You serve a 70B model with a 1B draft model. The draft proposes $k$ tokens one at a time, each draft step costing about 5% of a 70B forward pass, and the 70B model then scores all $k$ in one forward pass that costs about as much as generating one token, since decoding is memory-bound.

Left to right, each draft token is kept with probability $\min (1,p/q)$, where $p$ and $q$ are the 70B and draft models' probabilities for it. The first rejected token is resampled from the 70B model and the rest of the draft is discarded; if all $k$ survive, the 70B pass adds one more token.