prefill

v. to fill in information on a form or document before you give it to someone. This is often done by a computer to save you time.

v. to populate fields in a digital form or document with data before it is presented to the user. Often describes automated processes that use existing database records to streamline data entry.

“The website will prefill your address if you are logged in.”

“Modern web browsers can prefill sensitive payment information, provided the user has granted permission to store such data within their encrypted profile.”

v. to read the whole prompt in one pass before an AI model writes its first token. This is the first step of AI inference.

v. In LLM inference, to process all prompt tokens in parallel and compute the key-value cache before generation begins. This phase is compute-intensive and sets how long a user waits for the first token.

“The server will prefill the prompt before the model writes its first word.”

“Because the chip can prefill a long prompt in parallel, reading a large document takes far less time per token than writing the answer.”

Technical