InferQuest
← Quest map
Phase 2 · Engine Core

Speculative Decoding

Free tokens, provably distribution-preserving.

0/4 · 470 XP
You’re previewing the curriculum. Sign in (top right) to track progress, take knowledge checks, and unlock the verifiers.
Tasks
  • min(1, p/q) acceptance + residual resampling. Interviews test whether you can prove exactness.

    paper+70 XPresource ↗
  • EAGLE-3 is the shipping default in vLLM/SGLang/TRT-LLM. Also note MTP in DeepSeek-V3 as built-in speculation.

    paper+70 XPresource ↗
  • Implement the greedy draft-verify loop (draft k, verify in ONE target call, accept prefix, rollback, bonus token) against the harness's target + noisy draft. Graded on exact equality with pure target decoding and ≥1.5 tokens per verify call.

    build+250 XPauto-verified
  • Acceptance math, expected tokens/step, when speculation hurts. 75% to pass.

    drill+80 XPauto-verified