Put the steps of one round of vanilla speculative decoding in the correct order

**Speculative decoding** speeds up generation by pairing a small, fast draft model with the large, accurate target model. The draft does the guessing; the target does the checking. The round runs in a fixed order. First the draft model autoregressively proposes K candidate tokens, one after another, just like normal decoding but cheap. Then the target model runs a single forward pass over all K proposed tokens at once, computing its own logits at every position in parallel. Next comes verification. The verifier walks the positions left to right and checks each draft token against the target's distribution using the speculative sampling rule. It accepts the longest matching prefix. At the first rejected position, the target samples a corrected token from its own adjusted distribution. If all K are accepted, the target gets one bonus token for free from the final position. Finally the accepted tokens plus the correction are appended to the output. The draft model's KV cache is rewound to the new tip, and the round repeats. The big win: the expensive model runs once per round, not once per token.

**Speculative decoding** trades cheap draft compute for fewer expensive target passes. One round has five ordered steps. The draft model proposes K tokens autoregressively. The target then runs one parallel forward pass over those K positions. The verifier checks each draft token left to right under the speculative sampling rule. It accepts the longest matching prefix. At the first rejection it resamples a corrected token from an adjusted target distribution. Accepted tokens plus the correction are committed, caches are rewound to the new tip, and the next round begins. Order matters for a reason. The single target pass must come after the full K-token proposal, because the target needs all candidate positions to verify them in parallel. Verification must precede the resample, since you only correct the first token the target rejects. The subtle part is correctness. The accept rule keeps each token with probability proportional to the target over draft ratio. On rejection it draws from the residual of target minus draft. This makes the committed sequence exactly distributed as the target alone. So speculative decoding is not an approximation. It returns identical quality output, only faster when draft guesses land.