quantize_config.json

{
    "modes": [
        "fp16",
        "q8",
        "int8",
        "uint8",
        "q4",
        "q4f16",
        "bnb4"
    ],
    "per_channel": true,
    "reduce_range": true,
    "block_size": null,
    "is_symmetric": true,
    "accuracy_level": null,
    "quant_type": 1
}