"Quantization and Accelerated Inference" — a practical and applied guide to optimizing AI models for faster, lighter, and more cost-effective inference without complicating the architecture. The material helps to understand how reducing numerical precision in calculations speeds up model performance, reduces memory consumption, and decreases infrastructure costs with minimal loss of quality.
What You Will Learn from the Course
The material is structured as a complete optimization pipeline — from basic theory to production-ready solutions. The book explains the key concepts of quantization and demonstrates how to apply them in real ML projects.
Key Topics
- Post-training quantization (PTQ) — reducing precision without retraining;
- Quantization-aware training (QAT) — preparing models for quantization during training;
- Fake quantization and the use of straight-through estimators;
- Working with LLM: activation spikes, KV-cache optimization, formats below 8 bits (NF4, FP4);
- Constructing correct quantization mapping and analyzing trade-offs.
Approach and Learning Structure
The book is targeted at engineers and researchers working with Python and modern ML frameworks. The material is carefully structured and suitable for both implementing quantization from scratch and optimizing existing models.
Practical Orientation
- framework-agnostic methods and recommendations;
- cross-framework testing and strategy comparison;
- decision-making schemes for choosing the level of quantization;
- checklists for preparing models for deployment;
- real-world examples of optimizing large models.
Who It's For
The material will be especially useful for ML engineers, researchers, and practitioners aiming to reduce inference costs, speed up ML feature production, and implement modern model optimization methods without rewriting the architecture.
