quantization_config.json

{
  "file": "head_int8.safetensors",
  "scope": "head only; base is separate and unchanged",
  "weight_formula": "w[i].float() * scale[i][:, None]",
  "bias_dtype": "float32",
  "runtime": "inference.py dequantizes to float32; no INT8 acceleration is claimed"
}