mct-nightly 2.4.0.20250630.629__py3-none-any.whl → 2.4.0.20250702.605__py3-none-any.whl

model_compression_toolkit/core/pytorch/default_framework_info.py

@@ -21,12 +21,8 @@ from torch.nn import Conv2d, ConvTranspose2d, Linear
 from torch import sigmoid
 
 from model_compression_toolkit.core.common.framework_info import FrameworkInfo, set_fw_info, ChannelAxisMapping
-from mct_quantizers import QuantizationMethod
 from model_compression_toolkit.constants import SOFTMAX_THRESHOLD
 from model_compression_toolkit.core.pytorch.constants import KERNEL
-from model_compression_toolkit.core.pytorch.quantizer.fake_quant_builder import power_of_two_quantization, \
-    symmetric_quantization, uniform_quantization
-from model_compression_toolkit.core.pytorch.quantizer.lut_fake_quant import activation_lut_kmean_quantizer
 
 
 class PyTorchInfo(FrameworkInfo):
@@ -81,14 +77,6 @@ class PyTorchInfo(FrameworkInfo):
                               SiLU: (-0.279, None),
                               }
 
-    """
-    Mapping from a QuantizationMethod to an activation quantizer function.
-    """
-    activation_quantizer_mapping = {QuantizationMethod.POWER_OF_TWO: power_of_two_quantization,
-                                    QuantizationMethod.SYMMETRIC: symmetric_quantization,
-                                    QuantizationMethod.UNIFORM: uniform_quantization,
-                                    QuantizationMethod.LUT_POT_QUANTIZER: activation_lut_kmean_quantizer}
-
     @classmethod
     def get_kernel_channels(cls, node_type: Any) -> ChannelAxisMapping:
         """

model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/relu_bound_to_power_of_2.py

@@ -95,11 +95,11 @@ class ReLUBoundToPowerOfTwo(common.BaseSubstitution):
             else:
                 return graph
         elif non_linear_node.is_match_type(hardtanh):
-            if (non_linear_node.framework_attr[HARDTANH_MIN_VAL] == 0.0) and not \
-                    (np.log2(non_linear_node.framework_attr[HARDTANH_MAX_VAL]).astype(int) -
-                     np.log2(non_linear_node.framework_attr[HARDTANH_MAX_VAL]) == 0):
-                scale_factor = non_linear_node.framework_attr[HARDTANH_MAX_VAL] / self.threshold
-                non_linear_node.functional_op.__defaults__ = (0.0, self.threshold, non_linear_node.framework_attr[INPLACE])
+            kwargs = non_linear_node.op_call_kwargs
+            if (kwargs[HARDTANH_MIN_VAL] == 0.0) and not \
+                    (np.log2(kwargs[HARDTANH_MAX_VAL]).astype(int) - np.log2(kwargs[HARDTANH_MAX_VAL]) == 0):
+                scale_factor = kwargs[HARDTANH_MAX_VAL] / self.threshold
+                non_linear_node.functional_op.__defaults__ = (0.0, self.threshold, kwargs[INPLACE])
             else:
                 return graph
         else:

model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/shift_negative_activation.py

@@ -29,6 +29,7 @@ from model_compression_toolkit.core.common import BaseNode, Graph
 from model_compression_toolkit.core.common.graph.graph_matchers import EdgeMatcher
 from model_compression_toolkit.core.common.graph.graph_matchers import NodeOperationMatcher
 from model_compression_toolkit.core.common.substitutions.shift_negative_activation import apply_shift_negative_correction
+from model_compression_toolkit.core.pytorch.quantization.activation_quantization_fn_factory import get_activation_quantization_fn_factory
 from model_compression_toolkit.core.pytorch.constants import PAD, VALUE, PADDING, BIAS, USE_BIAS
 from model_compression_toolkit.core.pytorch.utils import to_torch_tensor
 
@@ -239,4 +240,5 @@ def pytorch_apply_shift_negative_correction(graph: Graph,
                                            PADDING,
                                            BIAS,
                                            USE_BIAS,
+                                           get_activation_quantization_fn_factory,
                                            params_search_quantization_fn=params_search_quantization_fn)

model_compression_toolkit/core/pytorch/hessian/weights_hessian_scores_calculator_pytorch.py

@@ -91,7 +91,7 @@ class WeightsHessianScoresCalculatorPytorch(HessianScoresCalculatorPytorch):
             for i, ipt_node in enumerate(self.hessian_request.target_nodes):  # Per Interest point weights tensor
 
                 # Check if the target node's layer type is supported.
-                if not ipt_node.is_kernel_op:
+                if not ipt_node.kernel_attr:
                     Logger.critical(f"Hessian information with respect to weights is not supported for "
                                     f"{ipt_node.type} layers.")  # pragma: no cover
 

model_compression_toolkit/core/pytorch/mixed_precision/configurable_activation_quantizer.py

@@ -20,6 +20,7 @@ from model_compression_toolkit.core.common.mixed_precision.configurable_quantize
     verify_candidates_descending_order, init_activation_quantizers
 from model_compression_toolkit.core.common.quantization.candidate_node_quantization_config import \
     CandidateNodeQuantizationConfig
+from model_compression_toolkit.core.pytorch.quantization.activation_quantization_fn_factory import get_activation_quantization_fn_factory
 from model_compression_toolkit.logger import Logger
 from mct_quantizers import QuantizationMethod
 from mct_quantizers import QuantizationTarget
@@ -67,7 +68,7 @@ class ConfigurableActivationQuantizer(BasePyTorchInferableQuantizer):
                 Logger.critical("Unsupported configuration: Mixing candidates with differing activation quantization states (enabled/disabled).")  # pragma: no cover
 
         # Setting layer's activation
-        self.activation_quantizers = init_activation_quantizers(self.node_q_cfg)
+        self.activation_quantizers = init_activation_quantizers(self.node_q_cfg, get_activation_quantization_fn_factory)
         self.active_quantization_config_index = max_candidate_idx  # initialize with first config as default
 
     def set_active_activation_quantizer(self, index: Optional[int]):

model_compression_toolkit/core/pytorch/pruning/pruning_pytorch_implementation.py

@@ -167,7 +167,7 @@ class PruningPytorchImplementation(PytorchImplementation, PruningFrameworkImplem
         """
 
         attributes_with_axis = {}
-        if node.is_kernel_op:
+        if node.kernel_attr:
             attributes_with_axis[node.kernel_attr] = (node.channel_axis.output, node.channel_axis.input)
 
             # Bias is a vector at the length of the number of output channels.

model_compression_toolkit/core/pytorch/pytorch_implementation.py

@@ -26,7 +26,7 @@ from torch.nn import Module, Sigmoid, Softmax
 
 import model_compression_toolkit.core.pytorch.constants as pytorch_constants
 from model_compression_toolkit.constants import HESSIAN_NUM_ITERATIONS
-from model_compression_toolkit.core import QuantizationConfig, FrameworkInfo, CoreConfig
+from model_compression_toolkit.core import QuantizationConfig, CoreConfig
 from model_compression_toolkit.core import common
 from model_compression_toolkit.core.common import Graph, BaseNode
 from model_compression_toolkit.core.common.framework_implementation import FrameworkImplementation

model_compression_toolkit/core/pytorch/quantization/activation_quantization_fn_factory.py

@@ -0,0 +1,45 @@
+# Copyright 2025 Sony Semiconductor Israel, Inc. All rights reserved.
+#
+# 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.abc import Callable
+
+from mct_quantizers import QuantizationMethod
+from model_compression_toolkit.core.pytorch.quantization.fake_quant_builder import power_of_two_quantization, \
+    symmetric_quantization, uniform_quantization
+from model_compression_toolkit.core.pytorch.quantization.lut_fake_quant import activation_lut_kmean_quantizer
+
+
+"""
+Mapping from a QuantizationMethod to an activation quantizer function.
+"""
+_activation_quantizer_factory_mapping = {
+    QuantizationMethod.POWER_OF_TWO: power_of_two_quantization,
+    QuantizationMethod.SYMMETRIC: symmetric_quantization,
+    QuantizationMethod.UNIFORM: uniform_quantization,
+    QuantizationMethod.LUT_POT_QUANTIZER: activation_lut_kmean_quantizer
+}
+
+
+def get_activation_quantization_fn_factory(quantization_method: QuantizationMethod) -> Callable[[int, dict], Callable]:
+    """
+    Get factory for activation quantizer.
+
+    Args:
+        quantization_method: quantization method for activation.
+
+    Returns:
+        Factory that accepts activation bitwidth and a dict of quantization params, and returns the quantizer.
+    """
+    return _activation_quantizer_factory_mapping[quantization_method]

model_compression_toolkit/core/pytorch/statistics_correction/pytorch_compute_activation_bias_correction_of_graph.py

@@ -18,7 +18,7 @@ from torch.nn import Conv2d, Linear, ConvTranspose2d
 from model_compression_toolkit.core import QuantizationConfig
 from model_compression_toolkit.core.common import Graph
 from model_compression_toolkit.core.common.framework_implementation import FrameworkImplementation
-from model_compression_toolkit.core.common.framework_info import FrameworkInfo
+from model_compression_toolkit.core.pytorch.quantization.activation_quantization_fn_factory import get_activation_quantization_fn_factory
 from model_compression_toolkit.core.common.graph.graph_matchers import NodeOperationMatcher
 from model_compression_toolkit.core.common.statistics_correction.compute_activation_bias_correction_of_graph import \
     compute_activation_bias_correction_of_graph
@@ -50,5 +50,6 @@ def pytorch_compute_activation_bias_correction_of_graph(graph: Graph,
                                                         fw_impl=fw_impl,
                                                         activation_bias_correction_node_matchers=
                                                         activation_bias_correction_node_matchers,
-                                                        kernel_size=KERNEL_SIZE)
+                                                        kernel_size=KERNEL_SIZE,
+                                                        get_activation_quantization_fn_factory=get_activation_quantization_fn_factory)
     return graph

model_compression_toolkit/core/runner.py

@@ -118,7 +118,7 @@ def core_runner(in_model: Any,
     if core_config.is_mixed_precision_enabled:
         if core_config.mixed_precision_config.configuration_overwrite is None:
 
-            filter_candidates_for_mixed_precision(graph, target_resource_utilization, fqc)
+            filter_candidates_for_mixed_precision(graph, target_resource_utilization)
             bit_widths_config = search_bit_width(tg,
                                                  fw_impl,
                                                  target_resource_utilization,

model_compression_toolkit/exporter/model_exporter/pytorch/fakely_quant_onnx_pytorch_exporter.py

@@ -12,14 +12,13 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-from typing import Callable
+from typing import Callable, Optional, List
 from io import BytesIO
 
 import torch.nn
 
 from mct_quantizers import PytorchActivationQuantizationHolder, PytorchQuantizationWrapper
 
-from model_compression_toolkit.core.pytorch.reader.node_holders import DummyPlaceHolder
 from model_compression_toolkit.verify_packages import FOUND_ONNX
 from model_compression_toolkit.logger import Logger
 from model_compression_toolkit.core.pytorch.utils import to_torch_tensor
@@ -65,11 +64,14 @@ if FOUND_ONNX:
             self._use_onnx_custom_quantizer_ops = use_onnx_custom_quantizer_ops
             self._onnx_opset_version = onnx_opset_version
 
-        def export(self, output_names=None) -> None:
+        def export(self, output_names: Optional[List[str]] = None) -> None:
             """
             Convert an exportable (fully-quantized) PyTorch model to a fakely-quant model
             (namely, weights that are in fake-quant format) and fake-quant layers for the activations.
 
+            Args:
+                output_names (Optional[List[str]]): Optional list of output node names for export compatibility.
+
             Returns:
                 Fake-quant PyTorch model.
             """
@@ -131,6 +133,8 @@ if FOUND_ONNX:
                     output_names = ['output']
                     dynamic_axes.update({'output': {0: 'batch_size'}})
             else:
+                assert isinstance(output_names, list), \
+                    f"`output_names` must be a list, but got {type(output_names).__name__}"
                 if isinstance(model_output, (list, tuple)):
                     num_of_outputs = len(model_output)
                 else:

model_compression_toolkit/exporter/model_exporter/pytorch/fakely_quant_torchscript_pytorch_exporter.py

@@ -49,7 +49,7 @@ class FakelyQuantTorchScriptPyTorchExporter(BasePyTorchExporter):
                          save_model_path,
                          repr_dataset)
 
-    def export(self) -> None:
+    def export(self, output_names=None) -> None:
         """
         Convert an exportable (fully-quantized) PyTorch model to a fakely-quant model
         (namely, weights that are in fake-quant format) and fake-quant layers for the activations.

model_compression_toolkit/exporter/model_exporter/pytorch/pytorch_export_facade.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-from typing import Callable
+from typing import Callable, Optional, List
 from packaging import version
 
 from model_compression_toolkit.verify_packages import FOUND_TORCH
@@ -49,7 +49,8 @@ if FOUND_TORCH:
                              is_layer_exportable_fn: Callable = is_pytorch_layer_exportable,
                              serialization_format: PytorchExportSerializationFormat = PytorchExportSerializationFormat.ONNX,
                              quantization_format: QuantizationFormat = QuantizationFormat.MCTQ,
-                             onnx_opset_version=DEFAULT_ONNX_OPSET_VERSION) -> None:
+                             onnx_opset_version=DEFAULT_ONNX_OPSET_VERSION,
+                             output_names: Optional[List[str]] = None) -> None:
         """
         Export a PyTorch quantized model to a torchscript or onnx model.
         The model will be saved to the path in save_model_path.
@@ -67,11 +68,19 @@ if FOUND_TORCH:
             PytorchExportSerializationFormat.ONNX).
             quantization_format: Format of how quantizers are exported (fakely-quant, int8, MCTQ quantizers).
             onnx_opset_version: ONNX opset version to use for exported ONNX model.
+            output_names (Optional[List[str]]): Optional list of output node names for export compatibility.
+            This argument is relevant only when using PytorchExportSerializationFormat.ONNX.
 
         """
         # Ensure 'metadata' is available directly on the model, if present in submodules
         find_and_assign_metadata_attr(model)
 
+        if output_names is not None and serialization_format != PytorchExportSerializationFormat.ONNX:
+            Logger.warning(
+                f'`output_names` is only applicable when exporting to ONNX. '
+                f'Current serialization format is {serialization_format}, so `output_names` will be ignored.'
+            )  # pragma: no cover
+
         if serialization_format == PytorchExportSerializationFormat.TORCHSCRIPT:
             if quantization_format in supported_serialization_quantization_export_dict[serialization_format]:
                 exporter = FakelyQuantTorchScriptPyTorchExporter(model,
@@ -109,7 +118,7 @@ if FOUND_TORCH:
                 f'Unsupported serialization {serialization_format} was used to export Pytorch model.'
                 f' Please see API for supported formats.')  # pragma: no cover
 
-        exporter.export()
+        exporter.export(output_names=output_names)
 
 else:
     def pytorch_export_model(*args, **kwargs):

model_compression_toolkit/pruning/keras/pruning_facade.py

@@ -17,6 +17,7 @@ from typing import Callable, Tuple, Union
 
 from model_compression_toolkit import get_target_platform_capabilities
 from model_compression_toolkit.constants import TENSORFLOW
+from model_compression_toolkit.quantization_preparation.load_fqc import load_fqc_configuration
 from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import TargetPlatformCapabilities
 from model_compression_toolkit.target_platform_capabilities.tpc_io_handler import load_target_platform_capabilities
 from model_compression_toolkit.verify_packages import FOUND_TF
@@ -24,10 +25,8 @@ from model_compression_toolkit.core.common.mixed_precision.resource_utilization_
 from model_compression_toolkit.core.common.pruning.pruner import Pruner
 from model_compression_toolkit.core.common.pruning.pruning_config import PruningConfig
 from model_compression_toolkit.core.common.pruning.pruning_info import PruningInfo
-from model_compression_toolkit.core.common.quantization.set_node_quantization_config import set_quantization_configuration_to_graph
 from model_compression_toolkit.core.graph_prep_runner import read_model_to_graph
 from model_compression_toolkit.logger import Logger
-from model_compression_toolkit.core.common.quantization.quantization_config import DEFAULTCONFIG
 from model_compression_toolkit.target_platform_capabilities.constants import DEFAULT_TP_MODEL
 
 if FOUND_TF:
@@ -117,20 +116,17 @@ if FOUND_TF:
 
         target_platform_capabilities = load_target_platform_capabilities(target_platform_capabilities)
         # Attach tpc model to framework
-        attach2keras = AttachTpcToKeras()
-        target_platform_capabilities = attach2keras.attach(target_platform_capabilities)
+        framework_platform_capabilities = AttachTpcToKeras().attach(target_platform_capabilities)
 
         # Convert the original Keras model to an internal graph representation.
         float_graph = read_model_to_graph(model,
                                           representative_data_gen,
-                                          target_platform_capabilities,
+                                          framework_platform_capabilities,
                                           fw_impl)
 
         # Apply quantization configuration to the graph. This step is necessary even when not quantizing,
         # as it prepares the graph for the pruning process.
-        float_graph_with_compression_config = set_quantization_configuration_to_graph(float_graph,
-                                                                                      quant_config=DEFAULTCONFIG,
-                                                                                      mixed_precision_enable=False)
+        float_graph_with_compression_config = load_fqc_configuration(float_graph, framework_platform_capabilities)
 
         # Create a Pruner object with the graph and configuration.
         pruner = Pruner(float_graph_with_compression_config,
@@ -138,7 +134,7 @@ if FOUND_TF:
                         target_resource_utilization,
                         representative_data_gen,
                         pruning_config,
-                        target_platform_capabilities)
+                        framework_platform_capabilities)
 
         # Apply the pruning process.
         pruned_graph = pruner.prune_graph()

model_compression_toolkit/pruning/pytorch/pruning_facade.py

@@ -23,10 +23,9 @@ from model_compression_toolkit.core.common.mixed_precision.resource_utilization_
 from model_compression_toolkit.core.common.pruning.pruner import Pruner
 from model_compression_toolkit.core.common.pruning.pruning_config import PruningConfig
 from model_compression_toolkit.core.common.pruning.pruning_info import PruningInfo
-from model_compression_toolkit.core.common.quantization.set_node_quantization_config import set_quantization_configuration_to_graph
+from model_compression_toolkit.quantization_preparation.load_fqc import load_fqc_configuration
 from model_compression_toolkit.core.graph_prep_runner import read_model_to_graph
 from model_compression_toolkit.logger import Logger
-from model_compression_toolkit.core.common.quantization.quantization_config import DEFAULTCONFIG
 from model_compression_toolkit.target_platform_capabilities.constants import DEFAULT_TP_MODEL
 
 
@@ -134,9 +133,7 @@ if FOUND_TORCH:
 
         # Apply quantization configuration to the graph. This step is necessary even when not quantizing,
         # as it prepares the graph for the pruning process.
-        float_graph_with_compression_config = set_quantization_configuration_to_graph(float_graph,
-                                                                                      quant_config=DEFAULTCONFIG,
-                                                                                      mixed_precision_enable=False)
+        float_graph_with_compression_config = load_fqc_configuration(float_graph, framework_platform_capabilities)
 
         # Create a Pruner object with the graph and configuration.
         pruner = Pruner(float_graph_with_compression_config,

model_compression_toolkit/ptq/keras/quantization_facade.py

@@ -122,7 +122,7 @@ if FOUND_TF:
 
             >>> quantized_model, quantization_info = mct.ptq.keras_post_training_quantization(model, repr_datagen, ru, core_config=config)
 
-            For more configuration options, please take a look at our `API documentation <https://sony.github.io/model_optimization/api/api_docs/modules/mixed_precision_quantization_config.html>`_.
+            For more configuration options, please take a look at our `API documentation <https://sonysemiconductorsolutions.github.io/mct-model-optimization/api/api_docs/modules/mixed_precision_quantization_config.html>`_.
 
          """
 

model_compression_toolkit/qat/keras/quantization_facade.py

@@ -167,7 +167,7 @@ if FOUND_TF:
 
              >>> quantized_model = tf.keras.models.load_model(model_file, custom_objects=custom_objects)
 
-             For more configuration options, please take a look at our `API documentation <https://sony.github.io/model_optimization/api/api_docs/modules/mixed_precision_quantization_config.html>`_.
+             For more configuration options, please take a look at our `API documentation <https://sonysemiconductorsolutions.github.io/mct-model-optimization/api/api_docs/modules/mixed_precision_quantization_config.html>`_.
 
          """
 

model_compression_toolkit/qat/pytorch/quantization_facade.py

@@ -136,7 +136,7 @@ if FOUND_TORCH:
 
              >>> quantized_model, quantization_info = mct.qat.pytorch_quantization_aware_training_init_experimental(model, repr_datagen, core_config=config)
 
-             For more configuration options, please take a look at our `API documentation <https://sony.github.io/model_optimization/api/api_docs/modules/mixed_precision_quantization_config.html>`_.
+             For more configuration options, please take a look at our `API documentation <https://sonysemiconductorsolutions.github.io/mct-model-optimization/api/api_docs/modules/mixed_precision_quantization_config.html>`_.
 
          """
         Logger.warning(

model_compression_toolkit/quantization_preparation/__init__.py

@@ -0,0 +1,14 @@
+# Copyright 2025 Sony Semiconductor Israel, Inc. All rights reserved.
+#
+# 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.
+# ==============================================================================

model_compression_toolkit/quantization_preparation/load_fqc.py

@@ -0,0 +1,223 @@
+# Copyright 2025 Sony Semiconductor Israel, Inc. All rights reserved.
+#
+# 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 typing import List, Optional
+
+from model_compression_toolkit.core.common import Graph, BaseNode
+from model_compression_toolkit.core.common.framework_info import ChannelAxisMapping
+from model_compression_toolkit.core.common.fusion.fusing_info import FusingInfoGenerator
+from model_compression_toolkit.core.common.quantization.candidate_node_quantization_config import \
+    CandidateNodeQuantizationConfig, NodeQuantizationConfig
+from model_compression_toolkit.core.common.quantization.node_quantization_config import \
+    NodeActivationQuantizationConfig, NodeWeightsQuantizationConfig, ActivationQuantizationMode
+from model_compression_toolkit.core.common.quantization.set_node_quantization_config import filter_node_qco_by_graph
+from model_compression_toolkit.logger import Logger
+from model_compression_toolkit.target_platform_capabilities import FrameworkQuantizationCapabilities, \
+    QuantizationConfigOptions, OpQuantizationConfig
+
+
+def load_fqc_configuration(graph: Graph, fqc: FrameworkQuantizationCapabilities):
+    """
+    Set-up graph for quantization per TPC.
+    Each node will contain quantization candidates for mixed precision and the base config for single precision.
+    The graph will contain the fusing info.
+
+    Args:
+        graph: graph.
+        fqc: framework quantization capabilities object.
+
+    Returns:
+        Updated graph.
+    """
+    graph = _set_nodes_quantization_configuration(graph, fqc)
+    graph = _set_fusion_info(graph, fqc)
+
+    return graph
+
+
+def set_quantization_configs_to_node(node: BaseNode,
+                                     graph: Graph,
+                                     fqc: FrameworkQuantizationCapabilities):
+    """
+    Create and set quantization configurations to a node (for both weights and activation).
+
+    Args:
+        node (BaseNode): Node to set its quantization configurations.
+        graph (Graph): Model's internal representation graph.
+        fqc (FrameworkQuantizationCapabilities): FrameworkQuantizationCapabilities to get default OpQuantizationConfig.
+    """
+    qc_options = fetch_qc_options_for_node(node, fqc)
+    base_config, candidates_qcs = filter_node_qco_by_graph(node, fqc, graph, qc_options)
+
+    node_attrs_list = node.get_node_weights_attributes()
+    mp_candidates = [_create_candidate(node.channel_axis, qc, node_attrs_list)
+                     for qc in candidates_qcs]
+    sp_cfg = _create_candidate(node.channel_axis, base_config, node_attrs_list)
+
+    node.quantization_cfg = NodeQuantizationConfig(base_quantization_cfg=sp_cfg,
+                                                   candidates_quantization_cfg=mp_candidates)
+
+    # TODO is not needed anymore as find min/max candidate look for a real max/min, but some tests still count on it
+    node.sort_node_candidates()
+
+    if not node.has_activation:
+        node.quantization_cfg.update_activation_quantization_mode(ActivationQuantizationMode.NO_QUANT)
+
+    _disable_unsupported_quant_preserving(node, graph)
+
+
+def fetch_qc_options_for_node(node: BaseNode,
+                              fqc: FrameworkQuantizationCapabilities,
+                              return_default=True) -> Optional[QuantizationConfigOptions]:
+    """
+    Get quantization configuration options for the node from TPC.
+
+    Args:
+        node: node for which to fetch quantization configuration.
+        fqc: framework quantization capabilities.
+        return_default: whether to return the default qco or None if node op is not in FQC.
+
+    Returns:
+        Quantization configuration options for the node.
+    """
+    # qcos by filters
+    filter_matches = [(fl, qco) for fl, qco in fqc.filterlayer2qco.items() if node.is_match_filter_params(fl)]
+    fls, filter_qcos = zip(*filter_matches) if filter_matches else (None, None)
+    if filter_qcos and any(qco != filter_qcos[0] for qco in filter_qcos[1:]):
+        raise ValueError(f'Cannot assign quantization configuration to {node} as it matches more than one filter with '
+                         f'conflicting configs: {fls}.')
+
+    # qco by opset
+    # must use is_match_type for functional op in TF2.15
+    matches = [(op_type, qco) for op_type, qco in fqc.layer2qco.items() if node.is_match_type(op_type)]
+    op_types, qcos = zip(*matches) if matches else (None, None)
+    if qcos and any(qco != qcos[0] for qco in qcos[1:]):
+        raise ValueError(f'Cannot assign quantization configuration to {node} as it matches more than one op type with '
+                         f'conflicting configs: {op_types}.')
+
+    # if node matches by both filter and opset, filter takes priority
+    if filter_qcos:
+        return filter_qcos[0]
+
+    if qcos:
+        return qcos[0]
+
+    return fqc.tpc.default_qco if return_default else None
+
+
+def _set_nodes_quantization_configuration(graph: Graph,
+                                          fqc: FrameworkQuantizationCapabilities) -> Graph:
+    """
+    Set quantization configuration for each graph node.
+
+    Args:
+        graph: graph to set with quantization configuration.
+        fqc: framework quantization capabilities.
+
+    Returns:
+        Graph: The graph with quantization configurations attached to each node in it.
+    """
+    _validate_custom_ops_have_qco(graph, fqc)
+
+    for n in graph.get_topo_sorted_nodes():
+        set_quantization_configs_to_node(node=n,
+                                         graph=graph,
+                                         fqc=fqc)
+    return graph
+
+
+def _set_fusion_info(graph: Graph, fqc: FrameworkQuantizationCapabilities) -> Graph:
+    """
+
+    Args:
+        graph: graph.
+        fqc: quantization capabilities with attached framework.
+
+    Returns:
+
+    """
+    # TODO fix the dict with const keys inside get_fusing_patterns. use named tuple or class
+    # TODO irena instead of storing fusion inside graph (including tpc objects) and then let graph convert tpc op config to
+    #  node config, do it here and only store in graph whatever is relevant after this stage.
+    fusing_info = FusingInfoGenerator(fqc.get_fusing_patterns()).generate_fusing_info(graph)
+    graph.fusing_info = fusing_info
+    graph.override_fused_node_activation_quantization_candidates()
+    return graph
+
+
+def _disable_unsupported_quant_preserving(node: BaseNode, graph: Graph):
+    """
+    Disable quantization for quantization preserving ops in cases it cannot be supported
+    (multiple inputs or un-quantized previous node).
+
+    Args:
+        node: current node.
+        graph: graph.
+    """
+    if not node.quantization_cfg.get_activation_quant_mode() == ActivationQuantizationMode.PRESERVE_QUANT:
+        return
+
+    prev_nodes = graph.get_prev_nodes(node)
+    if len(prev_nodes) != 1:
+        Logger.info(f'Disabling Quantization-Preserving for node {node.name} with {len(prev_nodes)} inputs.')
+        node.quantization_cfg.update_activation_quantization_mode(ActivationQuantizationMode.NO_QUANT)
+    elif prev_nodes[0].quantization_cfg.get_activation_quant_mode() == ActivationQuantizationMode.NO_QUANT:
+        Logger.info(f'Disabling Quantization-Preserving for node {node.name} since previous node activation '
+                    f'quantization is disabled.')
+        node.quantization_cfg.update_activation_quantization_mode(ActivationQuantizationMode.NO_QUANT)
+
+
+# TODO irena copied from graph.set_fqc as is. Why does it have Keras errors?
+def _validate_custom_ops_have_qco(graph, fqc):
+    custom_nodes = [n for n in graph.nodes if n.is_custom]
+    for n in custom_nodes:
+        qco = fetch_qc_options_for_node(n, fqc, return_default=False)
+        if not qco:
+            Logger.critical(f'MCT does not support optimizing Keras custom layers. Found a layer of type {n.type}. '
+                            ' Please add the custom layer to Framework Quantization Capabilities (FQC), or file a feature '
+                            'request or an issue if you believe this should be supported.')  # pragma: no cover
+        if any([qc.default_weight_attr_config.enable_weights_quantization for qc in qco.quantization_configurations]):
+            Logger.critical(f'Layer identified: {n.type}. MCT does not support weight quantization for Keras custom layers.')  # pragma: no cover
+
+
+def _create_candidate(weight_channel_axis: ChannelAxisMapping,
+                      op_cfg: OpQuantizationConfig,
+                      node_attrs_list: List[str]) -> CandidateNodeQuantizationConfig:
+    """
+    Create quantization configuration candidate.
+
+    Args:
+        weight_channel_axis: channels axes of the node's kernel.
+        op_cfg: quantization config for the op.
+        node_attrs_list: A list of the node's weights attributes names.
+
+    Returns:
+        Candidate quantization config.
+    """
+
+    aqc = NodeActivationQuantizationConfig(op_cfg=op_cfg)
+
+    # TODO: remove this validation and warning once enabling all attributes quantization by default
+    attrs_with_enabled_quantization = [attr for attr, cfg in op_cfg.attr_weights_configs_mapping.items()
+                                       if cfg.enable_weights_quantization]
+    if len(attrs_with_enabled_quantization) > 1:
+        Logger.warning(f"Multiple weights attributes quantization is enabled via the provided FQC."
+                       f"Quantizing any attribute other than the kernel is experimental "
+                       f"and may be subject to unstable behavior."
+                       f"Attributes with enabled weights quantization: {attrs_with_enabled_quantization}.")
+    wqc = NodeWeightsQuantizationConfig(op_cfg=op_cfg,
+                                        weights_channels_axis=weight_channel_axis,
+                                        node_attrs_list=node_attrs_list)
+
+    return CandidateNodeQuantizationConfig(activation_quantization_cfg=aqc, weights_quantization_cfg=wqc)

model_compression_toolkit/target_platform_capabilities/constants.py

@@ -29,7 +29,7 @@ QNNPACK_TP_MODEL = 'qnnpack'
 # TP Attributes
 KERNEL_ATTR = "kernel_attr"
 BIAS_ATTR = "bias_attr"
-POS_ATTR = "pos_attr"
+POSITIONAL_ATTR = "pos_attr"
 
 # TODO: this is duplicated from the core frameworks constants files, because the original consts can't be used here
 #  duo to circular dependency. It might be best to extract the constants from the core file and put them here (in a