Speculative Streaming and Multi-Token Prediction Quiz

Answers and explanations appear as you take the quiz.

1. Why is standard autoregressive decoding often slow on a GPU?
   • Tokens are produced one at a time, leaving the GPU underused per step
   • The GPU runs out of memory immediately
   • It must retrain on each token
   • The vocabulary is too small
2. What does a 'drafter' do in speculative decoding?
   • It writes the final answer alone
   • It proposes several candidate future tokens cheaply
   • It deletes low-probability tokens
   • It compresses the model weights
3. How is output quality preserved in speculative decoding?
   • By rounding probabilities to zero
   • By using a larger vocabulary
   • By verification that accepts only tokens matching the target distribution
   • By skipping verification entirely
4. What distinguishes Medusa-style multi-token prediction from classic speculative decoding?
   • It requires two separate large models
   • It only works for image generation
   • It removes the need for any forward pass
   • It adds extra lightweight prediction heads so one model drafts and verifies
5. Why can a transformer verify many candidate positions almost as cheaply as one during decoding?
   • Because decoding is memory-bandwidth bound, not compute bound
   • Because it caches the entire internet
   • Because verification skips the attention layers
   • Because the GPU clock speed doubles
6. What primarily determines the speedup from speculative or multi-token methods?
   • The color of the GPU
   • The average number of tokens accepted per step
   • The length of the model's name
   • The size of the disk
7. A draft proposes tokens A, B and C, but target verification first disagrees at B. Which proposal prefix can be accepted?
   • A, B and C
   • B and C only
   • A, the matching prefix before the first mismatch
   • None, even if A matches
8. Why can speculative decoding preserve the target model’s output distribution?
   • The drafter is guaranteed to be identical to the target
   • The target model is never run
   • All proposed tokens are accepted unconditionally
   • A verification and correction rule accepts or rejects draft tokens according to the target distribution