Move is_weight_compression_needed function to common

Author: aobolensk

src/nncf/experimental/torch/fx/quantization/backend_parameters.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
@@ -93,7 +92,7 @@ def quantize_impl(
     nncf_graph = NNCFGraphFactory.create(copied_model)
     quantized_model = quantization_algorithm.apply(copied_model, nncf_graph, dataset=calibration_dataset)
 
-    if is_weight_compression_needed(advanced_parameters):
+    if advanced_parameters is not None and advanced_parameters.is_weight_compression_needed():
         compress_post_quantize_transformation(quantized_model)
     else:
         fq_weights_transformation(quantized_model)

src/nncf/experimental/torch/fx/quantization/quantize_model.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/backend_parameters.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
@@ -123,7 +121,7 @@ def _extract_all_subgraphs(model: ov.Model, current_id: str) -> None:
         quantization_algorithm, model, graphs, main_model_graph_id, calibration_dataset, subset_size, 1
     )
 
-    if is_weight_compression_needed(advanced_parameters):
+    if advanced_parameters is None or advanced_parameters.is_weight_compression_needed():
         compress_quantize_weights_transformation(quantized_model)
 
     dump_parameters(
@@ -170,7 +168,7 @@ def native_quantize_impl(
     warning_model_no_batchwise_support(graph, advanced_parameters, model_type, OPERATIONS_OUTPUT_HAS_NO_BATCH_AXIS)
     quantized_model = quantization_algorithm.apply(model, graph, dataset=calibration_dataset)
 
-    if is_weight_compression_needed(advanced_parameters):
+    if advanced_parameters is None or advanced_parameters.is_weight_compression_needed():
         compress_quantize_weights_transformation(quantized_model)
 
     dump_parameters(
@@ -211,13 +209,13 @@ 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 = copied_parameters.is_weight_compression_needed()
+    copied_parameters.backend_params[AdvancedQuantizationParameters.COMPRESS_WEIGHTS] = False
 
     quantized_model = quantize_impl(
         model=model,

src/nncf/openvino/quantization/quantize_model.py

@@ -286,6 +286,22 @@ 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(self) -> 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 self.backend_params is not None:
+            return bool(self.backend_params.get(AdvancedQuantizationParameters.COMPRESS_WEIGHTS, True))
+        return True
+
 
 @api()
 @dataclass

src/nncf/quantization/advanced_parameters.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 +218,7 @@ 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, backend_params={AdvancedQuantizationParameters.COMPRESS_WEIGHTS: compress_weights}
     )
     quantization_parameters["subset_size"] = 1