A veterans-memorial kiosk sends raw English bytes into a TensorRT-LLM-optimized engine that expects token IDs, then prints integer IDs to the visitor. What is missing from the live pipeline?
Select an answer to reveal the explanation.
Short Explanation
The engine wants token IDs in and token IDs out. Encode with a tokenizer before it, decode after it. Posting raw English, printing integers to the visitor, or writing a custom kernel leaves the live path unfinished.
Full Explanation
An optimized LLM engine consumes token IDs and emits token IDs. The live path is therefore tokenize, infer, then decode, whether those stages live in the app or in a server in front of the engine. Posting raw English bytes, or showing integer IDs to a visitor, is a broken pipeline. Authoring a custom kernel is outside this associate serving task.