引导式测验 · 中等 等级

推测性解码草案模型测验

探索草稿模型提案、目标验证、拒绝采样、接受概率、奖励令牌和依赖于工作负载的解码速度。

相关引导路径推测性解码草案模型
问题 1 的 8

“草案”模型在推测解码中的主要作用是什么？

预览每个问题8 问题

当您参加测验时，就会出现答案和解释。

  1. “草案”模型在推测解码中的主要作用是什么？
    • 将大型模型永久压缩为更少的参数
    • 将提示翻译成其他语言
    • 快速提出多个候选token供大型模型验证
    • 存储过去的对话以供检索
  2. 为什么在合适的工作负载下，验证草稿 Token 块能够加速目标模型的生成？
    • Verification has no compute cost at any block size
    • A verification pass can process several positions together and amortize target work across accepted tokens
    • The draft model is guaranteed more accurate than the target
    • The target model is skipped for every accepted token
  3. 在目标模型拒绝第一个令牌之后，起草的令牌会发生什么？
    • 第一次不匹配之后的所有内容都将被丢弃
    • 无论如何他们都被接受了
    • 他们被送到第三个模型
    • 它们被缓存并在下一轮重用
  4. 标准的精确推测采样拒绝程序在其假设下保持了什么样的正确性特性？
    • An identical random sequence for every possible random seed
    • Guaranteed truth of every generated statement
    • The draft model's distribution regardless of the target
    • The target model's configured sampling distribution
  5. 什么机制将 EAGLE 的草稿生成方法与 Medusa 的附加解码头区分开来？
    • EAGLE performs only word-count-based lexical retrieval
    • EAGLE drafts autoregressively at the target model's feature level
    • EAGLE removes target verification from every algorithm
    • EAGLE and Medusa are the same architecture with different names
  6. Why should a speculative-decoding benchmark report the draft-target pair and acceptance rate?
    • Every model pair has the same guaranteed speedup
    • Acceptance rate measures whether generated facts are true
    • Speedup depends on drafting and verification costs and how many proposed tokens are accepted
    • The target model's hardware and context length cannot affect performance
  7. In exact speculative sampling, a proposed token has q equal to 0.2 and p equal to 0.1. What is its acceptance probability?
    • 0.1, because only p is used
    • 0.2, because only q is used
    • 1.0, because every proposed token is accepted
    • 0.5, from min(1, p/q)
  8. What can the standard speculative-sampling algorithm emit when every token in a drafted block is accepted?
    • The accepted block plus an additional token sampled from the target
    • Only the last drafted token, discarding the accepted prefix
    • An arbitrary extra draft token without target sampling
    • No output until the draft model is retrained