{
  "ok": true,
  "module_path": "model.layers.23.feed_forward.switch_mlp.down_proj",
  "stats": {
    "expert": 0,
    "changed_codes": 197683,
    "total_codes": 917504,
    "saturation_rate": 0.0,
    "loss_before": 1.0000001111620804e-06,
    "loss_after": 1.0007734090322629e-06
  },
  "peak_memory_gb": 0.323289108,
  "note": "This updates one quantized expert projection outside mx.gather_qmm and requantizes immediately."
}