onnx_int8/quantize_config.json

{
  "per_channel": true,
  "reduce_range": true,
  "is_symmetric": true,
  "weight_type": "QInt8"
}