Table of Contents

Also see the outline of the entire book as planned, including draft chapters that are not yet completed.

• Build DeepSeek from Scratch
  A comprehensive hands-on guide covering every major DeepSeek innovation — from KV Cache and Multi-Head Latent Attention to Mixture-of-Experts, Multi-Token Prediction, FP8 Quantization, Sparse Attention, Manifold-Constrained Hyper-Connections, Conditional Memory (Engram), and the million-token DeepSeek-V4 architecture.
• Part 1: Introduction
  • Chapter 1. Pengantar tentang DeepSeek
  • Chapter 2. Mengatasi Masalah Kemacetan Kinerja
  • Chapter 3. Mengatasi Masalah Kemacetan Kinerja
  • Chapter 4. Solusi Kemacetan Kinerja
  • Chapter 5. Jia Bin Huang DSA
  • Chapter 6. Jia Bin Huang DeepSeek V4
  • Chapter 7. Jia Bin Huang DeepSeek V4
• Part 2: Kv cache
  • Chapter 8. 2.1 The LLM inference loop: Generating text one token at a time
  • Chapter 9. Terobosan Hebat DeepSeek
  • Chapter 10. Meningkatkan Kecerdasan
  • Chapter 11. Menebak Banyak Kata Sekaligus
• Part 3: Mla
  • Chapter 12. 3.1 MLA: The best of both worlds
• Part 4: Moe
  • Chapter 13. 4.1 The intuition behind mixture of experts
• Part 5: Mtp fp8
  • Chapter 14. 5.1 The core idea: From single-token to multi-token prediction
  • Chapter 15. 5.2 The four key advantages of MTP
• Part 6: Dsa
  • Chapter 16. 6.1 DSA Prototype: Lightning Indexer and Fine-Grained Token Selection
  • Chapter 17. 6.2 DSA Continued Pre-Training: Warm-up and Sparse Stages
  • Chapter 18. 6.3 Parity Evaluation and Inference Cost Reduction
• Part 7: Papers
  • Chapter 19. Membedah DeepSeek-V4
  • Chapter 20. DeepSeek-v4 beyond basics
  • Chapter 21. DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models
  • Chapter 22. Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models
  • Chapter 23. mHC: Manifold-Constrained Hyper-Connections
  • Chapter 24. DeepSeek-V4
• 010 References