"AWQ Quantization: Shipping 4-Bit LLMs Without Quality Face-Plants"
Large language models rarely fail at 4-bit in obvious ways; they fail in production, under real prompts, on real hardware, and often only after teams have already celebrated the memory savings. This book is for experienced ML engineers, inference specialists, and platform builders who want to deploy AWQ-quantized models with confidence rather than folklore. It treats AWQ not as a buzzword or benchmark trick, but as a serious engineering discipline for production-grade LLM serving.
Across the book, readers will build a deep understanding of AWQ's activation-aware algorithm, calibration and search workflows, group size and zero-point choices, artifact formats, and the kernel realities that determine whether 4-bit models are actually faster. The coverage extends from quality evaluation and long-context failure modes to Hugging Face Transformers integration, ecosystem drift, legacy AutoAWQ migration, and serving-stack compatibility. By the end, readers will be able to judge when AWQ is appropriate, produce reproducible artifacts, benchmark honestly, and ship models that preserve quality under operational pressure.
The book assumes strong familiarity with modern LLM inference, GPU serving, and quantization basics. Its distinguishing feature is systems-level rigor: every major topic is tied to deployment decisions, failure analysis, and maintainable production workflows rather than isolated theory or toy examples.
