ai-engineering-from-scratch-learning/phases/19-capstone-projects/14-speculative-decoding-server/quiz.json at main · easterCat/ai-engineering-from-scratch-learning · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
{
  "lesson": "14-speculative-decoding-server",
  "title": "Capstone 14 — Speculative-Decoding Inference Server",
  "questions": [
    {
      "stage": "pre",
      "question": "What is the two-layer mechanism behind speculative decoding?",
      "options": [
        "Two target models alternate per token",
        "An embedding model and a synthesizer",
        "A retrieval cache and a reranker",
        "A draft model proposes k candidate tokens; the target model verifies them in a single pass"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "pre",
      "question": "What does EAGLE-3 do that ngram drafts do not?",
      "options": [
        "It replaces the target model",
        "It trains draft heads on the target model's hidden states for higher acceptance rates",
        "It avoids tokenization entirely",
        "It runs only on CPUs"
      ],
      "correct": 1,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Why is p99 tail latency reported across batch sizes 1, 8, and 32?",
      "options": [
        "Steady-state tokens-per-second can hide that the verify pass on rejection is more expensive than vanilla decoding",
        "vLLM cannot serve batch sizes below 8",
        "p99 only matters at batch size 32",
        "Tail latency is independent of batch size"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Why does acceptance rate drift when the traffic distribution shifts?",
      "options": [
        "vLLM rotates models hourly",
        "Draft alignment depends on the input distribution: ShareGPT, code, and domain data exercise different patterns",
        "The OS scheduler reshuffles GPU memory",
        "Quantization changes at runtime"
      ],
      "correct": 1,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Which Kubernetes HPA signal does the deployment scale on?",
      "options": [
        "Cluster autoscaler nodes",
        "CPU utilization",
        "Disk IOPS",
        "Queue-wait time on inference requests"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "What does P-EAGLE add over serial EAGLE-3?",
      "options": [
        "Quantization to FP4 weights",
        "Removal of the target model",
        "Parallel speculation across a tree of draft branches verified in one target pass",
        "Automatic dataset curation"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "Which throughput target does the rubric demand against the non-speculative baseline?",
      "options": [
        "At least 2.5x at matched quality on two models",
        "Any improvement above baseline",
        "Roughly 1.1x at matched quality",
        "Exactly 10x throughput at any quality"
      ],
      "correct": 0,
      "explanation": ""
    }
  ]
}