추론적 디코딩 퀴즈

퀴즈를 풀면 정답과 설명이 나타납니다.

1. How does speculative decoding speed up generation from a large model?
   • 소규모 초안 모델은 대규모 모델이 병렬로 검증하는 여러 토큰을 추측합니다.
   • 모델 크기를 영구적으로 절반으로 압축
   • 주의 메커니즘 제거
   • 더 많은 데이터로 모델 학습
2. For a large model at low batch size, why can one-token autoregressive decoding be slow?
   • Each token requires internet access
   • Repeated reads of large model weights can make decoding memory-bandwidth bound
   • The model permanently stores a full copy of every possible output
   • The tokenizer performs the model’s matrix multiplications
3. In greedy speculative decoding, what happens at the first token where draft and target predictions disagree?
   • The whole conversation is thrown away and restarted from scratch
   • 초안 모델이 즉시 재학습됩니다.
   • The matching prefix is kept, the target token is used at the mismatch, and later draft tokens are discarded
   • 출력 품질이 영구적으로 저하됩니다.
4. Under exact speculative sampling, how does the output distribution compare with sampling from the target model alone?
   • It is always lower quality but faster
   • It is identical only when the draft model is larger
   • It becomes deterministic regardless of target sampling
   • It is identical in distribution
5. 추측적 디코딩으로 인한 속도 향상을 가장 직접적으로 결정하는 것은 무엇입니까?
   • 초안 모델 추측의 합격률
   • GPU의 색상
   • 모델명의 길이
   • 시간
6. Medusa와 EAGLE의 접근 방식에는 어떤 공통점이 있나요?
   • 양자 하드웨어가 필요합니다
   • 별도의 독립형 초안 모델이 필요하지 않습니다.
   • 대상 모델을 삭제합니다.
   • 이미지에만 작업합니다
7. What should a serving team measure before claiming speculative decoding will speed up its workload?
   • Only the target model’s parameter count
   • Only whether the draft produces fluent text
   • Draft cost, target verification cost, accepted tokens and end-to-end latency under representative traffic
   • The number of tokenizer files
8. In exact speculative sampling, what happens when a proposed token is rejected?
   • The whole conversation is discarded
   • The target model is permanently changed
   • The rejected token is kept without verification
   • A correction step samples from the distribution needed to preserve the target model’s output distribution