Comments

nncf/quantization/algorithms/smooth_quant/algorithm.py

@@ -243,7 +243,7 @@ def _get_statistics_for_node(
         statistics_for_node = []
         for tensor_collector in statistic_points.get_algo_statistics_for_node(
             node_name,
-            self._backend_entity.get_filter_fn_for_statistics(act_port),
+            self._backend_entity.get_filter_fn_for_statistics(act_port, self._algorithm_key),
             self._algorithm_key,
         ):
             statistic = tensor_collector.get_statistics()[STATISTIC_BRANCH_KEY]

nncf/quantization/algorithms/smooth_quant/backend.py

@@ -207,9 +207,11 @@ def is_node_with_shared_weight(node: NNCFNode, nncf_graph: NNCFGraph) -> bool:
 
     @staticmethod
     @abstractmethod
-    def get_filter_fn_for_statistics(activation_port_id: int) -> Callable[[StatisticPoint], bool]:
+    def get_filter_fn_for_statistics(activation_port_id: int, algorithm_key: str) -> Callable[[StatisticPoint], bool]:
         """
         Returns backend-specific callable to filter statistic containers according to its statistic point.
 
         :param activation_port_id: Activation port id for the statistic collection target node.
+        :param algorithm_key: Current algorithm key.
+        :return: Backend-specific callable to filter statistic containers according to its statistic point.
         """

nncf/quantization/algorithms/smooth_quant/openvino_backend.py

@@ -38,6 +38,8 @@
 from nncf.openvino.statistics.collectors import OVNNCFCollectorTensorProcessor
 from nncf.quantization.algorithms.smooth_quant.backend import SmoothQuantAlgoBackend
 
+OV_PRE_LAYER_TARGET_TYPE = TargetType.PRE_LAYER_OPERATION
+
 
 class OVSmoothQuantAlgoBackend(SmoothQuantAlgoBackend):
     @property
@@ -54,7 +56,7 @@ def quantize_agnostic_metatypes(self) -> List[OperatorMetatype]:
 
     @staticmethod
     def pre_layer_target_type() -> TargetType:
-        return TargetType.PRE_LAYER_OPERATION
+        return OV_PRE_LAYER_TARGET_TYPE
 
     @staticmethod
     def target_point(target_type: TargetType, target_node_name: str, port_id: int) -> OVTargetPoint:
@@ -155,8 +157,12 @@ def is_node_with_shared_weight(node: NNCFNode, nncf_graph: NNCFGraph) -> bool:
         return len(nncf_graph.get_next_nodes(weight_node)) > 1
 
     @staticmethod
-    def get_filter_fn_for_statistics(activation_port_id: int) -> Callable[[StatisticPoint], bool]:
+    def get_filter_fn_for_statistics(activation_port_id: int, algorithm_key: str) -> Callable[[StatisticPoint], bool]:
         def filter_func(point: StatisticPoint) -> bool:
-            return point.target_point.port_id == activation_port_id
+            return (
+                algorithm_key in point.algorithm_to_tensor_collectors
+                and point.target_point.type == OV_PRE_LAYER_TARGET_TYPE
+                and point.target_point.port_id == activation_port_id
+            )
 
         return filter_func

nncf/quantization/algorithms/smooth_quant/torch_backend.py

@@ -46,6 +46,9 @@ def forward(self, x):
         return torch.mul(x, self._scale_value)
 
 
+PT_PRE_LAYER_TARGET_TYPE = TargetType.OPERATOR_PRE_HOOK
+
+
 class PTSmoothQuantAlgoBackend(SmoothQuantAlgoBackend):
     @property
     def convolution_metatypes(self) -> List[OperatorMetatype]:
@@ -65,22 +68,22 @@ def quantize_agnostic_metatypes(self) -> List[OperatorMetatype]:
 
     @staticmethod
     def pre_layer_target_type() -> TargetType:
-        return TargetType.OPERATOR_PRE_HOOK
+        return PT_PRE_LAYER_TARGET_TYPE
 
     @staticmethod
     def target_point(target_type: TargetType, target_node_name: str, port_id: int) -> PTTargetPoint:
         return PTTargetPoint(target_type, target_node_name, input_port_id=port_id)
 
     @staticmethod
     def is_node_with_weights(node: NNCFNode) -> bool:
-        # Metatypes of matmuls and convolutions guarantee
+        # Metatypes of linears and convolutions guarantee
         # all nodes with the metatypes have weights, we can skip
         # this check by returning True.
         return True
 
     @staticmethod
     def get_activations_port_id(node: NNCFNode, nncf_graph: NNCFGraph) -> int:
-        # Metatypes of matmuls and convolutions guarantee
+        # Metatypes of linears and convolutions guarantee
         # all nodes with the metatypes have 0 activation port id
         return 0
 
@@ -127,9 +130,7 @@ def scale_insertion_command(
         input_port_id = 0
         target_points = []
         for node in nodes:
-            target_points.append(
-                PTTargetPoint(TargetType.OPERATOR_PRE_HOOK, node.node_name, input_port_id=input_port_id)
-            )
+            target_points.append(PTTargetPoint(PT_PRE_LAYER_TARGET_TYPE, node.node_name, input_port_id=input_port_id))
 
         return PTSharedFnInsertionCommand(target_points, SQMultiply(scale_value), scale_node_name)
 
@@ -150,8 +151,12 @@ def is_node_with_shared_weight(node: NNCFNode, nncf_graph: NNCFGraph) -> bool:
         return node.is_shared()
 
     @staticmethod
-    def get_filter_fn_for_statistics(activation_port_id: int) -> Callable[[StatisticPoint], bool]:
+    def get_filter_fn_for_statistics(activation_port_id: int, algorithm_key: str) -> Callable[[StatisticPoint], bool]:
         def filter_func(point: StatisticPoint) -> bool:
-            return point.target_point.input_port_id == activation_port_id
+            return (
+                algorithm_key in point.algorithm_to_tensor_collectors
+                and point.target_point.type == PT_PRE_LAYER_TARGET_TYPE
+                and point.target_point.input_port_id == activation_port_id
+            )
 
         return filter_func