One Request， End to End: Tokenize， Prefill， Decode， Stream
This article explains the four key steps of a language model inference engine — tokenization, prefilling, decoding, and streaming — using a factory analogy to illustrate how a user's question is processed into a visible answer.