quantization_report.json

{
  "timestamp_utc": "2026-02-09T20:38:09.856840+00:00",
  "source_model_dir": "artifacts/medasr-mlx-fp16",
  "output_model_dir": "artifacts/medasr-mlx-int8",
  "quantization": {
    "bits": 8,
    "group_size": 64,
    "mode": "affine",
    "target_modules": "mlx.nn.quantize default predicate (Linear/Embedding layers)"
  },
  "timing": {
    "quantization_time_s": 0.0034
  },
  "size_mb": {
    "source_weights": 200.9,
    "output_weights": 121.108,
    "compression_ratio_x": 1.659,
    "reduction_percent": 39.72
  }
}