Learn how speculative decoding speeds up LLMs using a draft-and-verify pipeline. Discover the math behind rejection sampling, Medusa architecture, and implementation tips for production.