PyPI - mct-nightly - Versions diffs - 2.3.0.20250424.534__tar.gz → 2.3.0.20250426.524__tar.gz - Mend

@@ -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

@@ -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

@@ -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"

@@ -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

@@ -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)

@@ -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:

mct-nightly 2.3.0.20250424.534__tar.gz → 2.3.0.20250426.524__tar.gz

mct-nightly 2.3.0.20250424.534tar.gz → 2.3.0.20250426.524tar.gz