Press and sheet = prefillGalley slugs = KV cacheCarriage = one weight sweep per tokenOverflow tray = system RAMAn old print shop. The press prints your whole prompt in one stroke: prefill. A carriage sweeps the type case (the weights) once per letter: decode. Watch the sweep.Prefill sends every prompt token through the model in one parallel pass, so its time (TTFT) is set by the GPU's tensor FLOPs; a 32k-token prompt on a small card makes the platen strain for seconds. The printed sheet is laid into the galley as slugs: that is the KV cache, and one slug always holds 512 tokens (past 14 slugs the rows stack higher, and the plate on the galley says how high). Decode makes one token per pass, and every pass sweeps the entire type case, re-reading all the active weights plus the cache so far, so tokens per second are set by memory bandwidth; the counter on the crate counts the tokens (the crate is full at 2,048). A cache bigger than the memory left beside the weights runs past the stop bar into the overflow tray. Weights bigger than the card move part of the type case into the tray (system RAM, or the SSD on unified-memory machines), where the carriage crawls. Weights are GGUF Q4_K_M; prefill assumes 50% of peak dense FP16 tensor FLOPs, decode 70% of peak bandwidth.