"TensorRT-LLM Optimization: Quantization, Kernel Fusion, and Throughput Engineering"
Built for experienced ML systems engineers, inference specialists, and GPU performance practitioners, this book is a deep guide to making large language models run faster, cheaper, and more predictably with TensorRT-LLM. Rather than offering generic acceleration advice, it develops a precise mental model of the TensorRT-LLM stack so readers can understand where performance is won or lost: in quantization choices, graph compilation, fused kernels, KV-cache policy, and serving scheduler behavior.
The book covers the full optimization path from precision strategy and post-training quantization pipelines to engine build configuration, plugin-enabled fusion, attention specialization, and throughput-oriented serving design. Readers will learn how to choose among FP16, BF16, FP8, INT8, and INT4 in hardware-aware ways; validate deployable quantized artifacts; realize fused execution paths in compiled engines; engineer KV-cache behavior for long-context workloads; and benchmark and profile systems with enough rigor to attribute gains to the right layer.
Structured as an advanced, implementation-minded text, the book emphasizes cross-layer tradeoffs rather than isolated tricks. It assumes solid familiarity with transformer inference, CUDA-era GPU concepts, and production deployment concerns, and rewards readers who want durable optimization judgment instead of version-fragile recipes."
