{
  "lesson": "37-loading-pretrained-weights",
  "title": "Loading Pretrained Weights",
  "questions": [
    {
      "stage": "pre",
      "question": "Why does the loader need a name map?",
      "options": [
        "Published GPT-2 weights use names like wte, wpe, h.N.attn.c_attn; the local model uses tok_embed, pos_embed, blocks.N.attn.qkv; the loader translates one onto the other",
        "Safetensors files are encrypted",
        "Pretrained tensors and local parameters share the same name",
        "Torch requires it"
      ],
      "correct": 0,
      "explanation": "The two naming conventions are different histories; the map is a literal dict expanded per layer."
    },
    {
      "stage": "check",
      "question": "What is the conv1d transpose, and which tensors need it?",
      "options": [
        "All weight tensors",
        "Published GPT-2 stores c_attn, c_proj, and c_fc weights in tensorflow conv1d layout, which is the transpose of what nn.Linear.weight expects; the loader calls .t() on those during assignment",
        "Only the embedding",
        "Only the LayerNorm scales"
      ],
      "correct": 1,
      "explanation": "Three suffixes need transposing; biases and LayerNorms are already shaped correctly."
    },
    {
      "stage": "check",
      "question": "Why is the LM head not in the safetensors file?",
      "options": [
        "It is too large",
        "It is loaded from a separate file",
        "It is computed at runtime from logits",
        "Weight tying: the LM head shares storage with the token embedding (wte), so loading wte and aliasing lm_head.weight = tok_embed.weight after the load reconstructs the head"
      ],
      "correct": 3,
      "explanation": "Setting lm_head.weight = tok_embed.weight after wte lands restores the alias; copying instead doubles the parameter count."
    },
    {
      "stage": "check",
      "question": "What does the loader do on a shape mismatch?",
      "options": [
        "Pads the smaller tensor",
        "Crashes the process",
        "Records the mismatch in LoadReport.shape_mismatch and refuses to assign, leaving the local parameter untouched",
        "Truncates the larger tensor"
      ],
      "correct": 2,
      "explanation": "Half loaded models are silent failure machines; refusing to assign keeps the report honest."
    },
    {
      "stage": "post",
      "question": "Why is the demo's stub fixture not a real GPT-2 download?",
      "options": [
        "Real files are not in safetensors format",
        "Real weights are encrypted",
        "The demo must run offline and quickly; the stub uses the exact pretrained naming convention so the loader code path is identical, and swapping in a real file changes nothing about the loader",
        "Safetensors cannot store large tensors"
      ],
      "correct": 2,
      "explanation": "Same code path, smaller payload; the loader works on either."
    },
    {
      "stage": "post",
      "question": "Why generate a sample before and after the load?",
      "options": [
        "To measure inference speed",
        "To set the random seed",
        "If the post-load samples are identical to the pre-load samples, the load did not change the model, which means the mapping silently missed every tensor",
        "To warm up the GPU cache"
      ],
      "correct": 2,
      "explanation": "Sanity generation is the cheapest gate: a sample fingerprint that fails to change exposes a broken load that the LoadReport summary would not catch on its own."
    }
  ]
}