mct-nightly 2.3.0.20250424.534__py3-none-any.whl → 2.3.0.20250426.524__py3-none-any.whl

{mct_nightly-2.3.0.20250424.534.dist-info → mct_nightly-2.3.0.20250426.524.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.3.0.20250424.534
+Version: 2.3.0.20250426.524
 Summary: A Model Compression Toolkit for neural networks
 Classifier: Programming Language :: Python :: 3
 Classifier: License :: OSI Approved :: Apache Software License

{mct_nightly-2.3.0.20250424.534.dist-info → mct_nightly-2.3.0.20250426.524.dist-info}/RECORD

@@ -1,5 +1,5 @@
-mct_nightly-2.3.0.20250424.534.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-model_compression_toolkit/__init__.py,sha256=HmFpViJmJPVcQg5km-gnodqRgdt3lc5eqANLwoWrMqM,1557
+mct_nightly-2.3.0.20250426.524.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
+model_compression_toolkit/__init__.py,sha256=u4NAaox4mWCLO1P2VUBGICSwqHFGboDDlwkRKJJADa8,1557
 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
@@ -112,7 +112,7 @@ model_compression_toolkit/core/common/quantization/quantization_fn_selection.py,
 model_compression_toolkit/core/common/quantization/quantization_params_fn_selection.py,sha256=7eG7dl1TcbdnHwgmvyjarxLs0o6Lw_9VAjXAm4rsiBk,3791
 model_compression_toolkit/core/common/quantization/quantize_graph_weights.py,sha256=N005MSvx8UypVpa7XrxNrB2G732n2wHj3RmLyjTgd3I,2728
 model_compression_toolkit/core/common/quantization/quantize_node.py,sha256=cdzGNWfT4MRogIU8ehs0tr3lVjnzAI-jeoS9b4TwVBo,2854
-model_compression_toolkit/core/common/quantization/set_node_quantization_config.py,sha256=_hhRb5eeFwbtPddu2xdLi7qK1RsxoR7UHUfjO0ICM3Q,30586
+model_compression_toolkit/core/common/quantization/set_node_quantization_config.py,sha256=76BjqZbeRz7zIblKxGp-AUkHOt-KZMlYjsXP8Ph-1jc,30943
 model_compression_toolkit/core/common/quantization/quantization_params_generation/__init__.py,sha256=eCDGwsWYLU6z7qbEVb4TozMW_nd5VEP_iCJ6PcvyEPw,1486
 model_compression_toolkit/core/common/quantization/quantization_params_generation/error_functions.py,sha256=_m-XkEMJMHf0gYwVIXAoHVjdRa2NXt_gYdwBlw76ZR8,24031
 model_compression_toolkit/core/common/quantization/quantization_params_generation/lut_kmeans_params.py,sha256=RL-PklAjGyC-26anSt8fU07a6pB_LBQFQy9o4e9giN0,8739
@@ -145,7 +145,7 @@ model_compression_toolkit/core/common/substitutions/linear_collapsing_substituti
 model_compression_toolkit/core/common/substitutions/remove_identity.py,sha256=TKU1TIU52UIkVnl0EZvWnDhLV9nIVZ4hqi-w1i4NXMk,2637
 model_compression_toolkit/core/common/substitutions/residual_collapsing.py,sha256=N82mso5j3EJQlKt9EMHjjEJ67FmdGQeCfN8U5grOFXo,4830
 model_compression_toolkit/core/common/substitutions/scale_equalization.py,sha256=p57u25qdW2pimxzGwgMXEBV4S-LzXuTVAlIM7830WfU,10966
-model_compression_toolkit/core/common/substitutions/shift_negative_activation.py,sha256=1utreR5CkJYwaJS1LOCQi_EYkedsnxXzyJCnJ-ZeqQ0,30057
+model_compression_toolkit/core/common/substitutions/shift_negative_activation.py,sha256=zCkdyZHEkbxkORmd071_XWajkpIhnDq9D6FyeE4TQjc,30057
 model_compression_toolkit/core/common/substitutions/softmax_shift.py,sha256=R-0ZqhYAuZLEFWHvB2UTPm52L6gWHGdRdEnwGxKSeGI,2625
 model_compression_toolkit/core/common/substitutions/virtual_activation_weights_composition.py,sha256=w43dRmaG96a8SNECgghxoFCTSoZ-vUb33dXGm2PbomE,4251
 model_compression_toolkit/core/common/substitutions/weights_activation_split.py,sha256=gt07lXRUvYunJKiwv_w20zfXhcplSW4oT2C1dqiNNXc,4719
@@ -233,7 +233,7 @@ model_compression_toolkit/core/pytorch/back2framework/factory_model_builder.py,s
 model_compression_toolkit/core/pytorch/back2framework/float_model_builder.py,sha256=tLrlUyYhxVKVjkad1ZAtbRra0HedB3iVfIkZ_dYnQ-4,3419
 model_compression_toolkit/core/pytorch/back2framework/instance_builder.py,sha256=BBHBfTqeWm7L3iDyPBpk0jxvj-rBg1QWI23imkjfIl0,1467
 model_compression_toolkit/core/pytorch/back2framework/mixed_precision_model_builder.py,sha256=K4L8FzJFM8_Ge2MHYkSqzCtoZe-ejEhVq8C1RgecyOc,14531
-model_compression_toolkit/core/pytorch/back2framework/pytorch_model_builder.py,sha256=WccaNiHK12IIimYu29E1oJkQHUdhPCBcIRutefTQ3Ag,19903
+model_compression_toolkit/core/pytorch/back2framework/pytorch_model_builder.py,sha256=9l5ZsEd4lwCSzT6VNItj5rw996BI_eH71x_uV-4gIRM,21101
 model_compression_toolkit/core/pytorch/back2framework/quantized_model_builder.py,sha256=qZNNOlNTTV4ZKPG3q5GDXkIVTPUEr8dvxAS_YiMORmg,3456
 model_compression_toolkit/core/pytorch/back2framework/quantization_wrapper/__init__.py,sha256=cco4TmeIDIh32nj9ZZXVkws4dd9F2UDrmjKzTN8G0V0,697
 model_compression_toolkit/core/pytorch/back2framework/quantization_wrapper/quantized_layer_wrapper.py,sha256=q2JDw10NKng50ee2i9faGzWZ-IydnR2aOMGSn9RoZmc,5773
@@ -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-2.3.0.20250424.534.dist-info/METADATA,sha256=wMqM0-nGTBa189h4xpdr-iY2-QUxlm1vVnXkB7ogmzU,25560
-mct_nightly-2.3.0.20250424.534.dist-info/WHEEL,sha256=SmOxYU7pzNKBqASvQJ7DjX3XGUF92lrGhMb3R6_iiqI,91
-mct_nightly-2.3.0.20250424.534.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
-mct_nightly-2.3.0.20250424.534.dist-info/RECORD,,
+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,,

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__ = "2.3.0.20250424.000534"
+__version__ = "2.3.0.20250426.000524"

model_compression_toolkit/core/common/quantization/set_node_quantization_config.py

@@ -119,11 +119,16 @@ def filter_node_qco_by_graph(node: BaseNode,
             _next_nodes.extend(graph.get_next_nodes(n))
         next_nodes.append(n)
 
-    if len(next_nodes):
-        next_nodes_qc_options = [_node.get_qco(fqc) for _node in next_nodes]
-        next_nodes_supported_input_bitwidth = min([max_input_activation_n_bits(op_cfg)
+    if len(next_nodes) == 0:
+        return _base_config, _node_qc_options
+    next_nodes_qc_options = [_node.get_qco(fqc) for _node in next_nodes]
+    all_next_nodes_supported_input_bitwidth = [max_input_activation_n_bits(op_cfg)
                                                    for qc_opts in next_nodes_qc_options
-                                                   for op_cfg in qc_opts.quantization_configurations])
+                                                   for op_cfg in qc_opts.quantization_configurations
+                                               if op_cfg.enable_activation_quantization or op_cfg.quantization_preserving
+                                               ]
+    if len(all_next_nodes_supported_input_bitwidth):
+        next_nodes_supported_input_bitwidth = min(all_next_nodes_supported_input_bitwidth)
 
         # Filter node's QC options that match next nodes input bit-width.
         _node_qc_options = [_option for _option in _node_qc_options
@@ -205,7 +210,7 @@ def set_quantization_configs_to_node(node: BaseNode,
                 # Preserving the quantization of more than 1 previous node is ambiguous, so disable it.
                 Logger.info(f"Disabling Quantization-Preserving for node {node.name} because it has more than 1 input activations.")
                 candidate_qc.activation_quantization_cfg.quant_mode = ActivationQuantizationMode.NO_QUANT
-            elif not prev_nodes[0].is_quantization_preserving() or not prev_nodes[0].is_activation_quantization_enabled():
+            elif not prev_nodes[0].is_quantization_preserving() and not prev_nodes[0].is_activation_quantization_enabled():
                 # Preserving the quantization of an unquantized node isn't possible, so disable it.
                 Logger.info(f"Disabling Quantization-Preserving for node {node.name} because previous node activation quantization is disabled.")
                 candidate_qc.activation_quantization_cfg.quant_mode = ActivationQuantizationMode.NO_QUANT

model_compression_toolkit/core/common/substitutions/shift_negative_activation.py

@@ -343,6 +343,13 @@ def shift_negative_function(graph: Graph,
     graph.set_out_stats_collector_to_node(add_node, add_node_stats_collector)
     graph.shift_stats_collector(add_node, np.array(shift_value))
 
+    set_quantization_configs_to_node(fw_info=fw_info,
+                                     node=add_node,
+                                     graph=graph,
+                                     quant_config=core_config.quantization_config,
+                                     fqc=graph.fqc,
+                                     mixed_precision_enable=core_config.is_mixed_precision_enabled)
+
     if padding is not None:
         pad_node = create_pad_node(op2d_node.name,
                                    add_node.name,
@@ -373,13 +380,6 @@ def shift_negative_function(graph: Graph,
 
         op2d_node.input_shape = pad_node.output_shape
 
-    set_quantization_configs_to_node(fw_info=fw_info,
-                                     node=add_node,
-                                     graph=graph,
-                                     quant_config=core_config.quantization_config,
-                                     fqc=graph.fqc,
-                                     mixed_precision_enable=core_config.is_mixed_precision_enabled)
-
     original_non_linear_activation_nbits = non_linear_node_cfg_candidate.activation_n_bits
     # The non-linear node's output should be float, so we approximate it by using 16bits quantization.
     for candidate_qc in non_linear_node.candidates_quantization_cfg:

model_compression_toolkit/core/pytorch/back2framework/pytorch_model_builder.py

@@ -234,7 +234,9 @@ class PytorchModel(torch.nn.Module):
         self.wrapper = wrapper
         self.get_activation_quantizer_holder = get_activation_quantizer_holder_fn
         self.reuse_groups = {}
-        self._add_modules()
+        self._reused_nodes = []
+
+        self._add_all_modules()
 
     # todo: Move to parent class BaseModelBuilder
     @property
@@ -286,17 +288,37 @@ class PytorchModel(torch.nn.Module):
                 node_op = self.wrapper(node, node_builder(node))
         return node_op
 
-    def _add_modules(self):
+    def _add_all_modules(self):
+        """
+        Build and add the modules and functional nodes from node_sort list as attributes to PytorchModel.
+        To assure all required nodes for the reused nodes are already initialized, adds none-reused nodes first,
+        then adds the reused nodes.
+        """
+        self._add_modules(reused_nodes_only=False)  # add none-reused nodes
+        self._add_modules(reused_nodes_only=True)  # add reused nodes
+
+    def _add_modules(self, reused_nodes_only=False):
         """
         Build and add the modules and functional nodes from node_sort list as attributes to PytorchModel
+        Args:
+            reused_nodes_only: whether to go over the reuse nodes list or not.
+                   In case reuse_nodes_only is False - will go over all nodes, and add reused nodes to self._reused_nodes
+                   In case reuse_nodes_only is True - will go over self._reused_nodes only.
+
         """
-        for node in self.node_sort:
-            if node.reuse:
-                # If the node is reused, retrieve the original module
+        nodes = self._reused_nodes if reused_nodes_only else self.node_sort
+        for node in nodes:
+            if node.reuse and reused_nodes_only:
                 if node.reuse_group not in self.reuse_groups:
-                    Logger.critical(f"Reuse group {node.reuse_group} not found for node {node.name}")
+                    raise Exception(f"Reuse group {node.reuse_group} not found for node {node.name}. "
+                                    f"Make sure you first call the method with reused_nodes_only=False")
+                else:
+                    node_op = self.reuse_groups[node.reuse_group]  # retrieve the original module
+
+            elif node.reuse:  # add node to reused list, and go over the list after all other nodes were created
+                self._reused_nodes.append(node)
+                continue
 
-                node_op = self.reuse_groups[node.reuse_group]
             else:
                 # If it's not reused, create a new module
                 node_op = self.wrap(node)