quantize_config.json

{
    "per_channel": false,
    "reduce_range": false,
    "per_model_config": {
        "decoder_model": {
            "op_types": [
                "Unsqueeze",
                "Constant",
                "Gather",
                "Tanh",
                "LayerNormalization",
                "Shape",
                "Reshape",
                "Add",
                "Where",
                "Squeeze",
                "Cast",
                "Pow",
                "ConstantOfShape",
                "Split",
                "Gemm",
                "MatMul",
                "Slice",
                "Sub",
                "Transpose",
                "Concat",
                "Range",
                "Mul",
                "Softmax"
            ],
            "weight_type": "QInt8"
        }
    }
}