PyPI - mct-nightly - Versions diffs - 2.4.0.20250607.612__tar.gz → 2.4.0.20250609.615__tar.gz - Mend

Files changed (540) hide show

{mct_nightly-2.4.0.20250607.612 → mct_nightly-2.4.0.20250609.615}/model_compression_toolkit/core/common/fusion/fusing_info.py RENAMED Viewed

@@ -39,13 +39,13 @@ class FusingInfo:
     - 'fusing_patterns': The patterns to generate the fused operators from.
     - 'manual_fused_ops': List of sequence of node names to handle as fused ops (even if they are not part of the fusing patterns).
     - `fusing_data`: A dictionary mapping fused operation IDs to lists of nodes that belong to that operation.
-    - `node_to_fused_node_map`: A dictionary mapping each node name to the ID of the fused operation it belongs to.
+    - `node_name_to_fused_op_id`: A dictionary mapping each node name to the ID of the fused operation it belongs to.
     """
     fusing_patterns: List[list[any]] = None
     manual_fused_ops: List[List[str]] = None
     fusing_data: Dict[str, Tuple['BaseNode']] = field(default_factory=dict)
-    node_to_fused_node_map: Dict[str, str] = field(init=False, default_factory=dict)
+    node_name_to_fused_op_id: Dict[str, str] = field(init=False, default_factory=dict)
     fused_op_id_to_quant_config: Dict[str, OpQuantizationConfig] = field(default_factory=dict)
     def __post_init__(self):
@@ -64,10 +64,10 @@ class FusingInfo:
         """
         Init the node-to-fused-node mapping based on the initial fusing data.
         """
-        self.node_to_fused_node_map.clear()
+        self.node_name_to_fused_op_id.clear()
         for op_id, nodes in self.fusing_data.items():
             for node in nodes:
-                self.node_to_fused_node_map[node.name] = op_id
+                self.node_name_to_fused_op_id[node.name] = op_id
     def get_manual_nodes_to_fuse(self) -> List[List[str]]:
         """
@@ -115,7 +115,7 @@ class FusingInfo:
         self.fusing_data[op_id] = nodes
         # Update the mapping for these nodes
         for node in nodes:
-            self.node_to_fused_node_map[node.name] = op_id
+            self.node_name_to_fused_op_id[node.name] = op_id
         # Update the quantization config mapping for this operation
         if self.fusing_patterns is not None:
@@ -152,7 +152,7 @@ class FusingInfo:
             self._manual_fused_ops.remove(node_names)
         for node in nodes:
-            self.node_to_fused_node_map.pop(node.name, None)
+            self.node_name_to_fused_op_id.pop(node.name, None)
         del self.fusing_data[op_id]
         self.fused_op_id_to_quant_config.pop(op_id, None)
@@ -166,7 +166,7 @@ class FusingInfo:
         Returns:
             The name of the fused node containing this node, or None if not fused.
         """
-        return self.node_to_fused_node_map.get(node_name)
+        return self.node_name_to_fused_op_id.get(node_name)
     def get_node_to_fused_node_map(self) -> Dict[str, str]:
         """
@@ -175,7 +175,7 @@ class FusingInfo:
         Returns:
             A dictionary mapping each original node name to its fused node name.
         """
-        return self.node_to_fused_node_map.copy()
+        return self.node_name_to_fused_op_id.copy()
     def get_fusing_quantization_config_map(self) -> Dict[str, OpQuantizationConfig]:
         """
@@ -198,10 +198,12 @@ class FusingInfo:
         """
         return self.fusing_data.get(op_id)
-    def get_nodes_to_disable_activation_quantization(self) -> List['BaseNode']:
+    def get_inner_fln_nodes(self) -> List['BaseNode']:
         """
-        Returns a list of the nodes that their activation quantization is disabled due to fusing.
+        Returns a list of the nodes that are part but not the last node of an FLN.
         """
+        # TODO: the order of the nodes is not gurenteed when returned as dict from get_all_fused_operations -
+        #  then, removing the last one can cause issues
         return [node for nodes in self.get_all_fused_operations().values() for node in nodes[:-1]]
     def get_fused_op_quantization_config(self, op_id: str) -> OpQuantizationConfig:
@@ -228,6 +230,22 @@ class FusingInfo:
         """
         return any(node in nodes for nodes in self.fusing_data.values())
+    def is_quantized_node_in_fln(self, node: 'BaseNode') -> bool:
+        """
+        Check whether a node inside an FLN and should be quantized.
+        Args:
+            node (BaseNode): The node to check.
+        Returns:
+            bool: True if the node is in any fused operation and should be quantized.
+        """
+        if self.is_node_in_fused_op(node):
+            node_q_cfg = self.fused_op_id_to_quant_config[self.node_name_to_fused_op_id[node.name]]
+            return node_q_cfg is not None and node_q_cfg.enable_activation_quantization
+        return False
     def get_all_fused_operations(self) -> Dict[str, Tuple['BaseNode']]:
         """
         Retrieve fused information.
@@ -340,7 +358,7 @@ class FusingInfo:
             for op_id, nodes in self.fusing_data.items()
         )
         mapping_repr = ", ".join(
-            f"{node} -> {op_id}" for node, op_id in self.node_to_fused_node_map.items()
+            f"{node} -> {op_id}" for node, op_id in self.node_name_to_fused_op_id.items()
         )
         return (
             f"FusingInfo(\n"

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.4.0.20250607.612
+Version: 2.4.0.20250609.615
 Summary: A Model Compression Toolkit for neural networks
 Author-email: ssi-dnn-dev@sony.com
 Classifier: Programming Language :: Python :: 3

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.4.0.20250607.612
+Version: 2.4.0.20250609.615
 Summary: A Model Compression Toolkit for neural networks
 Author-email: ssi-dnn-dev@sony.com
 Classifier: Programming Language :: Python :: 3

@@ -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__ = "2.4.0.20250607.000612"
+__version__ = "2.4.0.20250609.000615"

@@ -908,7 +908,7 @@ class Graph(nx.MultiDiGraph, GraphSearches):
         Disable activation quantization for all nodes in fused operations,
         except for the last node in each fused group.
         """
-        nodes_to_disable = self.fusing_info.get_nodes_to_disable_activation_quantization()
+        nodes_to_disable = self.fusing_info.get_inner_fln_nodes()
         for node in nodes_to_disable:
             for qc in node.candidates_quantization_cfg:
                 qc.activation_quantization_cfg.quant_mode = ActivationQuantizationMode.FLN_QUANT

@@ -677,7 +677,7 @@ class ResourceUtilizationCalculator:
         elif target_criterion == TargetInclusionCriterion.AnyQuantizedNonFused:
             nodes = [n for n in nodes if n.is_activation_quantization_enabled() or n.is_quantization_preserving()]
             # remove fused nodes (due to SNC, where the non-linear is quantized, even though it should not be quantized)
-            nodes = [n for n in nodes if n not in self.graph.fusing_info.get_nodes_to_disable_activation_quantization()]
+            nodes = [n for n in nodes if n not in self.graph.fusing_info.get_inner_fln_nodes()]
         elif target_criterion == TargetInclusionCriterion.QNonConfigurable:
             nodes = [n for n in nodes if n.is_activation_quantization_enabled() and not n.has_configurable_activation()]
         elif target_criterion != TargetInclusionCriterion.Any:    # pragma: no cover

@@ -30,7 +30,8 @@ from model_compression_toolkit.core.common.collectors.statistics_collector impor
 def create_stats_collector_for_node(node: common.BaseNode,
-                                    fw_info: FrameworkInfo) -> BaseStatsCollector:
+                                    fw_info: FrameworkInfo,
+                                    quant_node_in_fln: bool) -> BaseStatsCollector:
     """
     Gets a node and a groups list and create and return a statistics collector for a node
     according to whether its statistics should be collected and the prior information we
@@ -44,7 +45,7 @@ def create_stats_collector_for_node(node: common.BaseNode,
         Statistics collector for statistics collection for the node.
     """
-    if node.is_activation_quantization_enabled() or node.is_fln_quantization():
+    if node.is_activation_quantization_enabled() or quant_node_in_fln:
         min_output = getattr(node.prior_info, 'min_output', None)
         max_output = getattr(node.prior_info, 'max_output', None)
         stats_collector = common.StatsCollector(out_channel_axis=fw_info.out_channel_axis_mapping.get(node.type),
@@ -160,7 +161,8 @@ class ModelCollector:
         # Assign statistics collectors to nodes
         for n in graph.get_topo_sorted_nodes():
-            sc = create_stats_collector_for_node(n, fw_info=fw_info)  # Get static collector for the node
+            quant_node_in_fln = n.is_fln_quantization() and graph.fusing_info.is_quantized_node_in_fln(n)
+            sc = create_stats_collector_for_node(n, fw_info=fw_info, quant_node_in_fln=quant_node_in_fln)  # Get static collector for the node
             # If we use bias correction, and the node has kernel weights to quantize, we need to make sure
             # its previous nodes' tensors are consistent with this node.
             kernel_attr = fw_info.get_kernel_op_attributes(n.type)[0]

@@ -78,9 +78,18 @@ if FOUND_ONNX:
             act_holder_list = [n for n, m in self.model.named_modules()
                                if isinstance(m, PytorchActivationQuantizationHolder) and
                                m.activation_holder_quantizer.num_bits > 8]
-            for act_holder in act_holder_list:  # pragma: no cover
-                delattr(self.model, act_holder)
-                setattr(self.model, act_holder, torch.nn.Identity())
+            for act_holder in act_holder_list: # pragma: no cover
+                obj = self.model
+                attrs = act_holder.split(".")
+                for a in attrs[:-1]:
+                    obj = getattr(obj, a)
+                if hasattr(obj, attrs[-1]):
+                    delattr(obj, attrs[-1])
+                    setattr(obj, attrs[-1], torch.nn.Identity())
+                else:
+                    Logger.info(f"During removal of activation quantization of a quantizer (with bits > 8) in ONNX FQ "
+                                f"export, deletion of activation holder '{act_holder}' failed — could not locate one or"
+                                f"more intermediate attributes in the path.")
             for layer in self.model.children():
                 self.is_layer_exportable_fn(layer)

mct-nightly 2.4.0.20250607.612__tar.gz → 2.4.0.20250609.615__tar.gz

mct-nightly 2.4.0.20250607.612tar.gz → 2.4.0.20250609.615tar.gz