quantize_config.json

{
    "fp16": {},
    "q8": {
        "per_model_config": {
            "model": {
                "op_types": [
                    "Add",
                    "Conv",
                    "Flatten",
                    "Gemm",
                    "GlobalAveragePool",
                    "HardSigmoid",
                    "Mul",
                    "ReduceMean",
                    "Relu"
                ],
                "weight_type": "QUInt8"
            }
        },
        "per_channel": true,
        "reduce_range": true
    },
    "int8": {
        "per_model_config": {
            "model": {
                "op_types": [
                    "Add",
                    "Conv",
                    "Flatten",
                    "Gemm",
                    "GlobalAveragePool",
                    "HardSigmoid",
                    "Mul",
                    "ReduceMean",
                    "Relu"
                ],
                "weight_type": "QInt8"
            }
        },
        "per_channel": true,
        "reduce_range": true
    },
    "uint8": {
        "per_model_config": {
            "model": {
                "op_types": [
                    "Add",
                    "Conv",
                    "Flatten",
                    "Gemm",
                    "GlobalAveragePool",
                    "HardSigmoid",
                    "Mul",
                    "ReduceMean",
                    "Relu"
                ],
                "weight_type": "QUInt8"
            }
        },
        "per_channel": true,
        "reduce_range": true
    },
    "q4": {
        "block_size": 32,
        "is_symmetric": true,
        "accuracy_level": null
    },
    "bnb4": {
        "block_size": 64,
        "quant_type": 1
    }
}