Dynamic OV model builder (#3137)

### Changes

- Added `ModelBuilder` class.
- Updated FastBC algorithm to utilize new approach.

### Reason for changes

- Algorithm speed up.

### Related tickets

- 122317

### Tests

- Added tests/openvino/native/test_model_builder.py.
- Conversion jobs were run using DLB scope.

Results (develop run - manual/post_training_quantization_performance/91,
92; branch - manual/post_training_quantization_performance/90)

Model | Backend | FBC time (develop, OV) | FBC time (develop, PT) | FBC
time (branch, OV) | Diff (develop - branch, OV)
-- | -- | -- | -- | -- | --
hf/bert-base-uncased | OV | 00:00:02 | - | 00:00:01 | **00:00:01**
torchvision/resnet18 | OV | 00:00:00 | 00:00:00 | 00:00:00 | 00:00:00
torchvision/mobilenet_v3_small_BC | OV | 00:00:01 | - | 00:00:01 |
00:00:00
torchvision/vit_b_16 | OV | 00:00:02 | - | 00:00:01 | **00:00:01**
torchvision/swin_v2_s | OV | 00:00:05 | - | 00:00:01 | **00:00:04**
timm/crossvit_9_240 | OV | 00:00:01 | 00:00:00 | **00:00:01** | 00:00:00
timm/darknet53 | OV | 00:00:01 | 00:00:00 | **00:00:01** | 00:00:00
timm/deit3_small_patch16_224 | OV | 00:00:01 | 00:00:00 | 00:00:00 |
**00:00:01**
timm/dla34 | OV | 00:00:00 | 00:00:00 | 00:00:00 | 00:00:00
timm/dpn68 | OV | 00:00:00 | 00:00:00 | 00:00:00 | 00:00:00
timm/efficientnet_b0 | OV | 00:00:00 | 00:00:00 | 00:00:00 | 00:00:00
timm/efficientnet_b0_BC | OV | 00:00:04 | - | 00:00:04 | 00:00:00
timm/efficientnet_lite0 | OV | 00:00:00 | 00:00:00 | 00:00:00 | 00:00:00
timm/hrnet_w18 | OV | 00:00:13 | 00:00:03 | **00:00:04** | **00:00:09**
timm/inception_resnet_v2 | OV | 00:00:08 | 00:00:03 | **00:00:04** |
**00:00:04**
timm/levit_128 | OV | 00:00:00 | 00:00:00 | 00:00:00 | 00:00:00
timm/mobilenetv2_050 | OV | 00:00:00 | 00:00:00 | 00:00:00 | 00:00:00
timm/mobilenetv2_050_BC | OV | 00:00:03 | - | 00:00:03 | 00:00:00
timm/mobilenetv3_small_050 | OV | 00:00:00 | 00:00:00 | 00:00:00 |
00:00:00
timm/mobilenetv3_small_050_BC | OV | 00:00:01 | - | 00:00:01 | 00:00:00
timm/regnetx_002 | OV | 00:00:00 | 00:00:00 | 00:00:00 | 00:00:00
timm/resnest14d | OV | 00:00:00 | 00:00:00 | 00:00:00 | 00:00:00
timm/swin_base_patch4_window7_224 | OV | 00:00:03 | 00:00:00 |
**00:00:02** | **00:00:01**
timm/tf_inception_v3 | OV | 00:00:01 | 00:00:01 | 00:00:01 | 00:00:00
timm/vgg11 | OV | 00:00:01 | 00:00:00 | **00:00:01** | 00:00:00
timm/visformer_small | OV | 00:00:00 | 00:00:00 | 00:00:00 | 00:00:00
timm/wide_resnet50_2 | OV | 00:00:01 | 00:00:00 | **00:00:01** |
00:00:00

Author: nikita-malininn

nncf/openvino/graph/model_builder.py

@@ -0,0 +1,226 @@
+# Copyright (c) 2025 Intel Corporation
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#      http://www.apache.org/licenses/LICENSE-2.0
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from collections import deque
+from typing import Dict, List, Tuple
+
+import openvino.runtime as ov
+from openvino.runtime import opset13 as opset
+from openvino.runtime.utils.node_factory import NodeFactory
+
+from nncf.openvino.graph.model_transformer import OVModelTransformer
+from nncf.openvino.graph.node_utils import get_parameter_node_name
+from nncf.openvino.graph.node_utils import get_result_node_name
+
+
+class OVModelBuilder:
+    """
+    The purpose of the ModelBuilder is to build a new OpenVINO model from input and output points.
+    This Builder was created to reduce the number of model cloning that is required for ModelTransformer to work.
+    """
+
+    def __init__(self):
+        self._node_factory = NodeFactory()
+
+    @staticmethod
+    def _create_parameter(node_name: str, node_input: ov.Input) -> ov.Node:
+        """
+        A method that contains steps to create a Parameter for a new model using a specific template.
+        """
+        port_id = node_input.get_index()
+        parameter_name = get_parameter_node_name(node_name, port_id)
+        return opset.parameter(
+            shape=node_input.get_partial_shape(),
+            dtype=node_input.get_element_type(),
+            name=parameter_name,
+        )
+
+    @staticmethod
+    def _create_result(node_name: str, node_output: ov.Input) -> ov.Node:
+        """
+        A method that contains steps to create a Result for a new model using a specific template.
+        """
+        port_id = node_output.get_index()
+        result_name = get_result_node_name(node_name, port_id=port_id)
+        result = opset.result(node_output, name=result_name)
+        result.get_output_tensor(0).set_names({result_name})
+        return result
+
+    def _collect_graph_nodes(
+        self,
+        input_ids: List[Tuple[str, int]],
+        output_ids: List[Tuple[str, int]],
+        node_mapping: Dict[str, ov.Node],
+    ) -> List[ov.Node]:
+        """
+        A method for aggregating layers to be further cloned.
+        Aggregation is designed in such a way that layers are listed from right to left,
+        as they pass from bottom to top. This is done in order to find all constants in the model and
+        to start graph creation from them (as well as Parameter layers), because
+        OpenVINO graph is created from top-down and cannot be created otherwise.
+
+        Legend: w - weigths, c - convert, il/ih - input low/high, ol/oh - output low/high
+        (w)
+         |
+        (c) (il) (ih) (ol) (oh)
+          \   |   |   /   /
+           (fake quantize) (parameter)
+                  \           /
+                  (convolution)
+                        |
+                    (result)
+        Based on the above graph, the return value would look like this:
+        [convolution, parameter, fake quantize, oh, ol, ih, il, c, w]
+
+        :param input_ids: List of the points in the special format - (node_name, port_id).
+            This helps to point to the precise part of the model that may be used to define the subgraph inputs.
+        :param output_ids: List of the points in the special format - (node_name, port_id).
+            This helps to point to the precise part of the model that may be used to define the subgraph outputs.
+        :param node_mapping: Original nodes mapping.
+        :return: List of the ov.Nodes to clone.
+        """
+        # Creating a list as a deque for FIFO layer acquisition and retrieval
+        lookup_nodes = deque(node_mapping[n] for n, _ in output_ids)
+        graph_nodes = []
+
+        while lookup_nodes:
+            lookup_node = lookup_nodes.popleft()
+            lookup_name = lookup_node.get_friendly_name()
+            node_inputs = lookup_node.inputs()
+            graph_nodes.append(lookup_node)
+            # Reversing to lookup nodes from right to left
+            for node_input in reversed(node_inputs):
+                port_id = node_input.get_index()
+                if (lookup_name, port_id) in input_ids:
+                    # We create Parameters here to avoid double creation in the future since it is not an original node,
+                    # but we need to have it as input for next node.
+                    parameter = self._create_parameter(lookup_name, node_input)
+                    lookup_nodes.append(parameter)
+                    continue
+                parent_node = node_input.get_source_output().get_node()
+                lookup_nodes.append(parent_node)
+
+        return graph_nodes
+
+    def build(
+        self,
+        input_ids: List[Tuple[str, int]],
+        output_ids: List[Tuple[str, int]],
+        node_mapping: Dict[str, ov.Node],
+    ) -> ov.Model:
+        """
+        The basic method of the algorithm. This method uses an aggregated list of layers to be recreated.
+        Let us take a graph of this kind as an example:
+
+        Legend: w - weigths, c - convert, il/ih - input low/high, ol/oh - output low/high
+        (w)
+         |
+        (c) (il) (ih) (ol) (oh)
+          \   |   |   /   /
+           (fake quantize) (parameter)
+                  \           /
+                  (convolution)
+                        |
+                    (result)
+
+        The externally collected list of layers will look like this:
+        [convolution, parameter, fake quantize, oh, ol, ih, il, c, w]
+
+        Next, this list will be circled from right to left. At the same time, the list of already created layers
+        will be filled from left to right, which will be used in the traversal step also, from left to right,
+        in order to keep the order of the original layer inputs.
+        For example:
+
+            graph_nodes = [convolution, parameter, fake quantize, oh, ol, ih, il, c, w]
+            clone_nodes = []
+
+        *creating w - weight node.*
+            graph_nodes = [convolution, parameter, fake quantize, oh, ol, ih, il, c]
+            clone_nodes = [w]
+
+        *creating c - convert node.
+        Based on the .inputs() output, we'll use the already created w-weight node to fill in the convert input.
+        As the result, weight node would be removed from the clone_nodes list and convert node would be placed here.*
+            graph_nodes = [convolution, parameter, fake quantize, oh, ol, ih, il]
+            clone_nodes = [c]
+
+        *creating il/ih - input low/high, ol/oh - output low/high nodes.
+        Since these nodes are constants and do not require any nodes as inputs, cloned nodes will not be used.*
+            graph_nodes = [convolution, parameter, fake quantize, oh, ol, ih, il]
+            clone_nodes = [c, il, ih, ol, oh]
+
+        *creating fake quantize node.
+        This node requires to have input values in a specific order.
+        All previous nodes will be connected/used for fake quantize, from left to right.*
+            graph_nodes = [convolution, parameter]
+            clone_nodes = [f]
+
+        *creating parameter node.
+        In this step, the list of parameters will also be filled out with the new node.*
+            graph_nodes = [convolution]
+            clone_nodes = [f, parameter]
+
+        *creating convolution node.
+        This node also requires to have inputs in a specific order.
+        All previous nodes will be connected/used for convolution, from left to right. Also,
+        the outputs verification step will show here that one of the convolution outputs is in the output_ids list.
+        This means that the Result node would be created and placed into the results list.*
+            graph_nodes = []
+            clone_nodes = [convolution]
+
+        The last step is to create a subgraph model based on the parameters & results lists.
+
+        :param input_ids: List of the points in the special format - (node_name, port_id).
+            This helps to point to the precise part of the model that may be used to define the subgraph inputs.
+        :param output_ids: List of the points in the special format - (node_name, port_id).
+            This helps to point to the precise part of the model that may be used to define the subgraph outputs.
+        :param node_mapping: Original nodes mapping.
+        :return: Builded ov.Model based on parameters.
+        """
+
+        parameters, results = [], []
+        clone_nodes = deque()
+
+        # Collecting nodes that declares the graph.
+        graph_nodes = self._collect_graph_nodes(input_ids, output_ids, node_mapping)
+
+        while graph_nodes:
+            graph_node = graph_nodes.pop()
+            node_type = graph_node.get_type_name()
+            node_name = graph_node.get_friendly_name()
+
+            # To create the new OpenVINO nodes, we need to provide all possible layer attributes.
+            attrs = graph_node.get_attributes()
+            attrs["name"] = node_name
+
+            if node_type == "Constant":
+                # Constants creation is apart due to specific behavior.
+                clone_node = OVModelTransformer._create_constant(
+                    graph_node.get_data(), dtype=graph_node.get_element_type(), name=attrs["name"]
+                )
+            elif node_type == "Parameter":
+                # We've created Parameter nodes on the previous step.
+                clone_node = graph_node
+                parameters.append(clone_node)
+            else:
+                # We have to have args as the inputs since all of them are nodes and are required to be as input.
+                args = [clone_nodes.popleft() for _ in graph_node.inputs()]
+
+                clone_node = self._node_factory.create(node_type, args, attrs)
+
+                for node_output in clone_node.outputs():
+                    port_id = node_output.get_index()
+                    if (node_name, port_id) in output_ids:
+                        result = self._create_result(node_name, node_output)
+                        results.append(result)
+
+            clone_nodes.append(clone_node)
+
+        return ov.Model(results, parameters)

nncf/openvino/graph/node_utils.py

@@ -121,18 +121,22 @@ def get_const_value(const_node: ov.Node) -> np.ndarray:
     return const_node.data
 
 
-def get_bias_value(node_with_bias: NNCFNode, nncf_graph: NNCFGraph, model: ov.Model) -> np.ndarray:
+def get_bias_value(
+    node_with_bias: NNCFNode, nncf_graph: NNCFGraph, model: ov.Model, node_mapping: Dict[str, ov.Node] = None
+) -> np.ndarray:
     """
     Returns the bias tensor for the biased node.
 
     :param node_with_bias: The node that corresponds to the operation with bias.
     :param nncf_graph: NNCFGraph instance.
     :param model: The model that contains this operation.
+    :param node_mapping: Original nodes mapping cache.
     :return: The bias value that is applied to the output tensor of the node's operation.
     """
-    ops_dict = {op.get_friendly_name(): op for op in model.get_ops()}
+    if node_mapping is None:
+        node_mapping = {op.get_friendly_name(): op for op in model.get_ops()}
     bias_constant = get_node_with_bias_value(get_add_bias_node(node_with_bias, nncf_graph), nncf_graph)
-    ov_bias_constant = ops_dict[bias_constant.node_name]
+    ov_bias_constant = node_mapping[bias_constant.node_name]
     return get_const_value(ov_bias_constant)
 
 

nncf/quantization/algorithms/fast_bias_correction/algorithm.py

@@ -17,7 +17,6 @@
 from nncf.common.factory import EngineFactory
 from nncf.common.factory import ModelTransformerFactory
 from nncf.common.graph.graph import NNCFGraph
-from nncf.common.graph.model_transformer import ModelTransformer
 from nncf.common.graph.transformations.commands import TargetPoint
 from nncf.common.graph.transformations.commands import TargetType
 from nncf.common.graph.transformations.layout import TransformationLayout
@@ -111,7 +110,7 @@ def _set_backend_entity(self, model: TModel) -> None:
                 OVFastBiasCorrectionAlgoBackend,
             )
 
-            self._backend_entity = OVFastBiasCorrectionAlgoBackend()
+            self._backend_entity = OVFastBiasCorrectionAlgoBackend(model)
         elif model_backend == BackendType.TORCH:
             from nncf.quantization.algorithms.fast_bias_correction.torch_backend import PTFastBiasCorrectionAlgoBackend
 
@@ -167,7 +166,7 @@ def apply(
             # Outputs of the subgraphs for the FastBiasCorrection are the same across the backends.
             output_id = (out_node_name, 0)
 
-            extracted_model = self._extract_submodel(model_transformer, input_id, output_id)
+            extracted_model = self._backend_entity.extract_submodel(model_transformer, input_id, output_id)
             if extracted_model is None:
                 nncf_logger.debug(f"Skipping node {node_name} because cant extract submodel")
                 continue
@@ -287,23 +286,6 @@ def output_filter_func(point):
             output_fp.extend(tensor_collector.get_statistics().mean_values)
         return output_fp
 
-    def _extract_submodel(
-        self, model_transformer: ModelTransformer, input_id: Tuple[str, int], output_id: Tuple[str, int]
-    ) -> TModel:
-        """
-        Extracts sub-model using backend-specific ModelTransformer.
-
-        :param model_transformer: Backend-specific ModelTransformer.
-        :param input_id: Input ID.
-        :param output_id: Output ID.
-        :return: Backend-specific sub-model.
-        """
-        model_extraction_command = self._backend_entity.model_extraction_command([input_id], [output_id])
-        me_transformation_layout = TransformationLayout()
-        me_transformation_layout.register(model_extraction_command)
-        extracted_model = model_transformer.transform(me_transformation_layout)
-        return extracted_model
-
     def _add_statistic_point(self, container: StatisticPointsContainer, point: TargetPoint, axis: int) -> None:
         """
         Adds specific statistic point.

nncf/quantization/algorithms/fast_bias_correction/backend.py

@@ -15,9 +15,11 @@
 
 from nncf.common.graph import NNCFGraph
 from nncf.common.graph import NNCFNode
+from nncf.common.graph.model_transformer import ModelTransformer
 from nncf.common.graph.transformations.commands import TargetPoint
 from nncf.common.graph.transformations.commands import TargetType
 from nncf.common.graph.transformations.commands import TransformationCommand
+from nncf.common.graph.transformations.layout import TransformationLayout
 from nncf.common.tensor_statistics.collectors import TensorStatisticCollectorBase
 from nncf.tensor import Tensor
 
@@ -194,3 +196,20 @@ def get_activation_channel_axis(node: NNCFNode, port_id: int, input_shape: Tuple
         :param input_shape: Shape of the input.
         :return: Channel axis number.
         """
+
+    def extract_submodel(
+        self, model_transformer: ModelTransformer, input_id: Tuple[str, int], output_id: Tuple[str, int]
+    ) -> TModel:
+        """
+        Extracts sub-model using backend-specific ModelTransformer.
+
+        :param model_transformer: Backend-specific ModelTransformer.
+        :param input_id: Input ID.
+        :param output_id: Output ID.
+        :return: Backend-specific sub-model.
+        """
+        model_extraction_command = self.model_extraction_command([input_id], [output_id])
+        me_transformation_layout = TransformationLayout()
+        me_transformation_layout.register(model_extraction_command)
+        extracted_model = model_transformer.transform(me_transformation_layout)
+        return extracted_model

nncf/quantization/algorithms/fast_bias_correction/openvino_backend.py

@@ -20,6 +20,7 @@
 from nncf.experimental.common.tensor_statistics.collectors import TensorCollector
 from nncf.openvino.graph.metatypes.groups import FAKE_QUANTIZE_OPERATIONS
 from nncf.openvino.graph.metatypes.groups import OPERATIONS_WITH_BIAS_REDUCED
+from nncf.openvino.graph.model_builder import OVModelBuilder
 from nncf.openvino.graph.node_utils import get_activation_channel_axis
 from nncf.openvino.graph.node_utils import get_bias_value
 from nncf.openvino.graph.node_utils import is_node_with_bias
@@ -33,6 +34,12 @@
 
 
 class OVFastBiasCorrectionAlgoBackend(FastBiasCorrectionAlgoBackend):
+
+    def __init__(self, model):
+        # Node mapping caching to reduce time for calculations
+        self._node_mapping = {op.get_friendly_name(): op for op in model.get_ops()}
+        self._model_builder = OVModelBuilder()
+
     @staticmethod
     def target_point(target_type: TargetType, target_node_name: str, port_id: int) -> OVTargetPoint:
         return OVTargetPoint(target_type, target_node_name, port_id)
@@ -73,9 +80,8 @@ def create_input_data(
         input_data = {input_name: blob}
         return input_data
 
-    @staticmethod
-    def get_bias_value(node: NNCFNode, nncf_graph: NNCFGraph, model: ov.Model) -> Tensor:
-        return Tensor(get_bias_value(node, nncf_graph, model))
+    def get_bias_value(self, node: NNCFNode, nncf_graph: NNCFGraph, model: ov.Model) -> Tensor:
+        return Tensor(get_bias_value(node, nncf_graph, model, node_mapping=self._node_mapping))
 
     @staticmethod
     def get_activation_port_ids_for_bias_node(node: NNCFNode) -> Tuple[int, int]:
@@ -113,3 +119,11 @@ def get_node_names_for_input_output_statistics(node: NNCFNode, nncf_graph: NNCFG
     @staticmethod
     def get_activation_channel_axis(node: NNCFNode, port_id: int, input_shape: Tuple[int]) -> int:
         return get_activation_channel_axis(node, port_id, input_shape)
+
+    def extract_submodel(self, model_transformer, input_id, output_id):
+
+        return self._model_builder.build(
+            input_ids=[input_id],
+            output_ids=[output_id],
+            node_mapping=self._node_mapping,
+        )

tests/openvino/native/data/2024.6/reference_graphs/original_nncf_graph/built_ConvModel.dot

@@ -0,0 +1,9 @@
+strict digraph  {
+"0 Parameter_Conv.0" [id=0, type=Parameter];
+"1 Convolution_57" [id=1, type=Convolution];
+"2 Result_Conv.0" [id=2, type=Result];
+"3 Conv/Constant_4" [id=3, type=Constant];
+"0 Parameter_Conv.0" -> "1 Convolution_57"  [label="[1, 3, 4, 2]", style=solid];
+"1 Convolution_57" -> "2 Result_Conv.0"  [label="[1, 3, 4, 2]", style=solid];
+"3 Conv/Constant_4" -> "1 Convolution_57"  [label="[3, 3, 1, 1]", style=solid];
+}