[PTQ][Torch][KQV self attention] Align FQ placement between OV and Torch backend

nncf/common/graph/graph.py

@@ -594,12 +594,22 @@ def get_graph_for_structure_analysis(self, extended: bool = False) -> nx.DiGraph
             attrs_edge = {}
             u = u.replace(__RESERVED_DOT_CHARACTER, __CHARACTER_REPLACE_TO)
             v = v.replace(__RESERVED_DOT_CHARACTER, __CHARACTER_REPLACE_TO)
+            label = {}
+            if edge[NNCFGraph.PARALLEL_INPUT_PORT_IDS_ATTR]:
+                label["parallel_input_port_ids"] = edge[NNCFGraph.PARALLEL_INPUT_PORT_IDS_ATTR]
+
             if extended:
                 if edge[NNCFGraph.DTYPE_EDGE_ATTR] is Dtype.INTEGER:
                     attrs_edge["style"] = "dashed"
                 else:
                     attrs_edge["style"] = "solid"
-                attrs_edge["label"] = edge[NNCFGraph.ACTIVATION_SHAPE_EDGE_ATTR]
+                label["shape"] = edge[NNCFGraph.ACTIVATION_SHAPE_EDGE_ATTR]
+
+            if label:
+                if len(label) == 1 and extended:
+                    attrs_edge["label"] = label.popitem()[1]
+                else:
+                    attrs_edge["label"] = ", ".join((f"{k}:{v}" for k, v in label.items()))
             out_graph.add_edge(u, v, **attrs_edge)
         return out_graph
 

nncf/quantization/algorithms/accuracy_control/ranker.py

@@ -28,7 +28,7 @@
 from nncf.quantization.algorithms.accuracy_control.evaluator import Evaluator
 from nncf.quantization.algorithms.accuracy_control.rank_functions import create_normalized_mse_func
 from nncf.quantization.algorithms.accuracy_control.subset_selection import select_subset
-from nncf.quantization.passes import remove_shapeof_subgraphs
+from nncf.quantization.passes import remove_shapeof_subgraphs_inplace
 
 TModel = TypeVar("TModel")
 TPModel = TypeVar("TPModel")
@@ -98,8 +98,9 @@ def find_groups_of_quantizers_to_rank(self, quantized_model_graph: NNCFGraph) ->
             if x.metatype in self._algo_backend.get_quantizer_metatypes()
         ]
 
-        quantized_model_graph_without_shapeof = remove_shapeof_subgraphs(
-            deepcopy(quantized_model_graph), self._algo_backend.get_shapeof_metatypes()
+        quantized_model_graph_without_shapeof = deepcopy(quantized_model_graph)
+        remove_shapeof_subgraphs_inplace(
+            quantized_model_graph_without_shapeof, self._algo_backend.get_shapeof_metatypes()
         )
 
         for quantizer_node in reversed(quantizers):

nncf/quantization/algorithms/min_max/algorithm.py

@@ -54,7 +54,6 @@
 from nncf.quantization.algorithms.min_max.backend import ALGO_BACKENDS
 from nncf.quantization.fake_quantize import calculate_quantizer_parameters
 from nncf.quantization.fake_quantize import get_quantizer_narrow_range
-from nncf.quantization.passes import transform_to_inference_graph
 from nncf.quantization.range_estimator import RangeEstimatorParameters
 from nncf.quantization.range_estimator import RangeEstimatorParametersSet
 from nncf.scopes import IgnoredScope
@@ -504,9 +503,7 @@ def _get_quantization_target_points(
         )
         hw_patterns = PatternsManager.get_full_hw_pattern_graph(backend=backend, device=device, model_type=model_type)
 
-        inference_nncf_graph = transform_to_inference_graph(
-            deepcopy(nncf_graph), self._backend_entity.shapeof_metatypes, self._backend_entity.read_variable_metatypes
-        )
+        inference_nncf_graph = self._backend_entity.transform_to_inference_graph(nncf_graph)
 
         quantizer_setup = self._get_quantizer_setup(nncf_graph, inference_nncf_graph, hw_patterns, ignored_patterns)
         self._apply_model_type_pass(self._model_type, quantizer_setup, nncf_graph)

nncf/quantization/algorithms/min_max/backend.py

@@ -49,13 +49,6 @@ def post_processing_metatypes(self) -> List[OperatorMetatype]:
         Property for the backend-specific post-processing metatypes (NonMaximumSupression, TopK, etc.).
         """
 
-    @property
-    @abstractmethod
-    def shapeof_metatypes(self) -> List[OperatorMetatype]:
-        """
-        Property for the backend-specific ShapeOf metatypes.
-        """
-
     @property
     @abstractmethod
     def conv_metatypes(self) -> List[OperatorMetatype]:
@@ -70,13 +63,6 @@ def overflow_fix_metatypes(self) -> List[OperatorMetatype]:
         Property for the backend-specific metatypes for which overflow_fix is applicable.
         """
 
-    @property
-    @abstractmethod
-    def read_variable_metatypes(self) -> List[OperatorMetatype]:
-        """
-        Property for the backend-specific metatypes that also can be interpreted as inputs (ReadValue).
-        """
-
     @property
     @abstractmethod
     def add_metatypes(self) -> List[OperatorMetatype]:
@@ -174,6 +160,13 @@ def get_statistic_collector(
         :return: Backend-specific TensorStatisticCollectorBase for the statistics calculation.
         """
 
+    @staticmethod
+    @abstractmethod
+    def transform_to_inference_graph(graph: NNCFGraph) -> NNCFGraph:
+        """
+        Returns inference NNCFGraph without constant flows and training time operations.
+        """
+
     @staticmethod
     @abstractmethod
     def get_weight_tensor_port_ids(node: NNCFNode) -> List[Optional[int]]:

nncf/quantization/algorithms/min_max/onnx_backend.py

@@ -9,6 +9,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from copy import deepcopy
 from typing import Dict, List, Optional, Set, Union
 
 import numpy as np
@@ -42,6 +43,8 @@
 from nncf.quantization.algorithms.min_max.backend import ALGO_BACKENDS
 from nncf.quantization.algorithms.min_max.backend import MinMaxAlgoBackend
 from nncf.quantization.fake_quantize import FakeQuantizeParameters
+from nncf.quantization.passes import filter_constant_nodes_inplace
+from nncf.quantization.passes import remove_shapeof_subgraphs_inplace
 from nncf.quantization.range_estimator import RangeEstimatorParameters
 
 
@@ -56,10 +59,6 @@ def mat_mul_metatypes(self) -> List[OperatorMetatype]:
     def post_processing_metatypes(self) -> List[OperatorMetatype]:
         return [om.ONNXTopKMetatype, om.ONNXNonMaxSuppressionMetatype]
 
-    @property
-    def shapeof_metatypes(self) -> List[OperatorMetatype]:
-        return [om.ONNXShapeMetatype]
-
     @property
     def conv_metatypes(self) -> List[OperatorMetatype]:
         return [om.ONNXConvolutionMetatype]
@@ -68,10 +67,6 @@ def conv_metatypes(self) -> List[OperatorMetatype]:
     def overflow_fix_metatypes(self) -> List[OperatorMetatype]:
         return [om.ONNXConvolutionMetatype, om.ONNXConvolutionTransposeMetatype, *MATMUL_METATYPES]
 
-    @property
-    def read_variable_metatypes(self) -> List[OperatorMetatype]:
-        return []
-
     @property
     def add_metatypes(self) -> List[OperatorMetatype]:
         return [om.ONNXAddLayerMetatype]
@@ -170,6 +165,15 @@ def get_statistic_collector(
             f"{str(range_estimator_params)}"
         )
 
+    @staticmethod
+    def transform_to_inference_graph(graph: NNCFGraph) -> NNCFGraph:
+        inference_graph = deepcopy(graph)
+        remove_shapeof_subgraphs_inplace(
+            nncf_graph=inference_graph, shapeof_metatypes=[om.ONNXShapeMetatype], read_variable_metatypes=[]
+        )
+        filter_constant_nodes_inplace(nncf_graph=inference_graph)
+        return inference_graph
+
     @staticmethod
     def get_weight_tensor_port_ids(node: NNCFNode) -> List[Optional[int]]:
         return list(node.layer_attributes.weight_attrs.keys())

nncf/quantization/algorithms/min_max/openvino_backend.py

@@ -9,6 +9,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from copy import deepcopy
 from typing import Dict, List, Optional, Set, Tuple
 
 import numpy as np
@@ -43,6 +44,8 @@
 from nncf.quantization.algorithms.min_max.backend import ALGO_BACKENDS
 from nncf.quantization.algorithms.min_max.backend import MinMaxAlgoBackend
 from nncf.quantization.fake_quantize import FakeQuantizeParameters
+from nncf.quantization.passes import filter_constant_nodes_inplace
+from nncf.quantization.passes import remove_shapeof_subgraphs_inplace
 
 
 # pylint:disable=too-many-public-methods
@@ -56,10 +59,6 @@ def mat_mul_metatypes(self) -> List[OperatorMetatype]:
     def post_processing_metatypes(self) -> List[OperatorMetatype]:
         return [om.OVTopKMetatype, om.OVNonMaxSuppressionMetatype]
 
-    @property
-    def shapeof_metatypes(self) -> List[OperatorMetatype]:
-        return [om.OVShapeOfMetatype]
-
     @property
     def conv_metatypes(self) -> List[OperatorMetatype]:
         return [om.OVConvolutionMetatype]
@@ -74,10 +73,6 @@ def overflow_fix_metatypes(self) -> List[OperatorMetatype]:
             om.OVMatMulMetatype,
         ]
 
-    @property
-    def read_variable_metatypes(self) -> List[OperatorMetatype]:
-        return [om.OVReadValueMetatype]
-
     @property
     def add_metatypes(self) -> List[OperatorMetatype]:
         return [om.OVAddMetatype]
@@ -200,6 +195,18 @@ def get_statistic_collector(
             collector.register_statistic_branch(container_key, reducer, aggregator)
         return collector
 
+    @staticmethod
+    def transform_to_inference_graph(graph: NNCFGraph) -> NNCFGraph:
+        inference_graph = deepcopy(graph)
+        read_variable_metatypes = [om.OVReadValueMetatype]
+        remove_shapeof_subgraphs_inplace(
+            nncf_graph=inference_graph,
+            shapeof_metatypes=[om.OVShapeOfMetatype],
+            read_variable_metatypes=read_variable_metatypes,
+        )
+        filter_constant_nodes_inplace(nncf_graph=inference_graph, read_variable_metatypes=read_variable_metatypes)
+        return inference_graph
+
     @staticmethod
     def get_weight_tensor_port_ids(node: NNCFNode) -> List[Optional[int]]:
         return node.layer_attributes.get_const_port_ids()

nncf/quantization/algorithms/min_max/torch_backend.py

@@ -9,6 +9,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from copy import deepcopy
 from typing import Dict, List, Optional, Set, Tuple
 
 import torch
@@ -32,6 +33,8 @@
 from nncf.quantization.algorithms.min_max.backend import ALGO_BACKENDS
 from nncf.quantization.algorithms.min_max.backend import MinMaxAlgoBackend
 from nncf.quantization.fake_quantize import FakeQuantizeParameters
+from nncf.quantization.passes import filter_constant_nodes_inplace
+from nncf.quantization.passes import remove_dropout_nodes_inplace
 from nncf.quantization.range_estimator import RangeEstimatorParameters
 from nncf.torch.graph.graph import PTTargetPoint
 from nncf.torch.graph.transformations.commands import PTQuantizerInsertionCommand
@@ -65,10 +68,6 @@ def mat_mul_metatypes(self) -> List[OperatorMetatype]:
     def post_processing_metatypes(self) -> List[OperatorMetatype]:
         return []
 
-    @property
-    def shapeof_metatypes(self) -> List[OperatorMetatype]:
-        return []
-
     @property
     def conv_metatypes(self) -> List[OperatorMetatype]:
         return [om.PTModuleConv1dMetatype, om.PTModuleConv2dMetatype, om.PTModuleConv3dMetatype]
@@ -85,10 +84,6 @@ def overflow_fix_metatypes(self) -> List[OperatorMetatype]:
             om.PTModuleConvTranspose3dMetatype,
         ]
 
-    @property
-    def read_variable_metatypes(self) -> List[OperatorMetatype]:
-        return []
-
     @property
     def add_metatypes(self) -> List[OperatorMetatype]:
         return [om.PTAddMetatype]
@@ -188,6 +183,13 @@ def get_statistic_collector(
             collector.register_statistic_branch(container_key, reducer, aggregator)
         return collector
 
+    @staticmethod
+    def transform_to_inference_graph(graph: NNCFGraph) -> NNCFGraph:
+        inference_graph = deepcopy(graph)
+        remove_dropout_nodes_inplace(nncf_graph=inference_graph, dropout_metatypes=[om.PTDropoutMetatype])
+        filter_constant_nodes_inplace(nncf_graph=inference_graph)
+        return inference_graph
+
     @staticmethod
     def get_weight_tensor_port_ids(node: NNCFNode) -> List[Optional[int]]:
         return [None]
@@ -307,6 +309,10 @@ def get_ignored_metatypes(model_type: ModelType, device: TargetDevice) -> List[O
                 om.PTDivMetatype,
                 om.PTMaxMetatype,
                 om.PTSqueezeMetatype,
+                om.PTLayerNormMetatype,
+                om.PTModuleLayerNormMetatype,
+                om.PTGroupNormMetatype,
+                om.PTModuleGroupNormMetatype,
             ]
             if device != TargetDevice.CPU_SPR:
                 types.append(om.PTMulMetatype)