mct-nightly 2.3.0.20250426.524__py3-none-any.whl → 2.3.0.20250428.605__py3-none-any.whl

{mct_nightly-2.3.0.20250426.524.dist-info → mct_nightly-2.3.0.20250428.605.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.3.0.20250426.524
+Version: 2.3.0.20250428.605
 Summary: A Model Compression Toolkit for neural networks
 Classifier: Programming Language :: Python :: 3
 Classifier: License :: OSI Approved :: Apache Software License

{mct_nightly-2.3.0.20250426.524.dist-info → mct_nightly-2.3.0.20250428.605.dist-info}/RECORD

@@ -1,5 +1,5 @@
-mct_nightly-2.3.0.20250426.524.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-model_compression_toolkit/__init__.py,sha256=u4NAaox4mWCLO1P2VUBGICSwqHFGboDDlwkRKJJADa8,1557
+mct_nightly-2.3.0.20250428.605.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
+model_compression_toolkit/__init__.py,sha256=KTddrZxT3r5G_WJ3NWOixbYFMVcLk032Ii0ssUccyic,1557
 model_compression_toolkit/constants.py,sha256=iJ6vfTjC2oFIZWt8wvHoxEw5YJi3yl0Hd4q30_8q0Zc,3958
 model_compression_toolkit/defaultdict.py,sha256=LSc-sbZYXENMCw3U9F4GiXuv67IKpdn0Qm7Fr11jy-4,2277
 model_compression_toolkit/logger.py,sha256=L3q7tn3Uht0i_7phnlOWMR2Te2zvzrt2HOz9vYEInts,4529
@@ -47,7 +47,7 @@ model_compression_toolkit/core/common/graph/memory_graph/compute_graph_max_cut.p
 model_compression_toolkit/core/common/graph/memory_graph/cut.py,sha256=ZUGgn-vDA7unzc9UWhK2v_2i5nfdkSG1xOpgpDmziEo,2870
 model_compression_toolkit/core/common/graph/memory_graph/max_cut_astar.py,sha256=1TWLVAOlT8g8q_YyOdjm5cQfiSDZ5EHGQcb509Gnzjg,17895
 model_compression_toolkit/core/common/graph/memory_graph/memory_element.py,sha256=ISD2BvJWj5mB91jrFjG8VQb0oOoLBoita_thCZWzCPI,4238
-model_compression_toolkit/core/common/graph/memory_graph/memory_graph.py,sha256=FCzK4HmX4lWI4qGoGv94wpGv7o6_f5wPBfeBPMerZ18,7752
+model_compression_toolkit/core/common/graph/memory_graph/memory_graph.py,sha256=dME1M0hOjBdW5SqUbl1BPxvFRs-ZtDiF1dDGJuWJbl8,7711
 model_compression_toolkit/core/common/hessian/__init__.py,sha256=E7LK3K_1AwMCQokanNc1JODMwUKNOKmwXQiGQ7GO10I,1033
 model_compression_toolkit/core/common/hessian/hessian_info_service.py,sha256=8NDC_WLe3ZnY_v3e_Vz_lseF22lrbvhFmArihpeWfuI,14291
 model_compression_toolkit/core/common/hessian/hessian_info_utils.py,sha256=1axmN0tjJSo_7hUr2d2KMv4y1pBi19cqWSQpi4BbdsA,1458
@@ -69,13 +69,13 @@ model_compression_toolkit/core/common/mixed_precision/mixed_precision_candidates
 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=-hOMBucYn12ePyLd0b1KxniPOIRu4b53SwEzv0bWToI,4943
 model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_facade.py,sha256=d5-3j2e_rdcQOT7c4s0p7640i3nSetjJ6MgMhhMM7dc,6152
-model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_manager.py,sha256=a0lyySRmQ1vKikx5YvDMA4l1Eha-W5BCPYScvDlL_6c,37300
+model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_manager.py,sha256=658DBP0sY6DRqEbFcK1gX4EGQMeaBSFE5-7_Py6sioE,37718
 model_compression_toolkit/core/common/mixed_precision/sensitivity_evaluation.py,sha256=4bkM8pYKvk18cxHbx973Dz6qWrNT0MRm44cuk__qVaI,27297
 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=S1ChgxtUjzXJufNWyRbKoNdyNC6fGUjPeComDMx8ZCo,9479
 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=PKkhc5q8pEPnNLXwo3U56EOCfYnPXIvPs0LlCGZOoKU,4426
-model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_calculator.py,sha256=D2sNbTPMDsDyUE18NUpVJN27AgdwwhpdOJ8UMLmhdPA,40420
+model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_calculator.py,sha256=PXBuUUuYDmukjhgyrwEe71egpT_iu-LQt5SqddgkRHo,40793
 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=6Z6nQL9UH7B8dbcUR0cuCTEYFOKZAlvOb-SCk_cAZFA,6670
@@ -528,7 +528,7 @@ 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-2.3.0.20250426.524.dist-info/METADATA,sha256=iIA-rsewTrMqkLA5MRDESRmAWOwkY25inok3LAYZvWM,25560
-mct_nightly-2.3.0.20250426.524.dist-info/WHEEL,sha256=SmOxYU7pzNKBqASvQJ7DjX3XGUF92lrGhMb3R6_iiqI,91
-mct_nightly-2.3.0.20250426.524.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
-mct_nightly-2.3.0.20250426.524.dist-info/RECORD,,
+mct_nightly-2.3.0.20250428.605.dist-info/METADATA,sha256=qHKhtkD9E5Npa0vcNQc376dwsvBE6iUM0aiTV1S76qg,25560
+mct_nightly-2.3.0.20250428.605.dist-info/WHEEL,sha256=ck4Vq1_RXyvS4Jt6SI0Vz6fyVs4GWg7AINwpsaGEgPE,91
+mct_nightly-2.3.0.20250428.605.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
+mct_nightly-2.3.0.20250428.605.dist-info/RECORD,,

{mct_nightly-2.3.0.20250426.524.dist-info → mct_nightly-2.3.0.20250428.605.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (79.0.1)
+Generator: setuptools (80.0.0)
 Root-Is-Purelib: true
 Tag: py3-none-any
 

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__ = "2.3.0.20250426.000524"
+__version__ = "2.3.0.20250428.000605"

model_compression_toolkit/core/common/graph/memory_graph/memory_graph.py

@@ -38,9 +38,6 @@ class MemoryGraph(DirectedBipartiteGraph):
         Args:
             model_graph: A graph representation of a model.
         """
-
-        self.model_graph = model_graph
-
         nodes = list(model_graph.nodes)
         memory_tensors = []
         node_to_tensor = []

model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_manager.py

@@ -78,14 +78,18 @@ class MixedPrecisionSearchManager:
         self.mp_topo_configurable_nodes = self.mp_graph.get_configurable_sorted_nodes(fw_info)
 
         self.ru_targets = target_resource_utilization.get_restricted_targets()
-        self.ru_helper = MixedPrecisionRUHelper(self.mp_graph, fw_info, fw_impl)
+        self.ru_helper = MixedPrecisionRUHelper(self.original_graph, fw_info, fw_impl)
 
         self.min_ru_config: Dict[BaseNode, int] = self.mp_graph.get_min_candidates_config(fw_info)
         self.max_ru_config: Dict[BaseNode, int] = self.mp_graph.get_max_candidates_config(fw_info)
-        self.min_ru = self.ru_helper.compute_utilization(self.ru_targets, self.min_ru_config)
 
         self.config_reconstruction_helper = ConfigReconstructionHelper(virtual_graph=self.mp_graph,
                                                                        original_graph=self.original_graph)
+        if self.using_virtual_graph:
+            real_min_ru_config: Dict[BaseNode, int] = self.config_reconstruction_helper.reconstruct_config_from_virtual_graph(self.min_ru_config)
+            self.min_ru = self.ru_helper.compute_utilization(self.ru_targets, real_min_ru_config)
+        else:
+            self.min_ru = self.ru_helper.compute_utilization(self.ru_targets, self.min_ru_config)
 
     def search(self) -> Dict[BaseNode, int]:
         """
@@ -251,7 +255,8 @@ class MixedPrecisionSearchManager:
                 else:
                     cfg = self.min_ru_config.copy()
                     cfg[node] = candidate_idx
-                    candidate_rus = self.ru_helper.compute_utilization(self.ru_targets, cfg)
+                    real_cfg = self.config_reconstruction_helper.reconstruct_config_from_virtual_graph(cfg)
+                    candidate_rus = self.ru_helper.compute_utilization(self.ru_targets, real_cfg)
 
                 for target, ru in candidate_rus.items():
                     rus_per_candidate[target].append(ru)

model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_calculator.py

@@ -18,6 +18,8 @@ from copy import deepcopy
 from enum import Enum, auto
 from typing import Dict, NamedTuple, Optional, Tuple, List, Iterable, Union, Literal, Sequence
 
+from model_compression_toolkit.core.common.fusion.graph_fuser import GraphFuser
+
 from model_compression_toolkit.constants import FLOAT_BITWIDTH
 from model_compression_toolkit.core import FrameworkInfo
 from model_compression_toolkit.core.common import Graph, BaseNode
@@ -145,8 +147,14 @@ class ResourceUtilizationCalculator:
                 raise RuntimeError("Failed to calculate activation memory cuts for graph.")
             cuts = [cut for cut in cuts if cut.mem_elements.elements]
             # cache cuts nodes for future use, so do not filter by target
-            self._cuts = {cut: [self.graph.find_node_by_name(m.node_name)[0] for m in cut.mem_elements.elements]
-                          for cut in cuts}
+            self._cuts = {
+                cut: [
+                    node
+                    for m in cut.mem_elements.elements
+                    for node in (self.graph.fusing_info.get_fused_nodes(m.node_name) or (self.graph.find_node_by_name(m.node_name)[0],))
+                ]
+                for cut in cuts
+            }
         return self._cuts
 
     def compute_resource_utilization(self,
@@ -580,7 +588,9 @@ class ResourceUtilizationCalculator:
 
     def _compute_cuts(self):
         """ Compute activation cuts of the graph. """
-        memory_graph = MemoryGraph(deepcopy(self.graph))
+        # Compute memory graph on fused graph with fused nodes
+        graph = GraphFuser().apply_node_fusion(self.graph)
+        memory_graph = MemoryGraph(deepcopy(graph))
         _, _, cuts = compute_graph_max_cut(memory_graph)
         return cuts