mct-nightly 2.3.0.20250322.517__py3-none-any.whl → 2.3.0.20250324.606__py3-none-any.whl

{mct_nightly-2.3.0.20250322.517.dist-info → mct_nightly-2.3.0.20250324.606.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.3.0.20250322.517
+Version: 2.3.0.20250324.606
 Summary: A Model Compression Toolkit for neural networks
 Classifier: Programming Language :: Python :: 3
 Classifier: License :: OSI Approved :: Apache Software License

{mct_nightly-2.3.0.20250322.517.dist-info → mct_nightly-2.3.0.20250324.606.dist-info}/RECORD

@@ -1,5 +1,5 @@
-mct_nightly-2.3.0.20250322.517.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-model_compression_toolkit/__init__.py,sha256=zwR3rimyZAfQQXE_6fl9E7U2oLz9pyRmT9EYhEEzQaU,1557
+mct_nightly-2.3.0.20250324.606.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
+model_compression_toolkit/__init__.py,sha256=cmlxKyxfMU7e0tT0g_dE3z_TjW63WzFf9t_y-H5dZ80,1557
 model_compression_toolkit/constants.py,sha256=i_R6uXBfO1ph_X6DNJych2x59SUojfJbn7dNjs_mZnc,3846
 model_compression_toolkit/defaultdict.py,sha256=LSc-sbZYXENMCw3U9F4GiXuv67IKpdn0Qm7Fr11jy-4,2277
 model_compression_toolkit/logger.py,sha256=L3q7tn3Uht0i_7phnlOWMR2Te2zvzrt2HOz9vYEInts,4529
@@ -9,7 +9,7 @@ model_compression_toolkit/core/__init__.py,sha256=8a0wUNBKwTdJGDk_Ho6WQAXjGuCqQZ
 model_compression_toolkit/core/analyzer.py,sha256=X-2ZpkH1xdXnISnw1yJvXnvV-ssoUh-9LkLISSWNqiY,3691
 model_compression_toolkit/core/graph_prep_runner.py,sha256=CVTjBaci8F6EP3IKDnRMfxkP-Sv8qY8GpkGt6FyII2U,11376
 model_compression_toolkit/core/quantization_prep_runner.py,sha256=DPevqQ8brkdut8K5f5v9g5lbT3r1GSmhLAk3NkL40Fg,6593
-model_compression_toolkit/core/runner.py,sha256=qblr8WM6R5v4jip94kBeWHKsjc-FUOteVgMtunGf8lU,13716
+model_compression_toolkit/core/runner.py,sha256=WjZMVXc-OGBTnkiH0PRjNdJEM5pKQRPvLHXor5tjwjk,13096
 model_compression_toolkit/core/common/__init__.py,sha256=Wh127PbXcETZX_d1PQqZ71ETK3J9XO5A-HpadGUbj6o,1447
 model_compression_toolkit/core/common/base_substitutions.py,sha256=xDFSmVVs_iFSZfajytI0cuQaNRNcwHX3uqOoHgVUvxQ,1666
 model_compression_toolkit/core/common/framework_implementation.py,sha256=s3yiqnbWkwfnAB1sSal_KAuqVg27rLhAJ2O8LHUbSHE,22494
@@ -34,7 +34,7 @@ model_compression_toolkit/core/common/fusion/__init__.py,sha256=Rf1RcYmelmdZmBV5
 model_compression_toolkit/core/common/fusion/graph_fuser.py,sha256=b41_4rL_Adiza4vpWlmmqgvkpUmWVdfdx0nEIB0p2n8,6195
 model_compression_toolkit/core/common/fusion/layer_fusing.py,sha256=-2fnjyC9q2RPw9st6RxROW-gdtT2mSRz0QZ_Gz1KDz4,5579
 model_compression_toolkit/core/common/graph/__init__.py,sha256=Xr-Lt_qXMdrCnnOaUS_OJP_3iTTGfPCLf8_vSrQgCs0,773
-model_compression_toolkit/core/common/graph/base_graph.py,sha256=0zsiEldkV_wjDoTjaGtL8DOMGEv2yQqhajwEAnFgqR8,37819
+model_compression_toolkit/core/common/graph/base_graph.py,sha256=VhniLTiMqL7i1Vqg2UBQuFFTvw2cYeJayssUJwabp3E,38112
 model_compression_toolkit/core/common/graph/base_node.py,sha256=kZbmAMh5cPAwYzlY8KYa8w0ipL58yApB09-WXQ8plrE,33763
 model_compression_toolkit/core/common/graph/edge.py,sha256=buoSEUZwilWBK3WeBKpJ-GeDaUA1SDdOHxDpxU_bGpk,3784
 model_compression_toolkit/core/common/graph/functional_node.py,sha256=GH5wStmw8SoAj5IdT_-ItN1Meo_P5NUTt_5bgJC4fak,3935
@@ -67,18 +67,18 @@ model_compression_toolkit/core/common/mixed_precision/configurable_quantizer_uti
 model_compression_toolkit/core/common/mixed_precision/distance_weighting.py,sha256=-x8edUyudu1EAEM66AuXPtgayLpzbxoLNubfEbFM5kU,2867
 model_compression_toolkit/core/common/mixed_precision/mixed_precision_candidates_filter.py,sha256=6pLUEEIqRTVIlCYQC4JIvY55KAvuBHEX8uTOQ-1Ac4Q,3859
 model_compression_toolkit/core/common/mixed_precision/mixed_precision_quantization_config.py,sha256=r1t025_QHshyoop-PZvL7x6UuXaeplCCU3h4VNBhJHo,4309
-model_compression_toolkit/core/common/mixed_precision/mixed_precision_ru_helper.py,sha256=k7LjEmcvlkiV995DU7S1CrNOllu6qPZrhUUKXcZDIUQ,7538
-model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_facade.py,sha256=4YH9tsFPOn6rCcedfyocZhZwDLNX5kB1tebu0-nvhyA,7226
-model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_manager.py,sha256=ItBWNZYOf-Zzi8FaRv1y170wYRXYcR3pJysClOtH8qc,32525
+model_compression_toolkit/core/common/mixed_precision/mixed_precision_ru_helper.py,sha256=2Pp4hiYvGW2I9YhloDxQNT0sZRg3TDp9CXObloF8IFU,4971
+model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_facade.py,sha256=LddWtLileazCOvVSz-7j-GA4yskcGD3UHQGo7XUzSTE,5661
+model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_manager.py,sha256=NBzzhkVI407S9cIiw7t7nsP3MrkOdSnweKQdPBXb8to,38180
 model_compression_toolkit/core/common/mixed_precision/sensitivity_evaluation.py,sha256=gsigifJ-ykWNafF4t7UMEC_-nd6YPERAk1_z0kT-Y88,27172
 model_compression_toolkit/core/common/mixed_precision/set_layer_to_bitwidth.py,sha256=P8QtKgFXtt5b2RoubzI5OGlCfbEfZsAirjyrkFzK26A,2846
-model_compression_toolkit/core/common/mixed_precision/solution_refinement_procedure.py,sha256=MQZnBcpBDMd5y6rOunUtH3t41GQH0aBmxVB4muoxNfk,9477
+model_compression_toolkit/core/common/mixed_precision/solution_refinement_procedure.py,sha256=fk7PWiZ6Na5O_Z_dymk_UfDCTqW_X_4EROU7DZknQnc,9444
 model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/__init__.py,sha256=Rf1RcYmelmdZmBV5qOKvKWF575ofc06JFQSq83Jz99A,696
-model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization.py,sha256=T5yVr7lay-6QLuTDBZNI1Ufj02EMBWuY_yHjC8eHx5I,3998
-model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_calculator.py,sha256=Oj-tVGUyBXtTpxNFQVPja8fFcUOpi6B2PdpNKHkAlbc,39314
-model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_data.py,sha256=J7gqUGs4ITo4ufl84A5vACxm670LG6RhQyXkejfpbn8,8834
+model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization.py,sha256=PKkhc5q8pEPnNLXwo3U56EOCfYnPXIvPs0LlCGZOoKU,4426
+model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_calculator.py,sha256=xCYL36K0nK41VSsLcy52uDA7zVfoLxhubmOrtXbqw7s,39140
+model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_data.py,sha256=QQwtl08DiDxUOQGpYPnek_RlZjWm1Ky7tL2ESHXMK78,4050
 model_compression_toolkit/core/common/mixed_precision/search_methods/__init__.py,sha256=sw7LOPN1bM82o3SkMaklyH0jw-TLGK0-fl2Wq73rffI,697
-model_compression_toolkit/core/common/mixed_precision/search_methods/linear_programming.py,sha256=9Hh85pr0VL65umhf9mPnrrssJXwJPAsIkBwCZnfzjHY,17575
+model_compression_toolkit/core/common/mixed_precision/search_methods/linear_programming.py,sha256=rSSN5MhH5BO5b58d8pe2pY9wc5HbfescoUStfg-nWfk,7263
 model_compression_toolkit/core/common/network_editors/__init__.py,sha256=vZmu55bYqiaOQs3AjfwWDXHmuKZcLHt-wm7uR5fPEqg,1307
 model_compression_toolkit/core/common/network_editors/actions.py,sha256=nid0_j-Cn10xvmztT8yCKW_6uA7JEnom9SW9syx7wc0,19594
 model_compression_toolkit/core/common/network_editors/edit_network.py,sha256=dfgawi-nB0ocAJ0xcGn9E-Zv203oUnQLuMiXpX8vTgA,1748
@@ -526,7 +526,7 @@ model_compression_toolkit/xquant/pytorch/model_analyzer.py,sha256=b93o800yVB3Z-i
 model_compression_toolkit/xquant/pytorch/pytorch_report_utils.py,sha256=UVN_S9ULHBEldBpShCOt8-soT8YTQ5oE362y96qF_FA,3950
 model_compression_toolkit/xquant/pytorch/similarity_functions.py,sha256=CERxq5K8rqaiE-DlwhZBTUd9x69dtYJlkHOPLB54vm8,2354
 model_compression_toolkit/xquant/pytorch/tensorboard_utils.py,sha256=mkoEktLFFHtEKzzFRn_jCnxjhJolK12TZ5AQeDHzUO8,9767
-mct_nightly-2.3.0.20250322.517.dist-info/METADATA,sha256=aYBvUc3xM30BNiWYY4eIwtN5--e4Scbr-bHnYGl6lAk,27098
-mct_nightly-2.3.0.20250322.517.dist-info/WHEEL,sha256=1tXe9gY0PYatrMPMDd6jXqjfpz_B-Wqm32CPfRC58XU,91
-mct_nightly-2.3.0.20250322.517.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
-mct_nightly-2.3.0.20250322.517.dist-info/RECORD,,
+mct_nightly-2.3.0.20250324.606.dist-info/METADATA,sha256=xCfGClT5tOs76bxlrQX92fjTkmtPi-QvD6CBI8TM-EA,27098
+mct_nightly-2.3.0.20250324.606.dist-info/WHEEL,sha256=1tXe9gY0PYatrMPMDd6jXqjfpz_B-Wqm32CPfRC58XU,91
+mct_nightly-2.3.0.20250324.606.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
+mct_nightly-2.3.0.20250324.606.dist-info/RECORD,,

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__ = "2.3.0.20250322.000517"
+__version__ = "2.3.0.20250324.000606"

model_compression_toolkit/core/common/graph/base_graph.py

@@ -706,14 +706,24 @@ class Graph(nx.MultiDiGraph, GraphSearches):
         """
         self.fused_nodes.append(fusion)
 
-    def is_single_activation_cfg(self):
+    def has_any_configurable_activation(self) -> bool:
         """
-        Checks whether all nodes in the graph that have activation quantization are quantized with the same bit-width.
+        Checks whether any node in the graph has a configurable activation quantization.
 
-        Returns: True if all quantization config candidates of all nodes have the same activation quantization bit-width.
+        Returns:
+            Whether any node in the graph has a configurable activation quantization.
+        """
+        return any([n.has_configurable_activation() for n in self.nodes])
+
+    def has_any_configurable_weights(self):
+        """
+        Checks whether any node in the graph has any configurable weights quantization.
 
+        Returns:
+            Whether any node in the graph has any configurable weights quantization.
         """
-        return all([n.is_all_activation_candidates_equal() for n in self.nodes])
+
+        return any([n.has_any_configurable_weight() for n in self.nodes])
 
     def replace_node(self, node_to_replace: BaseNode, new_node: BaseNode):
         """

model_compression_toolkit/core/common/mixed_precision/mixed_precision_ru_helper.py

@@ -12,12 +12,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-from typing import List, Set, Dict, Optional, Tuple, Any
+from typing import List, Set, Dict, Tuple
 
 import numpy as np
 
 from model_compression_toolkit.core import FrameworkInfo
-from model_compression_toolkit.core.common import Graph, BaseNode
+from model_compression_toolkit.core.common import Graph
 from model_compression_toolkit.core.common.framework_implementation import FrameworkImplementation
 from model_compression_toolkit.core.common.mixed_precision.resource_utilization_tools.resource_utilization import \
     RUTarget
@@ -36,42 +36,46 @@ class MixedPrecisionRUHelper:
         self.fw_impl = fw_impl
         self.ru_calculator = ResourceUtilizationCalculator(graph, fw_impl, fw_info)
 
-    def compute_utilization(self, ru_targets: Set[RUTarget], mp_cfg: Optional[List[int]]) -> Dict[RUTarget, np.ndarray]:
+    def compute_utilization(self, ru_targets: Set[RUTarget], mp_cfg: List[int]) -> Dict[RUTarget, np.ndarray]:
         """
-        Compute utilization of requested targets for a specific configuration in the format expected by LP problem
-        formulation namely a vector of ru values for relevant memory elements (nodes or cuts) in a constant order
-        (between calls).
+        Compute utilization of requested targets for a specific configuration:
+          for weights and bops - total utilization,
+          for activations and total - utilization per cut.
 
         Args:
             ru_targets: resource utilization targets to compute.
             mp_cfg: a list of candidates indices for configurable layers.
 
         Returns:
-            Dict of the computed utilization per target.
+            Dict of the computed utilization per target, as 1d vector.
         """
-
-        ru = {}
-        act_qcs, w_qcs = self.get_quantization_candidates(mp_cfg) if mp_cfg else (None, None)
-        if RUTarget.WEIGHTS in ru_targets:
-            wu = self._weights_utilization(w_qcs)
-            ru[RUTarget.WEIGHTS] = np.array(list(wu.values()))
-
-        if RUTarget.ACTIVATION in ru_targets:
-            au = self._activation_utilization(act_qcs)
-            ru[RUTarget.ACTIVATION] = np.array(list(au.values()))
-
-        if RUTarget.BOPS in ru_targets:
-            ru[RUTarget.BOPS] = self._bops_utilization(act_qcs=act_qcs, w_qcs=w_qcs)
-
-        if RUTarget.TOTAL in ru_targets:
-            raise ValueError('Total target should be computed based on weights and activations targets.')
-
-        assert len(ru) == len(ru_targets), (f'Mismatch between the number of computed and requested metrics.'
-                                            f'Requested {ru_targets}')
-        return ru
+        act_qcs, w_qcs = self.get_quantization_candidates(mp_cfg)
+
+        ru, detailed_ru = self.ru_calculator.compute_resource_utilization(TargetInclusionCriterion.AnyQuantized,
+                                                                          BitwidthMode.QCustom,
+                                                                          act_qcs=act_qcs,
+                                                                          w_qcs=w_qcs,
+                                                                          ru_targets=ru_targets,
+                                                                          allow_unused_qcs=True,
+                                                                          return_detailed=True)
+
+        ru_dict = {k: np.array([v]) for k, v in ru.get_resource_utilization_dict(restricted_only=True).items()}
+        # For activation and total we need utilization per cut, as different mp configurations might result in
+        # different cuts to be maximal.
+        for target in [RUTarget.ACTIVATION, RUTarget.TOTAL]:
+            if target in ru_dict:
+                ru_dict[target] = np.array(list(detailed_ru[target].values()))
+
+        assert all(v.ndim == 1 for v in ru_dict.values())
+        if RUTarget.ACTIVATION in ru_targets and RUTarget.TOTAL in ru_targets:
+            assert ru_dict[RUTarget.ACTIVATION].shape == ru_dict[RUTarget.TOTAL].shape
+
+        assert len(ru_dict) == len(ru_targets), (f'Mismatch between the number of computed and requested metrics.'
+                                                 f'Requested {ru_targets}')
+        return ru_dict
 
     def get_quantization_candidates(self, mp_cfg) \
-            -> Tuple[Dict[BaseNode, NodeActivationQuantizationConfig], Dict[BaseNode, NodeWeightsQuantizationConfig]]:
+            -> Tuple[Dict[str, NodeActivationQuantizationConfig], Dict[str, NodeWeightsQuantizationConfig]]:
         """
         Retrieve quantization candidates objects for weights and activations from the configuration list.
 
@@ -87,71 +91,3 @@ class MixedPrecisionRUHelper:
         act_qcs = {n.name: cfg.activation_quantization_cfg for n, cfg in node_qcs.items()}
         w_qcs = {n.name: cfg.weights_quantization_cfg for n, cfg in node_qcs.items()}
         return act_qcs, w_qcs
-
-    def _weights_utilization(self, w_qcs: Optional[Dict[BaseNode, NodeWeightsQuantizationConfig]]) -> Dict[BaseNode, float]:
-        """
-        Compute weights utilization for configurable weights if configuration is passed,
-        or for non-configurable nodes otherwise.
-
-        Args:
-            w_qcs: nodes quantization configuration to compute, or None.
-
-        Returns:
-            Weight utilization per node.
-        """
-        if w_qcs:
-            target_criterion = TargetInclusionCriterion.QConfigurable
-            bitwidth_mode = BitwidthMode.QCustom
-        else:
-            target_criterion = TargetInclusionCriterion.QNonConfigurable
-            bitwidth_mode = BitwidthMode.QDefaultSP
-
-        _, nodes_util, _ = self.ru_calculator.compute_weights_utilization(target_criterion=target_criterion,
-                                                                          bitwidth_mode=bitwidth_mode,
-                                                                          w_qcs=w_qcs)
-        nodes_util = {n: u.bytes for n, u in nodes_util.items()}
-        return nodes_util
-
-    def _activation_utilization(self, act_qcs: Optional[Dict[BaseNode, NodeActivationQuantizationConfig]]) \
-            -> Optional[Dict[Any, float]]:
-        """
-        Compute activation utilization using MaxCut for all quantized nodes if configuration is passed.
-
-        Args:
-            act_qcs: nodes activation configuration or None.
-
-        Returns:
-            Activation utilization per cut, or empty dict if no configuration was passed.
-        """
-        # Maxcut activation utilization is computed for all quantized nodes, so non-configurable memory is already
-        # covered by the computation of configurable activations.
-        if not act_qcs:
-            return {}
-
-        _, cuts_util, *_ = self.ru_calculator.compute_activation_utilization_by_cut(
-            TargetInclusionCriterion.AnyQuantized, bitwidth_mode=BitwidthMode.QCustom, act_qcs=act_qcs)
-        cuts_util = {c: u.bytes for c, u in cuts_util.items()}
-        return cuts_util
-
-    def _bops_utilization(self,
-                          act_qcs: Optional[Dict[BaseNode, NodeActivationQuantizationConfig]],
-                          w_qcs: Optional[Dict[BaseNode, NodeWeightsQuantizationConfig]]) -> np.ndarray:
-        """
-        Computes a resource utilization vector with the respective bit-operations (BOPS) count
-        according to the given mixed-precision configuration.
-
-        Args:
-            act_qcs: nodes activation configuration or None.
-            w_qcs: nodes quantization configuration to compute, or None.
-              Either both are provided, or both are None.
-
-        Returns:
-            A vector of node's BOPS count.
-        """
-        assert [act_qcs, w_qcs].count(None) in [0, 2], 'act_qcs and w_qcs should both be provided or both be None.'
-        if act_qcs is None:
-            return np.array([])
-
-        _, detailed_bops = self.ru_calculator.compute_bops(TargetInclusionCriterion.Any, BitwidthMode.QCustom,
-                                                           act_qcs=act_qcs, w_qcs=w_qcs)
-        return np.array(list(detailed_bops.values()))

model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_facade.py

@@ -13,37 +13,27 @@
 # limitations under the License.
 # ==============================================================================
 
-import copy
 from enum import Enum
-import numpy as np
-from typing import List, Callable, Dict
+from typing import List, Callable
 
 from model_compression_toolkit.core import MixedPrecisionQuantizationConfig
 from model_compression_toolkit.core.common import Graph
-from model_compression_toolkit.core.common.hessian import HessianInfoService
-from model_compression_toolkit.core.common.mixed_precision.resource_utilization_tools.resource_utilization import ResourceUtilization, RUTarget
 from model_compression_toolkit.core.common.framework_implementation import FrameworkImplementation
-from model_compression_toolkit.core.common.mixed_precision.mixed_precision_search_manager import MixedPrecisionSearchManager
-from model_compression_toolkit.core.common.mixed_precision.search_methods.linear_programming import \
-    mp_integer_programming_search
 from model_compression_toolkit.core.common.framework_info import FrameworkInfo
+from model_compression_toolkit.core.common.hessian import HessianInfoService
+from model_compression_toolkit.core.common.mixed_precision.mixed_precision_search_manager import \
+    MixedPrecisionSearchManager
+from model_compression_toolkit.core.common.mixed_precision.resource_utilization_tools.resource_utilization import \
+    ResourceUtilization
 from model_compression_toolkit.core.common.mixed_precision.solution_refinement_procedure import \
     greedy_solution_refinement_procedure
-from model_compression_toolkit.core.common.substitutions.apply_substitutions import substitute
-from model_compression_toolkit.logger import Logger
 
 
 class BitWidthSearchMethod(Enum):
-    # When adding a new search_methods MP configuration method, these enum and factory dictionary
-    # should be updated with it's kind and a search_method implementation.
     INTEGER_PROGRAMMING = 0
 
 
-search_methods = {
-    BitWidthSearchMethod.INTEGER_PROGRAMMING: mp_integer_programming_search}
-
-
-def search_bit_width(graph_to_search_cfg: Graph,
+def search_bit_width(graph: Graph,
                      fw_info: FrameworkInfo,
                      fw_impl: FrameworkImplementation,
                      target_resource_utilization: ResourceUtilization,
@@ -60,7 +50,7 @@ def search_bit_width(graph_to_search_cfg: Graph,
     target_resource_utilization have to be passed. If it was not passed, the facade is not supposed to get here by now.
 
     Args:
-        graph_to_search_cfg: Graph to search a MP configuration for.
+        graph: Graph to search a MP configuration for.
         fw_info: FrameworkInfo object about the specific framework (e.g., attributes of different layers' weights to quantize).
         fw_impl: FrameworkImplementation object with specific framework methods implementation.
         target_resource_utilization: Target Resource Utilization to bound our feasible solution space s.t the configuration does not violate it.
@@ -75,17 +65,7 @@ def search_bit_width(graph_to_search_cfg: Graph,
         bit-width index on the node).
 
     """
-
-    # target_resource_utilization have to be passed. If it was not passed, the facade is not supposed to get here by now.
-    if target_resource_utilization is None:
-        Logger.critical("Target ResourceUtilization is required for the bit-width search method's configuration.")  # pragma: no cover
-
-    # Set graph for MP search
-    graph = copy.deepcopy(graph_to_search_cfg)  # Copy graph before searching
-    if target_resource_utilization.bops_restricted():
-        # TODO: we only need the virtual graph is both activations and weights are configurable
-        # Since Bit-operations count target resource utilization is set, we need to reconstruct the graph for the MP search
-        graph = substitute(graph, fw_impl.get_substitutions_virtual_weights_activation_coupling())
+    assert target_resource_utilization.is_any_restricted()
 
     # If we only run weights compression with MP than no need to consider activation quantization when computing the
     # MP metric (it adds noise to the computation)
@@ -93,33 +73,28 @@ def search_bit_width(graph_to_search_cfg: Graph,
     weight_only_restricted = tru.weight_restricted() and not (tru.activation_restricted() or
                                                               tru.total_mem_restricted() or
                                                               tru.bops_restricted())
-    disable_activation_for_metric = weight_only_restricted or graph_to_search_cfg.is_single_activation_cfg()
+    disable_activation_for_metric = weight_only_restricted or not graph.has_any_configurable_activation()
 
     # Set Sensitivity Evaluator for MP search. It should always work with the original MP graph,
     # even if a virtual graph was created (and is used only for BOPS utilization computation purposes)
     se = fw_impl.get_sensitivity_evaluator(
-        graph_to_search_cfg,
+        graph,
         mp_config,
         representative_data_gen=representative_data_gen,
         fw_info=fw_info,
         disable_activation_for_metric=disable_activation_for_metric,
         hessian_info_service=hessian_info_service)
 
-    # Instantiate a manager object
+    if search_method != BitWidthSearchMethod.INTEGER_PROGRAMMING:
+        raise NotImplementedError()
+
+    # Search manager and LP are highly coupled, so LP search method was moved inside search manager.
     search_manager = MixedPrecisionSearchManager(graph,
                                                  fw_info,
                                                  fw_impl,
                                                  se,
-                                                 target_resource_utilization,
-                                                 original_graph=graph_to_search_cfg)
-
-    if search_method not in search_methods:
-        raise NotImplementedError()  # pragma: no cover
-
-    search_method_fn = search_methods[search_method]
-    # Search for the desired mixed-precision configuration
-    result_bit_cfg = search_method_fn(search_manager,
-                                      target_resource_utilization)
+                                                 target_resource_utilization)
+    result_bit_cfg = search_manager.search()
 
     if mp_config.refine_mp_solution:
         result_bit_cfg = greedy_solution_refinement_procedure(result_bit_cfg, search_manager, target_resource_utilization)