{
  "id": "int8-quantization",
  "title": "Symmetric int8 quantization",
  "kind": "implement",
  "summary": "Absmax int8 quantization with one scale per tensor and one scale per row, with dequantization and a zero-row rule.",
  "difficulty": "medium",
  "tags": ["quantization", "int8", "inference"],
  "lesson": "phases/10-llms-from-scratch/11-quantization",
  "cheatsheet": "quantization",
  "minutes": 20,
  "entry": ["quantize_symmetric", "dequantize_symmetric", "quantize_per_channel", "dequantize_per_channel"],
  "collections": ["llm-inference"],
  "time_limit_s": 10,
  "mutants": {
    "columns_not_rows": "Takes the maximum over axis 0, so it gives one scale for each column, not for each row.",
    "max_not_absmax": "Uses the maximum of the values, not of their magnitudes, for the per-tensor scale.",
    "round_half_up": "Rounds with floor(x + 0.5), so values that are exactly halfway round up and not to the even integer.",
    "scale_by_128": "Divides the maximum by 128, not by 127.",
    "truncate_per_row": "Casts the per-row values to int8 without a round and without a clip, so the fraction is cut off.",
    "zero_row_divides": "Has no rule for a row of zeros, so 0 / 0 gives NaN."
  }
}