Glossary
Speculative decoding
A cheap drafter, often a small model, proposes several tokens and the large model verifies them in one pass. The output distribution is preserved, and the speed-up depends on how often drafts are accepted.
Where it is explained
See also
Terms explained in the same lessons.