Prefill vs Decode: LLM Inference Phases Explained

Every LLM request runs in two distinct phases: prefill, where the model reads your prompt in one parallel burst, and decode, where it generates the response one token at a time, each one depending on the last. These two phases have different performan...