mct-nightly 2.3.0.20250408.522__tar.gz → 2.3.0.20250410.526__tar.gz

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.3.0.20250408.522
+Version: 2.3.0.20250410.526
 Summary: A Model Compression Toolkit for neural networks
 Classifier: Programming Language :: Python :: 3
 Classifier: License :: OSI Approved :: Apache Software License
@@ -22,6 +22,7 @@ Requires-Dist: scipy
 Requires-Dist: protobuf
 Requires-Dist: mct-quantizers-nightly
 Requires-Dist: pydantic<2.0
+Requires-Dist: sony-custom-layers-dev==0.4.0.dev6
 Dynamic: classifier
 Dynamic: description
 Dynamic: description-content-type

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/mct_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.3.0.20250408.522
+Version: 2.3.0.20250410.526
 Summary: A Model Compression Toolkit for neural networks
 Classifier: Programming Language :: Python :: 3
 Classifier: License :: OSI Approved :: Apache Software License
@@ -22,6 +22,7 @@ Requires-Dist: scipy
 Requires-Dist: protobuf
 Requires-Dist: mct-quantizers-nightly
 Requires-Dist: pydantic<2.0
+Requires-Dist: sony-custom-layers-dev==0.4.0.dev6
 Dynamic: classifier
 Dynamic: description
 Dynamic: description-content-type

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/mct_nightly.egg-info/requires.txt

@@ -11,3 +11,4 @@ scipy
 protobuf
 mct-quantizers-nightly
 pydantic<2.0
+sony-custom-layers-dev==0.4.0.dev6

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/__init__.py

@@ -27,4 +27,4 @@ from model_compression_toolkit import data_generation
 from model_compression_toolkit import pruning
 from model_compression_toolkit.trainable_infrastructure.keras.load_model import keras_load_quantized_model
 
-__version__ = "2.3.0.20250408.000522"
+__version__ = "2.3.0.20250410.000526"

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/common/framework_implementation.py

@@ -93,6 +93,17 @@ class FrameworkImplementation(ABC):
         raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s to_tensor method.')  # pragma: no cover
 
+    @abstractmethod
+    def is_tuple_of_tensors(self, obj: Any) -> bool:
+        """
+        Check if a given object if a tuple of tensors
+        :param obj: Object to check its type
+        :return: True if obj is a tuple of tensors, False otherwise
+        """
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
+                             f'framework\'s is_tuple_of_tensors method.')  # pragma: no cover
+
+
     @abstractmethod
     def model_reader(self,
                      model: Any,

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/common/fusion/fusing_info.py

@@ -150,7 +150,6 @@ class FusingInfo:
         """
         return self.fusing_data
 
-
     @staticmethod
     def generate_fused_op_id(nodes: List['BaseNode']) -> str:
         """
@@ -166,7 +165,7 @@ class FusingInfo:
         id = FUSED_OP_ID_PREFIX + '_'.join([node.name for node in nodes])
         return id
 
-    def validate(self, graph) -> None:
+    def validate(self, graph: 'Graph') -> None:
         """
         Validate that the fusing information is consistent with the given graph and generation logic.
 
@@ -267,7 +266,7 @@ class FusingInfoGenerator:
     def __init__(self, fusing_patterns):
         self._fusing_patterns = fusing_patterns
 
-    def generate_fusing_info(self, graph) -> FusingInfo:
+    def generate_fusing_info(self, graph: 'Graph') -> FusingInfo:
         """
         Generate fusing information based on the graph and fusing patterns.
 
@@ -289,7 +288,7 @@ class FusingInfoGenerator:
             return FusingInfo(fusing_patterns=self._fusing_patterns)
 
         # Find max fusion
-        max_layers_fusing = 0 if len(self._fusing_patterns) == 0 else max([len(fusing_pattern) for fusing_pattern in self._fusing_patterns])
+        max_layers_fusing = max([len(fusing_pattern) for fusing_pattern in self._fusing_patterns])
 
         # Travel along the graph to find layers for fusing
         nodes = graph.get_topo_sorted_nodes()
@@ -331,7 +330,7 @@ def get_valid_fusing_patterns_for_node(fusing_patterns: List[List[Any]],
     Returns only the fusing patterns where a specific layer (at index idx) matches the given node — either by type or filter params.
 
     Args:
-        fusing_patterns: supported fusings
+        fusing_patterns: supported fusing patterns
         node: node to decide if it can be a part of fusion
         idx: index of layer in the fusion
 

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/common/graph/base_graph.py

@@ -33,6 +33,7 @@ from model_compression_toolkit.core.common.collectors.statistics_collector impor
 from model_compression_toolkit.core.common.collectors.statistics_collector import scale_statistics, shift_statistics
 from model_compression_toolkit.core.common.pruning.pruning_section import PruningSection
 from model_compression_toolkit.core.common.user_info import UserInformation
+from model_compression_toolkit.core.common.quantization.node_quantization_config import ActivationQuantizationMode
 from model_compression_toolkit.logger import Logger
 from model_compression_toolkit.target_platform_capabilities.targetplatform2framework import LayerFilterParams
 from model_compression_toolkit.target_platform_capabilities.targetplatform2framework.framework_quantization_capabilities import \
@@ -920,7 +921,7 @@ class Graph(nx.MultiDiGraph, GraphSearches):
         nodes_to_disable = [node for nodes in self.fusing_info.get_all_fused_operations().values() for node in nodes[:-1]]
         for node in nodes_to_disable:
             for qc in node.candidates_quantization_cfg:
-                qc.activation_quantization_cfg.enable_activation_quantization = False
+                qc.activation_quantization_cfg.quant_mode = ActivationQuantizationMode.FLN_QUANT
 
     def validate(self):
         """

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/common/graph/base_node.py

@@ -20,7 +20,8 @@ import numpy as np
 
 from model_compression_toolkit.constants import WEIGHTS_NBITS_ATTRIBUTE, CORRECTED_BIAS_ATTRIBUTE, \
     ACTIVATION_N_BITS_ATTRIBUTE, FP32_BYTES_PER_PARAMETER
-from model_compression_toolkit.core.common.quantization.node_quantization_config import WeightsAttrQuantizationConfig
+from model_compression_toolkit.core.common.quantization.node_quantization_config import WeightsAttrQuantizationConfig, \
+    ActivationQuantizationMode
 from model_compression_toolkit.logger import Logger
 from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import QuantizationConfigOptions, \
     OpQuantizationConfig
@@ -116,33 +117,28 @@ class BaseNode:
         """
         return any(isinstance(key, int) for key in self.weights.keys())
 
+    def _is_single_quant_mode(self, q_mode: ActivationQuantizationMode) -> bool:
+        """ Check whether all candidates have the same unique quantization mode, and if it is 'q_mode'. """
+
+        if self.final_activation_quantization_cfg:
+            # if we have a final configuration, then we only care to check if it enables activation quantization.
+            return self.final_activation_quantization_cfg.quant_mode == q_mode
+
+        q_modes = {qc.activation_quantization_cfg.quant_mode for qc in self.candidates_quantization_cfg}
+        assert len(q_modes) == 1
+        return q_modes.pop() == q_mode
+
     def is_activation_quantization_enabled(self) -> bool:
         """
-
         Returns: Whether node activation quantization is enabled or not.
-
         """
-        if self.final_activation_quantization_cfg:
-            # if we have a final configuration, then we only care to check if it enables activation quantization
-            return self.final_activation_quantization_cfg.enable_activation_quantization
-
-        for qc in self.candidates_quantization_cfg:
-            assert self.candidates_quantization_cfg[0].activation_quantization_cfg.enable_activation_quantization == \
-                   qc.activation_quantization_cfg.enable_activation_quantization
-        return self.candidates_quantization_cfg[0].activation_quantization_cfg.enable_activation_quantization
+        return self._is_single_quant_mode(ActivationQuantizationMode.QUANT)
 
     def is_quantization_preserving(self) -> bool:
         """
         Returns: Whether node activation quantization information is preserved from its inputs.
         """
-        if self.final_activation_quantization_cfg:
-            # if we have a final configuration, then we only care to check if it enables activation quantization.
-            return self.final_activation_quantization_cfg.quantization_preserving
-
-        for qc in self.candidates_quantization_cfg:
-            assert self.candidates_quantization_cfg[0].activation_quantization_cfg.quantization_preserving == \
-                   qc.activation_quantization_cfg.quantization_preserving
-        return self.candidates_quantization_cfg[0].activation_quantization_cfg.quantization_preserving
+        return self._is_single_quant_mode(ActivationQuantizationMode.PRESERVE_QUANT)
 
     def is_weights_quantization_enabled(self, attr_name: str) -> bool:
         """

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/common/hessian/hessian_scores_calculator.py

@@ -72,8 +72,7 @@ class HessianScoresCalculator(ABC):
         """
         raise NotImplemented(f'{self.__class__.__name__} have to implement compute method.')  # pragma: no cover
 
-    @staticmethod
-    def unfold_tensors_list(tensors_to_unfold: Any) -> List[Any]:
+    def unfold_tensors_list(self, tensors_to_unfold: Any) -> List[Any]:
         """
         Unfold (flatten) a nested tensors list.
         Given a mixed list of single tensors and nested tensor lists,
@@ -85,6 +84,9 @@ class HessianScoresCalculator(ABC):
         """
         unfold_tensors = []
         for tensor in tensors_to_unfold:
+            if self.fw_impl.is_tuple_of_tensors(tensor):
+                tensor = list(tensor)  # converts named tuple to list
+
             if isinstance(tensor, List):
                 unfold_tensors += tensor
             else:

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_calculator.py

@@ -31,7 +31,7 @@ from model_compression_toolkit.core.common.graph.virtual_activation_weights_node
 from model_compression_toolkit.core.common.mixed_precision.resource_utilization_tools.resource_utilization import \
     RUTarget, ResourceUtilization
 from model_compression_toolkit.core.common.quantization.node_quantization_config import NodeWeightsQuantizationConfig, \
-    NodeActivationQuantizationConfig, BaseNodeQuantizationConfig
+    NodeActivationQuantizationConfig, BaseNodeQuantizationConfig, ActivationQuantizationMode
 from model_compression_toolkit.core.common.substitutions.virtual_activation_weights_composition import \
     get_input_activation_if_composable
 
@@ -710,7 +710,7 @@ class ResourceUtilizationCalculator:
         """
         if act_qc:
             assert bitwidth_mode == BitwidthMode.QCustom
-            return act_qc.activation_n_bits if act_qc.enable_activation_quantization else FLOAT_BITWIDTH
+            return act_qc.activation_n_bits if act_qc.quant_mode == ActivationQuantizationMode.QUANT else FLOAT_BITWIDTH
 
         if bitwidth_mode == BitwidthMode.Float or not (n.is_activation_quantization_enabled() or
                                                        n.is_quantization_preserving()):

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/common/mixed_precision/sensitivity_evaluation.py

@@ -20,6 +20,7 @@ from typing import Callable, Any, List, Tuple
 from model_compression_toolkit.constants import AXIS
 from model_compression_toolkit.core import FrameworkInfo, MixedPrecisionQuantizationConfig
 from model_compression_toolkit.core.common import Graph, BaseNode
+from model_compression_toolkit.core.common.quantization.node_quantization_config import ActivationQuantizationMode
 from model_compression_toolkit.core.common.graph.functional_node import FunctionalNode
 from model_compression_toolkit.core.common.similarity_analyzer import compute_kl_divergence
 from model_compression_toolkit.core.common.model_builder_mode import ModelBuilderMode
@@ -207,7 +208,7 @@ class SensitivityEvaluation:
         if self.disable_activation_for_metric:
             for n in evaluation_graph.get_topo_sorted_nodes():
                 for c in n.candidates_quantization_cfg:
-                    c.activation_quantization_cfg.enable_activation_quantization = False
+                    c.activation_quantization_cfg.quant_mode = ActivationQuantizationMode.NO_QUANT
 
         model_mp, _, conf_node2layers = self.fw_impl.model_builder(evaluation_graph,
                                                                    mode=ModelBuilderMode.MIXEDPRECISION,

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/common/quantization/debug_config.py

@@ -29,8 +29,10 @@ class DebugConfig:
          enabled) or not. Can be used to pinpoint problematic layers in the quantization process.
         network_editor (List[EditRule]): A list of rules and actions to edit the network for quantization.
         simulate_scheduler (bool): Simulate scheduler behavior to compute operators' order and cuts.
+        bypass (bool): A flag to enable MCT bypass, which skips MCT runner and returns the input model unchanged.
     """
 
     analyze_similarity: bool = False
     network_editor: List[EditRule] = field(default_factory=list)
     simulate_scheduler: bool = False
+    bypass: bool = False

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/common/quantization/node_quantization_config.py

@@ -15,7 +15,7 @@
 
 
 from typing import Callable, Any, List, Tuple, Union, Dict, TYPE_CHECKING
-
+from enum import Enum, auto
 import numpy as np
 
 from model_compression_toolkit.core.common.quantization.quantization_fn_selection import get_weights_quantization_fn
@@ -40,6 +40,14 @@ if TYPE_CHECKING:
 ##########################################
 
 
+class ActivationQuantizationMode(Enum):
+    """ An enum defining the output activation quantization mode of  a node. """
+    QUANT = auto()
+    FLN_QUANT = auto()
+    PRESERVE_QUANT = auto()
+    NO_QUANT = auto()
+
+
 class BaseNodeQuantizationConfig(object):
     """
     Base class for node quantization configuration
@@ -100,8 +108,14 @@ class NodeActivationQuantizationConfig(BaseNodeQuantizationConfig):
         self.activation_n_bits = op_cfg.activation_n_bits
         self.relu_bound_to_power_of_2 = qc.relu_bound_to_power_of_2
         self.activation_bias_correction_term = None
-        self.enable_activation_quantization = op_cfg.enable_activation_quantization
-        self.quantization_preserving = op_cfg.quantization_preserving
+        if op_cfg.enable_activation_quantization and op_cfg.quantization_preserving:
+            raise ValueError("An OpQuantizationConfig can't have both enable_activation_quantization and quantization_preserving enabled.")
+        if op_cfg.enable_activation_quantization:
+            self.quant_mode = ActivationQuantizationMode.QUANT
+        elif op_cfg.quantization_preserving:
+            self.quant_mode = ActivationQuantizationMode.PRESERVE_QUANT
+        else:
+            self.quant_mode = ActivationQuantizationMode.NO_QUANT
         self.signedness = op_cfg.signedness
         self.activation_channel_equalization = qc.activation_channel_equalization
         self.input_scaling = qc.input_scaling
@@ -113,6 +127,17 @@ class NodeActivationQuantizationConfig(BaseNodeQuantizationConfig):
         self.shift_negative_threshold_recalculation = qc.shift_negative_threshold_recalculation
         self.concat_threshold_update = qc.concat_threshold_update
 
+    @property
+    def enable_activation_quantization(self):
+        return self.quant_mode == ActivationQuantizationMode.QUANT
+
+    @property
+    def quantization_preserving(self):
+        return self.quant_mode == ActivationQuantizationMode.PRESERVE_QUANT
+
+    def fln_quantization(self):
+        return self.quant_mode == ActivationQuantizationMode.FLN_QUANT
+
     def quantize_node_output(self,
                              tensors: Any) -> Any:
         """
@@ -181,7 +206,7 @@ class NodeActivationQuantizationConfig(BaseNodeQuantizationConfig):
             activation_params: Dictionary that contains weight quantization params.
 
         """
-        assert self.enable_activation_quantization
+        assert self.quant_mode == ActivationQuantizationMode.QUANT
         for param_name, param_value in activation_params.items():
             self.activation_quantization_params[param_name] = param_value
 
@@ -203,7 +228,7 @@ class NodeActivationQuantizationConfig(BaseNodeQuantizationConfig):
                self.activation_error_method == other.activation_error_method and \
                self.activation_quantization_method == other.activation_quantization_method and \
                self.activation_n_bits == other.activation_n_bits and \
-               self.enable_activation_quantization == other.enable_activation_quantization and \
+               self.quant_mode == other.quant_mode and \
                self.activation_channel_equalization == other.activation_channel_equalization and \
                self.input_scaling == other.input_scaling and \
                self.min_threshold == other.min_threshold and \
@@ -219,7 +244,7 @@ class NodeActivationQuantizationConfig(BaseNodeQuantizationConfig):
                      self.activation_error_method,
                      self.activation_quantization_method,
                      self.activation_n_bits,
-                     self.enable_activation_quantization,
+                     self.quant_mode,
                      self.activation_channel_equalization,
                      self.input_scaling,
                      self.min_threshold,

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/common/quantization/set_node_quantization_config.py

@@ -25,7 +25,8 @@ from model_compression_toolkit.core.common.framework_info import FrameworkInfo
 from model_compression_toolkit.core.common.graph.base_graph import Graph
 from model_compression_toolkit.core.common.quantization.candidate_node_quantization_config import \
     CandidateNodeQuantizationConfig
-from model_compression_toolkit.core.common.quantization.node_quantization_config import NodeActivationQuantizationConfig
+from model_compression_toolkit.core.common.quantization.node_quantization_config import NodeActivationQuantizationConfig, \
+    ActivationQuantizationMode
 from model_compression_toolkit.core.common.quantization.quantization_config import QuantizationConfig, \
     QuantizationErrorMethod
 from model_compression_toolkit.core.common.quantization.quantization_params_fn_selection import \
@@ -191,8 +192,9 @@ def set_quantization_configs_to_node(node: BaseNode,
     node.sort_node_candidates(fw_info)
 
     for candidate_qc in node.candidates_quantization_cfg:
-        candidate_qc.activation_quantization_cfg.enable_activation_quantization = \
-            candidate_qc.activation_quantization_cfg.enable_activation_quantization and node.get_has_activation()
+        if candidate_qc.activation_quantization_cfg.quant_mode == ActivationQuantizationMode.QUANT and \
+                not node.get_has_activation():
+            candidate_qc.activation_quantization_cfg.quant_mode = ActivationQuantizationMode.NO_QUANT
 
 
 def create_node_activation_qc(qc: QuantizationConfig,

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/common/statistics_correction/compute_activation_bias_correction_of_graph.py

@@ -45,8 +45,7 @@ def get_previous_node_with_activation_quantization(linear_node: BaseNode,
     activation_quantization_config = prev_node.final_activation_quantization_cfg
 
     # Search for node with activation quantization
-    if (activation_quantization_config.enable_activation_quantization and
-            not activation_quantization_config.quantization_preserving):
+    if activation_quantization_config.enable_activation_quantization:
         return prev_node
     else:
         return get_previous_node_with_activation_quantization(prev_node, graph)

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/common/substitutions/batchnorm_reconstruction.py

@@ -22,7 +22,8 @@ import numpy as np
 from model_compression_toolkit.core.common import Graph
 from model_compression_toolkit.core.common.quantization.quantization_config import QuantizationConfig
 from model_compression_toolkit.core import common
-from model_compression_toolkit.core.common.quantization.node_quantization_config import WeightsAttrQuantizationConfig
+from model_compression_toolkit.core.common.quantization.node_quantization_config import WeightsAttrQuantizationConfig, \
+    ActivationQuantizationMode
 from model_compression_toolkit.logger import Logger
 from model_compression_toolkit.core.common.graph.base_node import BaseNode
 from model_compression_toolkit.core.common.graph.graph_matchers import NodeOperationMatcher
@@ -127,7 +128,7 @@ class BatchNormalizationReconstruction(common.BaseSubstitution):
         bn_node.candidates_quantization_cfg = copy.deepcopy(source_node.candidates_quantization_cfg)
 
         for qc in bn_node.candidates_quantization_cfg:
-            qc.activation_quantization_cfg.enable_activation_quantization = False
+            qc.activation_quantization_cfg.quant_mode = ActivationQuantizationMode.NO_QUANT
             for attr in bn_node.get_node_weights_attributes():
                 if qc.weights_quantization_cfg.has_attribute_config(attr):
                     # we only create a BN layer to collect statistics, so we don't need to quantize anything,

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/common/substitutions/shift_negative_activation.py

@@ -17,7 +17,8 @@ import numpy as np
 from typing import List, Tuple, Any, Callable
 
 from model_compression_toolkit.core.common.quantization.quantization_config import QuantizationConfig
-from model_compression_toolkit.core.common.quantization.node_quantization_config import WeightsAttrQuantizationConfig
+from model_compression_toolkit.core.common.quantization.node_quantization_config import WeightsAttrQuantizationConfig, \
+    ActivationQuantizationMode
 from model_compression_toolkit.logger import Logger
 from model_compression_toolkit.core.common import FrameworkInfo, Graph, BaseNode
 from model_compression_toolkit.constants import THRESHOLD, SIGNED, SHIFT_NEGATIVE_NON_LINEAR_NUM_BITS
@@ -363,7 +364,7 @@ def shift_negative_function(graph: Graph,
                                          mixed_precision_enable=core_config.is_mixed_precision_enabled)
 
         for candidate_qc in pad_node.candidates_quantization_cfg:
-            candidate_qc.activation_quantization_cfg.enable_activation_quantization = False
+            candidate_qc.activation_quantization_cfg.quant_mode = ActivationQuantizationMode.NO_QUANT
             for attr in pad_node.get_node_weights_attributes():
                 candidate_qc.weights_quantization_cfg.get_attr_config(attr).enable_weights_quantization = False
 

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/keras/keras_implementation.py

@@ -159,6 +159,19 @@ class KerasImplementation(FrameworkImplementation):
         """
         return to_tf_tensor(tensor)
 
+    def is_tuple_of_tensors(self, obj: Any) -> bool:
+        """
+        Check if a given object if a tuple of tensors
+        :param obj: Object to check its type
+        :return: True if obj is a tuple of tensors, False otherwise
+        """
+        if not isinstance(obj, tuple):
+            return False
+        for item in obj:
+            if not isinstance(item, tf.Tensor):
+                return False
+        return True
+
     def model_builder(self,
                       graph: Graph,
                       mode: ModelBuilderMode,
@@ -454,7 +467,7 @@ class KerasImplementation(FrameworkImplementation):
             return True
 
         return any([node.is_match_type(_type) for _type in [Conv2D, DepthwiseConv2D, Conv2DTranspose, Dense,
-                                                            Concatenate, tf.concat, Add, tf.add]])
+                                                            Concatenate, tf.concat, Add, tf.add, tf.stack]])
 
     def get_mp_node_distance_fn(self, n: BaseNode,
                                 compute_distance_fn: Callable = None,

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/keras/mixed_precision/configurable_activation_quantizer.py

@@ -64,8 +64,7 @@ class ConfigurableActivationQuantizer(BaseKerasInferableQuantizer):
         verify_candidates_descending_order(self.node_q_cfg, kernel_attr)
 
         for qc in node_q_cfg:
-            if qc.activation_quantization_cfg.enable_activation_quantization != \
-                    node_q_cfg[0].activation_quantization_cfg.enable_activation_quantization:
+            if qc.activation_quantization_cfg.quant_mode != node_q_cfg[0].activation_quantization_cfg.quant_mode:
                 Logger.critical("Unsupported configuration: Mixing candidates with differing activation quantization states (enabled/disabled).")  # pragma: no cover
 
         self.activation_quantizers = init_activation_quantizers(self.node_q_cfg)

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/scaled_dot_product_attention.py

@@ -198,7 +198,6 @@ class ScaledDotProductDecomposition(BaseSubstitution):
         :param attention_node: the node to replace
         :return: A graph after the substitution
         """
-        print("In scale_dot_product_attention substitution@@@@@@@@")
         input_nodes = self._get_attention_input_nodes(graph, attention_node)
         q_node, k_node, v_node = input_nodes["q"], input_nodes["k"], input_nodes["v"]
         transpose_k_node = self._get_transpose_k_node(attention_node.name, k_node)

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/pytorch/mixed_precision/configurable_activation_quantizer.py

@@ -63,8 +63,7 @@ class ConfigurableActivationQuantizer(BasePyTorchInferableQuantizer):
         verify_candidates_descending_order(self.node_q_cfg, kernel_attr)
 
         for qc in self.node_q_cfg:
-            if qc.activation_quantization_cfg.enable_activation_quantization != \
-                   self.node_q_cfg[0].activation_quantization_cfg.enable_activation_quantization:
+            if qc.activation_quantization_cfg.quant_mode != self.node_q_cfg[0].activation_quantization_cfg.quant_mode:
                 Logger.critical("Unsupported configuration: Mixing candidates with differing activation quantization states (enabled/disabled).")  # pragma: no cover
 
         # Setting layer's activation

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/pytorch/pytorch_implementation.py

@@ -15,12 +15,12 @@
 import operator
 from copy import deepcopy
 from functools import partial
-from typing import List, Any, Tuple, Callable, Type, Dict, Generator
+from typing import List, Any, Tuple, Callable, Generator
 
 import numpy as np
 import torch
 from mct_quantizers import PytorchQuantizationWrapper, PytorchActivationQuantizationHolder
-from torch import sigmoid, softmax, add, cat, argmax, concat, concatenate
+from torch import sigmoid, softmax, add, cat, argmax, concat, concatenate, stack
 from torch.nn import Conv2d, ConvTranspose2d, Linear
 from torch.nn import Module, Sigmoid, Softmax
 
@@ -144,6 +144,19 @@ class PytorchImplementation(FrameworkImplementation):
         """
         return to_torch_tensor(tensor)
 
+    def is_tuple_of_tensors(self, obj: Any) -> bool:
+        """
+        Check if a given object if a tuple of tensors
+        :param obj: Object to check its type
+        :return: True if obj is a tuple of tensors, False otherwise
+        """
+        if not isinstance(obj, tuple):
+            return False
+        for item in obj:
+            if not isinstance(item, torch.Tensor):
+                return False
+        return True
+
     def model_reader(self,
                      module: Module,
                      representative_data_gen: Callable) -> Graph:
@@ -449,7 +462,7 @@ class PytorchImplementation(FrameworkImplementation):
 
         return any(node.is_match_type(_type) for _type in [Conv2d, Linear, ConvTranspose2d, Sigmoid, sigmoid, Softmax,
                                                            softmax, operator.add, add, cat, concat, concatenate,
-                                                           operator.concat])
+                                                           operator.concat, stack])
 
     def get_mp_node_distance_fn(self, n: BaseNode,
                                 compute_distance_fn: Callable = None,

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/pytorch/reader/reader.py

@@ -13,19 +13,40 @@
 # limitations under the License.
 # ==============================================================================
 
-
-import logging
-from typing import Callable, Dict
-
-import numpy as np
 import torch
-from torch.fx import symbolic_trace
+import logging
+from typing import Callable, Dict, Union, Any
 from torch.fx.passes.shape_prop import ShapeProp
+from torch.fx import Tracer, GraphModule, symbolic_trace
 
 from model_compression_toolkit.logger import Logger
 from model_compression_toolkit.core.common import Graph
 from model_compression_toolkit.core.pytorch.reader.graph_builders import edges_builder, nodes_builder
 from model_compression_toolkit.core.pytorch.utils import set_model
+from sony_custom_layers.pytorch import CustomLayer
+
+
+def _trace_model(root: Union[torch.nn.Module, Callable[..., Any]]) -> GraphModule:
+    """
+    Given an ``nn.Module`` or function instance ``root``, this function will return a ``GraphModule``
+    constructed by recording operations seen while tracing through ``root``.
+    This function replaces torch.fx.symbolic_trace in order to handle custom layers tracing - treating them as graph
+    leafs.
+    :param root: Module or function to be traced and converted into a Graph representation.
+    :return: GraphModule: a Module created from the recorded operations from ``root``.
+    """
+
+    class MCTTracer(Tracer):
+        def is_leaf_module(self, m: torch.nn.Module, module_qualified_name: str) -> bool:
+            if isinstance(m, CustomLayer):
+                return True
+            return super().is_leaf_module(m, module_qualified_name)
+
+    tracer = MCTTracer()
+    graph = tracer.trace(root)
+    # handling the possibility that the model (root) might be a torch.nn.Module or a function
+    model_name = (root.__class__.__name__ if isinstance(root, torch.nn.Module) else root.__name__)
+    return GraphModule(tracer.root, graph, model_name)
 
 
 def generate_module_dict(model: torch.nn.Module) -> Dict:
@@ -87,7 +108,7 @@ def fx_graph_module_generation(pytorch_model: torch.nn.Module,
     set_model(pytorch_model)
 
     try:
-        symbolic_traced = symbolic_trace(pytorch_model)
+        symbolic_traced = _trace_model(pytorch_model)
     except torch.fx.proxy.TraceError as e:
         Logger.critical(f'Error parsing model with torch.fx\n'
                         f'fx error: {e}')

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/core/pytorch/utils.py

@@ -15,7 +15,7 @@
 import torch
 from torch import Tensor
 import numpy as np
-from typing import Union, Sequence, Optional, List, Tuple
+from typing import Union, Optional, List, Tuple, Any
 
 from model_compression_toolkit.core.pytorch.constants import MAX_FLOAT16, MIN_FLOAT16
 from model_compression_toolkit.core.pytorch.pytorch_device_config import get_working_device
@@ -112,4 +112,4 @@ def clip_inf_values_float16(tensor: Tensor) -> Tensor:
     # Replace inf values with max float16 value
     tensor[inf_mask] = MAX_FLOAT16 * torch.sign(tensor[inf_mask])
 
-    return tensor
+    return tensor

{mct_nightly-2.3.0.20250408.522 → mct_nightly-2.3.0.20250410.526}/model_compression_toolkit/gptq/keras/quantization_facade.py

@@ -14,7 +14,7 @@
 # ==============================================================================
 import copy
 
-from typing import Callable, Tuple, Union
+from typing import Callable, Tuple, Union, Optional
 from packaging import version
 
 from model_compression_toolkit.core.common.visualization.tensorboard_writer import init_tensorboard_writer
@@ -158,7 +158,7 @@ if FOUND_TF:
                                                   target_resource_utilization: ResourceUtilization = None,
                                                   core_config: CoreConfig = CoreConfig(),
                                                   target_platform_capabilities: Union[TargetPlatformCapabilities, str]
-                                                  = DEFAULT_KERAS_TPC) -> Tuple[Model, UserInformation]:
+                                                  = DEFAULT_KERAS_TPC) -> Tuple[Model, Optional[UserInformation]]:
         """
         Quantize a trained Keras model using post-training quantization. The model is quantized using a
         symmetric constraint quantization thresholds (power of two).
@@ -230,6 +230,10 @@ if FOUND_TF:
             >>> quantized_model, quantization_info = mct.gptq.keras_gradient_post_training_quantization(model, repr_datagen, gptq_config, target_resource_utilization=ru, core_config=config)
 
         """
+
+        if core_config.debug_config.bypass:
+            return in_model, None
+
         KerasModelValidation(model=in_model,
                              fw_info=DEFAULT_KERAS_INFO).validate()