{
  "lesson": "09-production-quantization",
  "title": "Production Quantization — AWQ, GPTQ, GGUF K-quants, FP8, MXFP4/NVFP4",
  "questions": [
    {
      "stage": "pre",
      "question": "Which quantization format does the lesson call the production default for CPU and edge serving?",
      "options": [
        "GGUF Q4_K_M / Q5_K_M",
        "NVFP4",
        "FP8",
        "AWQ INT4"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Which format is the lesson's pick for datacenter GPU serving when multi-LoRA is required in vLLM?",
      "options": [
        "NVFP4",
        "GGUF Q4_K_M",
        "GPTQ with Marlin kernels",
        "AWQ"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "What is the \"my model is 4 GB now\" trap with AWQ?",
      "options": [
        "AWQ does not actually shrink weights",
        "AWQ is incompatible with vLLM",
        "AWQ only shrinks weights; KV cache and activations are separate and can add 30-50 GB at production batch sizes",
        "AWQ requires INT8 KV cache"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Why does calibrating AWQ on generic web text hurt domain models?",
      "options": [
        "It increases model size",
        "It only works on AMD GPUs",
        "The algorithm makes wrong decisions about which weights to protect, dropping domain accuracy (for example several Pass@1 points on HumanEval)",
        "It disables Marlin kernels"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "For a reasoning-heavy workload where quality is non-negotiable, which precision does the lesson recommend by default?",
      "options": [
        "NVFP4 weights",
        "GPTQ INT4",
        "FP8 weights",
        "INT2 GGUF"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "Which 2026 quantization limitation does the lesson call out for NVFP4 in early 2026?",
      "options": [
        "Cannot be combined with FP8 KV cache",
        "Only runs on CPU",
        "No LoRA support yet",
        "Not supported on H100"
      ],
      "correct": 2,
      "explanation": ""
    }
  ]
}