Why no execution provider selection present for quantize_static when available in CalibratorBase?

[prikmm]

Hi Community,

I am currently understanding the quantization api of ONNXRuntime and got curious when I found no execution provider selection is given in quantize_static whereas CalibratorBase has set_execution_provider function present.

onnxruntime/onnxruntime/python/tools/quantization/calibrate.py

Line 70 in b713855

def set_execution_providers(self, execution_providers=['CPUExecutionProvider']):

Won't adding this support to quantize_static make quantization for an execution provider better as inferred output, through which calibration range is obtained, become more execution provider/hardware driven?

I was thinking of something like

def quantize_static(
    model_input,
    model_output,
    calibration_data_reader: CalibrationDataReader,
    quant_format=QuantFormat.QDQ,
    op_types_to_quantize=[],
    per_channel=False,
    reduce_range=False,
    activation_type=QuantType.QInt8,
    weight_type=QuantType.QInt8,
    nodes_to_quantize=[],
    nodes_to_exclude=[],
    optimize_model=True,
    use_external_data_format=False,
    calibrate_method=CalibrationMethod.MinMax,
    extra_options = {},
    execution_providers = ["CPUExecutionProvider"],  # -----> This way
):
    .
    .
    .
    .
    .
    calibrator = create_calibrator(
        model,
        op_types_to_quantize,
        calibrate_method=calibrate_method,
        use_external_data_format=use_external_data_format,
        extra_options=calib_extra_options
    )
    calibrator.set_execution_providers(execution_providers)  # ---------> And then this
    calibrator.collect_data(calibration_data_reader)
    .
    .
    .
    .
    .
 

I am really curious why the ONNXRuntime contributors didn't add this as it seems very simple. Please let me know if I am missing something.

[prikmm]

Let me know if I can do a PR regarding this.