Make compress_weights boolean parameter in AdvancedQuantizationParameters

Author: aobolensk

src/nncf/openvino/quantization/quantize_model.py

@@ -216,7 +216,7 @@ def quantize_with_accuracy_control_impl(
         copied_parameters = deepcopy(advanced_quantization_parameters)
 
     compress_weights = is_weight_compression_needed(copied_parameters)
-    copied_parameters.backend_params[AdvancedQuantizationParameters.COMPRESS_WEIGHTS] = False
+    copied_parameters.compress_weights = False
 
     quantized_model = quantize_impl(
         model=model,

src/nncf/quantization/advanced_parameters.py

@@ -256,10 +256,8 @@ class AdvancedQuantizationParameters:
     :type smooth_quant_alpha: float
     :param backend_params: Backend-specific parameters.
     :type backend_params: dict[str, Any]
-    :param COMPRESS_WEIGHTS: A key in the `backend_params` dictionary that indicates whether
-        weight compression should be applied. If set to False, weight compression is disabled.
-        By default, weight compression is enabled (True).
-    :type COMPRESS_WEIGHTS: str
+    :param compress_weights: Indicates whether to apply weight compression after quantization.
+    :type compress_weights: bool
     """
 
     # General parameters
@@ -274,6 +272,7 @@ class AdvancedQuantizationParameters:
     activations_quantization_params: Optional[Union[QuantizationParameters, FP8QuantizationParameters]] = None
     weights_quantization_params: Optional[Union[QuantizationParameters, FP8QuantizationParameters]] = None
     quantizer_propagation_rule: QuantizerPropagationRule = QuantizerPropagationRule.MERGE_ALL_IN_ONE
+    compress_weights: bool = True
 
     # Range estimator parameters
     activations_range_estimator_params: RangeEstimatorParameters = field(default_factory=RangeEstimatorParameters)
@@ -290,9 +289,6 @@ class AdvancedQuantizationParameters:
     # Backend specific parameters
     backend_params: dict[str, Any] = field(default_factory=dict)
 
-    # Backend parameter names
-    COMPRESS_WEIGHTS = "compress_weights"
-
 
 def is_weight_compression_needed(advanced_parameters: Optional[AdvancedQuantizationParameters]) -> bool:
     """
@@ -304,7 +300,7 @@ def is_weight_compression_needed(advanced_parameters: Optional[AdvancedQuantizat
     :return: True if weight compression is needed, False otherwise.
     """
     if advanced_parameters.backend_params is not None:
-        return bool(advanced_parameters.backend_params.get(AdvancedQuantizationParameters.COMPRESS_WEIGHTS, True))
+        return advanced_parameters.compress_weights
     return True
 
 

tests/torch2/fx/test_models.py

@@ -218,7 +218,8 @@ def transform_fn(data_item):
     calibration_dataset = nncf.Dataset([example_input], transform_fn)
 
     quantization_parameters["advanced_parameters"] = AdvancedQuantizationParameters(
-        disable_bias_correction=True, backend_params={AdvancedQuantizationParameters.COMPRESS_WEIGHTS: compress_weights}
+        disable_bias_correction=True,
+        compress_weights=compress_weights,
     )
     quantization_parameters["subset_size"] = 1