Skip to content
← Speculative & Constrained Decoding / 19
expertNewSnowflakeBasetenQualcomm

Block verification accepts more drafts than token-by-token rejection with the same drafter. Why, and when does it gain nothing?

Standard speculative sampling throws away accepted mass every time one token's ratio is clipped at 1. The answer derives the joint rule that keeps it, works the two-token example by hand, quotes the paper's measured gain with its setup, and proves the case where vLLM's default drafting makes the switch a no-op.

Updated Oct 2026 · Worked answers on inference mechanisms, with calculations, implementation details, and follow-up questions.

Standard speculative sampling throws away accepted mass every time one token's ratio is clipped at 1. The answer derives the joint rule that keeps it, works the two-token example by hand, quotes the paper's measured gain with its setup, and proves the case where vLLM's default drafting makes the switch a no-op.

10 answers per topic instead of 5, plus saved progress and bookmarks · costs nothingor unlock all 213 remaining answers · $25

Every worked answer and derivation in all 12 topics, every concept page, the full practice-test bank, and every code drill’s edge-case tests and reference solution.

6 months · One payment · No auto-renewal

UP NEXT ON YOUR JOURNEY
Next in this trackImplement a JSON-schema token mask: walk the vocabulary through the automaton, pack it into int32 words, apply it to logits.Next in this trackA diffusion LLM emits 15 to 20 tokens per forward. Is it 15x faster than autoregressive decode? Derive where it wins.Next in this trackAn RL rollout engine speculates with an EAGLE head trained on the step-0 policy, and rollouts are no faster now. Why?
EDITOR'S NOTE

Most candidates can recite min(1, p/q). The screen is whether they can say why the per-token clip is lossy and why a one-hot draft never clips.

DISCUSSION · 0

Nobody has commented on this answer. Add a correction or another way to work it.