quantize_config.json

{
    "per_channel": true,
    "reduce_range": true,
    "per_model_config": {
        "decoder_with_past_model": {
            "op_types": [
                "MatMul",
                "Gemm",
                "Where",
                "ConstantOfShape",
                "Split",
                "Slice",
                "Pow",
                "Shape",
                "Unsqueeze",
                "Squeeze",
                "ReduceMean",
                "Reshape",
                "Mul",
                "Add",
                "Concat",
                "Sub",
                "Transpose",
                "Constant",
                "Div",
                "Gather",
                "Cast",
                "Tanh",
                "Sqrt",
                "Softmax",
                "Range"
            ],
            "weight_type": "QInt8"
        },
        "decoder_model": {
            "op_types": [
                "MatMul",
                "Gemm",
                "Where",
                "ConstantOfShape",
                "Split",
                "Slice",
                "Pow",
                "Shape",
                "Unsqueeze",
                "Squeeze",
                "ReduceMean",
                "Reshape",
                "Mul",
                "Add",
                "Concat",
                "Sub",
                "Transpose",
                "Constant",
                "Div",
                "Gather",
                "Cast",
                "Tanh",
                "Sqrt",
                "Softmax",
                "Range"
            ],
            "weight_type": "QInt8"
        },
        "encoder_model": {
            "op_types": [
                "MatMul",
                "Where",
                "ConstantOfShape",
                "Expand",
                "Slice",
                "Pow",
                "Shape",
                "Unsqueeze",
                "ReduceMean",
                "Reshape",
                "Equal",
                "Mul",
                "Add",
                "Conv",
                "Concat",
                "Sub",
                "Transpose",
                "Constant",
                "Gather",
                "Div",
                "Sqrt",
                "Softmax",
                "Erf"
            ],
            "weight_type": "QUInt8"
        },
        "decoder_model_merged": {
            "op_types": [
                "MatMul",
                "Gemm",
                "Where",
                "ConstantOfShape",
                "Split",
                "Slice",
                "Pow",
                "Shape",
                "Unsqueeze",
                "Squeeze",
                "ReduceMean",
                "Reshape",
                "Mul",
                "Add",
                "Concat",
                "Sub",
                "Transpose",
                "Constant",
                "Div",
                "Gather",
                "Cast",
                "Tanh",
                "Sqrt",
                "Softmax",
                "Range",
                "If"
            ],
            "weight_type": "QInt8"
        }
    }
}