Fix tests

nncf/openvino/graph/layer_attributes.py

@@ -71,7 +71,7 @@ def get_backend_agnostic_attributes(self):
 
 
 def get_weighted_layer_attributes(
-    ov_node: ov.Node, ov_metatype: OVOpMetatype, constant_attributes: Dict[str, Any]
+    ov_node: ov.Node, ov_metatype: OVOpMetatype, constant_attributes: Dict[int, Any]
 ) -> WeightedLayerAttributes:
     """
     Funciton retrieves common layer attributes from the given node.

nncf/openvino/quantization/weights_compression.py

@@ -18,6 +18,7 @@
 from nncf.common.graph.layer_attributes import LayoutElem
 from nncf.common.graph.layer_attributes import LinearLayerAttributes
 from nncf.common.graph.operator_metatypes import OperatorMetatype
+from nncf.openvino.graph.layer_attributes import get_weighted_layer_attributes
 from nncf.openvino.graph.metatypes.openvino_metatypes import OVEmbeddingMetatype
 from nncf.openvino.graph.metatypes.openvino_metatypes import OVMatMulMetatype
 from nncf.openvino.graph.metatypes.openvino_metatypes import get_node_metatype
@@ -57,7 +58,8 @@ def insert_pre_compression_operations(model: ov.Model, bits: int = 8) -> None:
                 continue
 
             weight = get_const_value(weight_node)
-            axes = _get_reduction_axes(metatype, node, const_port_id)
+            weight_shape = weight_node.get_output_shape(0)
+            axes = _get_reduction_axes(metatype, node, const_port_id, weight_shape)
             min_values = np.min(weight, axis=axes, keepdims=True)
             max_values = np.max(weight, axis=axes, keepdims=True)
 
@@ -78,18 +80,25 @@ def insert_pre_compression_operations(model: ov.Model, bits: int = 8) -> None:
                 target_input.replace_source_output(mul.output(0))
 
 
-def _get_reduction_axes(metatype: Type[OperatorMetatype], node: ov.Node, weight_port_id: int) -> Union[int, Tuple[int]]:
+def _get_reduction_axes(
+    metatype: Type[OperatorMetatype], node: ov.Node, weight_port_id: int, weight_shape: ov.Shape
+) -> Union[int, Tuple[int]]:
     """
     Determines reduction axes by given metatype and node information.
 
     :param metatype: The metatype of the operator.
     :param node: The OpenVINO node.
     :param weight_port_id: The weight port ID.
+    :param weight_shape: Shape of the target node weight.
 
     :return: The reduction axes as an integer or a tuple of integers.
     """
     if metatype is OVMatMulMetatype:
-        layer_attributes = node.layer_attributes.get_backend_agnostic_attributes()
+        transpose = node.get_attributes()[f"transpose_{'a' if weight_port_id == 0 else 'b'}"]
+        constant_attributes = {weight_port_id: {"shape": weight_shape, "transpose": transpose}}
+        layer_attributes = get_weighted_layer_attributes(
+            ov_node=node, ov_metatype=OVMatMulMetatype, constant_attributes=constant_attributes
+        )
         assert isinstance(layer_attributes, LinearLayerAttributes)
         axes = tuple(idx for idx, elem in enumerate(layer_attributes.weights_layout) if elem == LayoutElem.C_IN)
     elif metatype is OVEmbeddingMetatype:

nncf/quantization/algorithms/channel_alignment/algorithm.py

@@ -29,7 +29,6 @@
 from nncf.common.tensor_statistics.statistic_point import StatisticPointsContainer
 from nncf.common.utils.backend import BackendType
 from nncf.common.utils.backend import get_backend
-from nncf.openvino.graph.node_utils import get_channel_agnostic_reduction_shape
 from nncf.quantization.algorithms.algorithm import Algorithm
 from nncf.quantization.algorithms.channel_alignment.backend import ALGO_BACKENDS
 from nncf.quantization.algorithms.channel_alignment.backend import ChannelAlignmentAlgoBackend
@@ -392,7 +391,9 @@ def get_statistic_points(self, model: TModel, graph: NNCFGraph) -> StatisticPoin
 
             channel_axis = conv_in.metatype.output_channel_axis
             activation_shape = list(range(len(graph.get_output_edges(node_in)[0].tensor_shape)))
-            reduction_shape = get_channel_agnostic_reduction_shape([channel_axis], activation_shape)
+            reduction_shape = self._backend_entity.get_channel_agnostic_reduction_shape(
+                [channel_axis], activation_shape
+            )
 
             statistic_collector = self._backend_entity.get_statistic_collector(
                 tuple(reduction_shape), self._quantile, self.subset_size, self.inplace_statistics

nncf/quantization/algorithms/channel_alignment/backend.py

@@ -133,3 +133,14 @@ def create_bias_tensor(node: NNCFNode, nncf_graph: NNCFGraph, value: Any) -> np.
         :param value: Value to fill bias constant array.
         :return: Bias value constant array filled by given value.
         """
+
+    @staticmethod
+    @abstractmethod
+    def get_channel_agnostic_reduction_shape(channel_axis: int, shape: Tuple[int]) -> Tuple[int]:
+        """
+        Returns filtered reduction shape without axes that corresponds channels.
+
+        :param channel_axes: List of the channel axes.
+        :param shape: Shape that need to be filtered.
+        :return: Reduction shape in tuple format.
+        """

nncf/quantization/algorithms/channel_alignment/openvino_backend.py

@@ -29,6 +29,7 @@
 from nncf.openvino.graph.metatypes.openvino_metatypes import OVSubtractMetatype
 from nncf.openvino.graph.node_utils import create_bias_tensor
 from nncf.openvino.graph.node_utils import get_bias_value
+from nncf.openvino.graph.node_utils import get_channel_agnostic_reduction_shape
 from nncf.openvino.graph.node_utils import get_node_with_bias_value
 from nncf.openvino.graph.node_utils import get_weight_value
 from nncf.openvino.graph.transformations.commands import OVTargetPoint
@@ -101,3 +102,7 @@ def is_node_with_bias(node: NNCFNode, nncf_graph: NNCFGraph) -> bool:
     @staticmethod
     def create_bias_tensor(node: NNCFNode, nncf_graph: NNCFGraph, value: Any) -> np.ndarray:
         return create_bias_tensor(node, nncf_graph, value)
+
+    @staticmethod
+    def get_channel_agnostic_reduction_shape(channel_axis: int, shape: Tuple[int]) -> Tuple[int]:
+        return get_channel_agnostic_reduction_shape(channel_axes=channel_axis, shape=shape)