Move is_weight_compression_needed function to common

src/nncf/experimental/torch/fx/quantization/quantize_model.py

@@ -25,7 +25,6 @@
 from nncf.common.logging import nncf_logger
 from nncf.common.quantization.structs import QuantizationPreset
 from nncf.data import Dataset
-from nncf.experimental.torch.fx.quantization.backend_parameters import is_weight_compression_needed
 from nncf.experimental.torch.fx.transformations import DuplicateDQPassNoAnnotations
 from nncf.experimental.torch.fx.transformations import apply_quantization_transformations
 from nncf.experimental.torch.fx.transformations import compress_post_quantize_transformation
@@ -39,6 +38,7 @@
 from nncf.parameters import TargetDevice
 from nncf.quantization.advanced_parameters import AdvancedCompressionParameters
 from nncf.quantization.advanced_parameters import AdvancedQuantizationParameters
+from nncf.quantization.advanced_parameters import is_weight_compression_needed
 from nncf.quantization.algorithms.post_training.algorithm import PostTrainingQuantization
 from nncf.quantization.algorithms.weight_compression.algorithm import WeightCompression
 from nncf.scopes import IgnoredScope

src/nncf/onnx/quantization/backend_parameters.py

@@ -19,28 +19,11 @@ class BackendParameters:
     """
     :param EXTERNAL_DATA_DIR: An absolute path to the directory where the external data
         files are stored. All external data files must be located in the same folder.
-    :param COMPRESS_WEIGHTS: If `True` compresses constant quantized weights by folding
-        `QuantizeLinear` nodes into pre-quantized initializers. If `False`, this transformation
-        is skipped.
     """
 
-    COMPRESS_WEIGHTS = "compress_weights"
     EXTERNAL_DATA_DIR = "external_data_dir"
 
 
-def is_weight_compression_needed(advanced_parameters: Optional[AdvancedQuantizationParameters]) -> bool:
-    """
-    Determines whether weight compression is needed based on the provided
-    advanced quantization parameters.
-
-    :param advanced_parameters: Advanced quantization parameters.
-    :return: `True` if weight compression is needed, `False` otherwise.
-    """
-    if advanced_parameters is not None and advanced_parameters.backend_params is not None:
-        return advanced_parameters.backend_params.get(BackendParameters.COMPRESS_WEIGHTS, True)
-    return True
-
-
 def get_external_data_dir(
     advanced_parameters: Optional[Union[AdvancedQuantizationParameters, AdvancedCompressionParameters]],
 ) -> Optional[str]:

src/nncf/onnx/quantization/quantize_model.py

@@ -32,9 +32,7 @@
 from nncf.onnx.graph.nncf_graph_builder import GraphConverter
 from nncf.onnx.graph.passes import apply_preprocess_passes
 from nncf.onnx.graph.passes import compress_quantize_weights_transformation
-from nncf.onnx.quantization.backend_parameters import BackendParameters
 from nncf.onnx.quantization.backend_parameters import get_external_data_dir
-from nncf.onnx.quantization.backend_parameters import is_weight_compression_needed
 from nncf.parameters import BackupMode
 from nncf.parameters import CompressionFormat
 from nncf.parameters import CompressWeightsMode
@@ -47,6 +45,7 @@
 from nncf.quantization.advanced_parameters import AdvancedCompressionParameters
 from nncf.quantization.advanced_parameters import AdvancedQuantizationParameters
 from nncf.quantization.advanced_parameters import QuantizationParameters
+from nncf.quantization.advanced_parameters import is_weight_compression_needed
 from nncf.quantization.algorithms.accuracy_control.algorithm import QuantizationAccuracyRestorer
 from nncf.quantization.algorithms.accuracy_control.algorithm import calculate_accuracy_drop
 from nncf.quantization.algorithms.accuracy_control.evaluator import Evaluator
@@ -215,7 +214,7 @@ def quantize_with_accuracy_control_impl(
         copied_parameters = AdvancedQuantizationParameters()
     else:
         copied_parameters = deepcopy(advanced_quantization_parameters)
-    copied_parameters.backend_params[BackendParameters.COMPRESS_WEIGHTS] = False
+    copied_parameters.compress_weights = False
 
     quantized_model = quantize_impl(
         model=model,

src/nncf/openvino/quantization/backend_parameters.py

@@ -9,29 +9,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Optional
-
-from nncf.quantization.advanced_parameters import AdvancedQuantizationParameters
-
 
 class BackendParameters:
-    COMPRESS_WEIGHTS = "compress_weights"
     STAT_REQUESTS_NUMBER = "stat_requests_number"
     EVAL_REQUESTS_NUMBER = "eval_requests_number"
     ACTIVATIONS = "activations"
     WEIGHTS = "weights"
     LEVEL_LOW = "level_low"
     LEVEL_HIGH = "level_high"
-
-
-def is_weight_compression_needed(advanced_parameters: Optional[AdvancedQuantizationParameters]) -> bool:
-    """
-    Determines whether weight compression is needed based on the provided
-    advanced quantization parameters.
-
-    :param advanced_parameters: Advanced quantization parameters.
-    :return: True if weight compression is needed, False otherwise.
-    """
-    if advanced_parameters is not None and advanced_parameters.backend_params is not None:
-        return advanced_parameters.backend_params.get(BackendParameters.COMPRESS_WEIGHTS, True)
-    return True

src/nncf/openvino/quantization/quantize_model.py

@@ -27,8 +27,6 @@
 from nncf.openvino.graph.model_utils import remove_friendly_name_duplicates
 from nncf.openvino.graph.nncf_graph_builder import GraphConverter
 from nncf.openvino.graph.node_utils import get_number_if_op
-from nncf.openvino.quantization.backend_parameters import BackendParameters
-from nncf.openvino.quantization.backend_parameters import is_weight_compression_needed
 from nncf.openvino.quantization.quantize_ifmodel import apply_algorithm_if_bodies
 from nncf.openvino.rt_info import dump_parameters
 from nncf.parameters import BackupMode
@@ -43,6 +41,7 @@
 from nncf.quantization.advanced_parameters import AdvancedCompressionParameters
 from nncf.quantization.advanced_parameters import AdvancedQuantizationParameters
 from nncf.quantization.advanced_parameters import convert_to_dict_recursively
+from nncf.quantization.advanced_parameters import is_weight_compression_needed
 from nncf.quantization.algorithms.accuracy_control.algorithm import QuantizationAccuracyRestorer
 from nncf.quantization.algorithms.accuracy_control.algorithm import calculate_accuracy_drop
 from nncf.quantization.algorithms.accuracy_control.evaluator import Evaluator
@@ -211,13 +210,13 @@ def quantize_with_accuracy_control_impl(
     if advanced_accuracy_restorer_parameters is None:
         advanced_accuracy_restorer_parameters = AdvancedAccuracyRestorerParameters()
 
-    compress_weights = is_weight_compression_needed(advanced_quantization_parameters)
-
     if advanced_quantization_parameters is None:
         copied_parameters = AdvancedQuantizationParameters()
     else:
         copied_parameters = deepcopy(advanced_quantization_parameters)
-    copied_parameters.backend_params[BackendParameters.COMPRESS_WEIGHTS] = False
+
+    compress_weights = is_weight_compression_needed(copied_parameters)
+    copied_parameters.compress_weights = False
 
     quantized_model = quantize_impl(
         model=model,

src/nncf/quantization/advanced_parameters.py

@@ -256,6 +256,8 @@ class AdvancedQuantizationParameters:
     :type smooth_quant_alpha: float
     :param backend_params: Backend-specific parameters.
     :type backend_params: dict[str, Any]
+    :param compress_weights: Indicates whether to apply weight compression after quantization.
+    :type compress_weights: bool
     """
 
     # General parameters
@@ -270,6 +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)
@@ -287,6 +290,20 @@ class AdvancedQuantizationParameters:
     backend_params: dict[str, Any] = field(default_factory=dict)
 
 
+def is_weight_compression_needed(advanced_parameters: Optional[AdvancedQuantizationParameters]) -> bool:
+    """
+    Determine whether weight compression is needed based on advanced quantization parameters.
+
+    If `advanced_parameters` or its `backend_params` are not provided, defaults to True.
+
+    :param advanced_parameters: Advanced quantization parameters.
+    :return: True if weight compression is needed, False otherwise.
+    """
+    if advanced_parameters is not None:
+        return advanced_parameters.compress_weights
+    return True
+
+
 @api()
 @dataclass
 class AdvancedAWQParameters:

tests/onnx/test_passes.py

@@ -14,7 +14,6 @@
 import nncf
 from nncf.onnx.graph.passes import apply_preprocess_passes
 from nncf.onnx.graph.passes import compress_quantize_weights_transformation
-from nncf.onnx.quantization.backend_parameters import BackendParameters
 from tests.onnx.common import ModelBuilder
 from tests.onnx.models import build_matmul_model_with_nop_cast
 
@@ -60,7 +59,7 @@ def test_compress_quantize_weights_transformation():
         model,
         calibration_dataset,
         advanced_parameters=nncf.AdvancedQuantizationParameters(
-            backend_params={BackendParameters.COMPRESS_WEIGHTS: False}
+            compress_weights=False,
         ),
     )
 

tests/torch2/fx/test_models.py

@@ -33,7 +33,6 @@
 from nncf.common.utils.os import safe_open
 from nncf.experimental.torch.fx.nncf_graph_builder import GraphConverter
 from nncf.experimental.torch.fx.node_utils import get_tensor_constant_from_node
-from nncf.experimental.torch.fx.quantization.backend_parameters import FXBackendParameters
 from nncf.experimental.torch.fx.transformations import DEQUANTIZE_NODE_TARGETS
 from nncf.experimental.torch.fx.transformations import _get_node_inputs
 from nncf.quantization.advanced_parameters import AdvancedQuantizationParameters
@@ -219,7 +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={FXBackendParameters.COMPRESS_WEIGHTS: compress_weights}
+        disable_bias_correction=True,
+        compress_weights=compress_weights,
     )
     quantization_parameters["subset_size"] = 1