"Speculative Decoding Systems: Faster Generation with Draft Models and Safety Checks"
Large language models have made generation powerful, but not fast enough for many serious systems. This book is written for experienced ML engineers, inference researchers, and platform architects who need to understand why decoding remains the dominant bottleneck—and how speculative decoding changes the performance equation without surrendering correctness. Rather than treating speedup as a black-box trick, it approaches speculative decoding as a full systems discipline spanning algorithms, serving infrastructure, and operational constraints.
Readers will learn the exact mechanics of lossless draft-and-verify decoding, the acceptance rules that preserve target-model behavior, and the design trade-offs behind high-performance draft models. The book then moves into performance modeling, scheduler and KV-cache interactions, self-speculation, Medusa-style multi-token heads, tree verification, and safety-aware guarded generation. It also translates theory into practice through implementation guidance, framework realities such as vLLM support, benchmarking strategy, and version-sensitive operational caveats, equipping readers to evaluate, deploy, and tune speculative systems with rigor.
The presentation assumes strong familiarity with modern transformer inference, sampling, and production serving concepts. Its distinguishing focus is depth: every chapter connects formal guarantees to real deployment regimes, hidden failure modes, and decision criteria that matter in production.
