onnx/quantize_config.json

{
    "per_channel": true,
    "reduce_range": true,
    "per_model_config": {
        "model": {
            "op_types": [
                "Erf",
                "Transpose",
                "Cast",
                "Softmax",
                "Gather",
                "Mul",
                "ReduceMean",
                "Unsqueeze",
                "Reshape",
                "Sqrt",
                "Slice",
                "Constant",
                "MatMul",
                "Concat",
                "Div",
                "Add",
                "Pow",
                "Sub",
                "Shape"
            ],
            "weight_type": "QInt8"
        }
    }
}