Speculative decoding

Decode faster without changing the output: a small draft model proposes several tokens, the big model verifies them in one parallel pass, accepting the prefix that matches its own distribution. Exactness preserved, 2-3× latency wins when the draft model guesses well.