grouped Query Attention and Neural Memory Indexers to achieve 500K context windows with a 90% VRAM reduction.