def postprocess_quantization_parameters()

in src/optimum/nvidia/utils/cli.py [0:0]

• 18 lines of code
• 3 McCabe index (conditional complexity)

def postprocess_quantization_parameters(params: Namespace) -> Namespace:
    # Only support FP8 quantization for now
    qconfig = QuantMode.from_description(
        quantize_weights=False,
        quantize_activations=False,
        per_token=False,
        per_channel=False,
        per_group=False,
        use_int4_weights=False,
        use_int8_kv_cache=False,
        use_fp8_kv_cache=params.fp8_cache,
        use_fp8_qdq=params.fp8,
    )

    params.has_quantization_step = qconfig != QuantMode(0)
    params.quantization_config = qconfig

    # If we do have the output path, then let's create the calibration path
    if "output" in params:
        from pathlib import Path

        params.calibration_output = Path(params.output).joinpath("calibration")

    return params