base_model_int8/quantization_config.json

{
  "quantization_method": "int8_per_channel_symmetric",
  "skip_patterns": [
    "final_layer.linear"
  ],
  "description": "Weight-only INT8 quantization with per-channel symmetric scaling"
}