Optimization

Explains speculative decoding, which pairs a small draft model with a large target model to accelerate LLM inference without changing outputs.

Explains prefix caching for reusing attention KV computations to speed up shared-prefix AI inference.