PyPI - mct-nightly - Versions diffs - 2.4.0.20250626.617__tar.gz → 2.4.0.20250628.555__tar.gz - Mend

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.4.0.20250626.617
+Version: 2.4.0.20250628.555
 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: 2.4.0.20250626.617
+Version: 2.4.0.20250628.555
 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__ = "2.4.0.20250626.000617"
+__version__ = "2.4.0.20250628.000555"

@@ -32,13 +32,13 @@ from model_compression_toolkit.core.common.collectors.statistics_collector impor
 from model_compression_toolkit.core.common.collectors.statistics_collector import scale_statistics, shift_statistics
 from model_compression_toolkit.core.common.pruning.pruning_section import PruningSection
 from model_compression_toolkit.core.common.user_info import UserInformation
-from model_compression_toolkit.core.common.quantization.node_quantization_config import ActivationQuantizationMode
+from model_compression_toolkit.core.common.quantization.node_quantization_config import \
+    NodeActivationQuantizationConfig, ActivationQuantizationMode
 from model_compression_toolkit.logger import Logger
 from model_compression_toolkit.target_platform_capabilities.targetplatform2framework import LayerFilterParams
 from model_compression_toolkit.target_platform_capabilities.targetplatform2framework.framework_quantization_capabilities import \
     FrameworkQuantizationCapabilities
 def validate_graph_after_change(method: Callable) -> Callable:
     """
     Decorator for graph-mutating methods. After the decorated method executes,
@@ -876,15 +876,32 @@ class Graph(nx.MultiDiGraph, GraphSearches):
         return intermediate_nodes, next_node
-    def disable_fused_nodes_activation_quantization(self):
+    def override_fused_node_activation_quantization_candidates(self):
         """
-        Disable activation quantization for all nodes in fused operations,
+        Override fused node activation quantization candidates for all nodes in fused operations,
         except for the last node in each fused group.
-        """
-        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
+        Update the value of quantization_config with the value of op_quaitization_cfg from FusingInfo.
+        """
+        from model_compression_toolkit.core.common.quantization.candidate_node_quantization_config import CandidateNodeQuantizationConfig
+        nodes_in_fln = self.fusing_info.get_inner_fln_nodes()
+        for node in nodes_in_fln:
+            fused_node_op_id = self.fusing_info.get_fused_op_id_for_node(node.name)
+            fusiong_op_quaitization_cfg = self.fusing_info.get_fused_op_quantization_config(fused_node_op_id)
+            org_candidate = node.candidates_quantization_cfg[0]
+            if fusiong_op_quaitization_cfg is not None and fusiong_op_quaitization_cfg.enable_activation_quantization:
+                # Set ActivationQuantizationMode to FLN_QUANT and update the value of quantization_config
+                activation_quantization_cfg = NodeActivationQuantizationConfig(qc=org_candidate,
+                                                                               op_cfg=fusiong_op_quaitization_cfg,
+                                                                               activation_quantization_fn=org_candidate.activation_quantization_cfg.activation_quantization_fn,
+                                                                               activation_quantization_params_fn=org_candidate.activation_quantization_cfg.activation_quantization_params_fn)
+                activation_quantization_cfg.quant_mode = ActivationQuantizationMode.FLN_QUANT
+                for qc in node.candidates_quantization_cfg:
+                    qc.activation_quantization_cfg = activation_quantization_cfg
+            else:
+                # Set ActivationQuantizationMode to FLN_NO_QUANT
+                for qc in node.candidates_quantization_cfg:
+                    qc.activation_quantization_cfg.quant_mode = ActivationQuantizationMode.FLN_NO_QUANT
     def validate(self):
         """
@@ -908,4 +925,4 @@ class Graph(nx.MultiDiGraph, GraphSearches):
         """
         Wrap networkx functions (that modifies the graph) with our validate decorator.
         """
-        return super().remove_edge(*args, **kwargs)
+        return super().remove_edge(*args, **kwargs)

@@ -216,19 +216,31 @@ class BaseNode:
         Returns: Whether node activation quantization is enabled or not.
         """
         return self._is_single_quant_mode(ActivationQuantizationMode.QUANT)
-    def is_fln_quantization(self) -> bool:
+    def is_fln_no_quantization(self) -> bool:
         """
-        Returns: Whether the node's activation quantization is FLN
+        Returns: Whether node is FLN no quantization.
         """
-        return self._is_single_quant_mode(ActivationQuantizationMode.FLN_QUANT)
+        return self._is_single_quant_mode(ActivationQuantizationMode.FLN_NO_QUANT)
     def is_quantization_preserving(self) -> bool:
         """
         Returns: Whether node activation quantization information is preserved from its inputs.
         """
         return self._is_single_quant_mode(ActivationQuantizationMode.PRESERVE_QUANT)
+    def is_no_quantization(self) -> bool:
+        """
+        Returns: Whether node is no quantization.
+        """
+        return self._is_single_quant_mode(ActivationQuantizationMode.NO_QUANT)
+    def is_fln_quantization(self) -> bool:
+        """
+        Returns: Whether the node's activation quantization is FLN
+        """
+        return self._is_single_quant_mode(ActivationQuantizationMode.FLN_QUANT)
     def is_weights_quantization_enabled(self, attr_name: str) -> bool:
         """
         Checks whether a node's weights attribute quantization is enabled.

@@ -21,7 +21,6 @@ from model_compression_toolkit.constants import FLOAT_BITWIDTH
 from model_compression_toolkit.core.common.quantization.candidate_node_quantization_config import \
     CandidateNodeQuantizationConfig
 def filter_nodes_candidates(graph: Graph):
     """
     Filters the graph's nodes candidates configuration list.
@@ -87,7 +86,7 @@ def filter_node_candidates(node: BaseNode) -> List[CandidateNodeQuantizationConf
     filtered_candidates = copy.deepcopy(node.candidates_quantization_cfg)
     final_candidates = copy.deepcopy(node.candidates_quantization_cfg)
-    if (node.kernel_attr is None or not node.is_weights_quantization_enabled(node.kernel_attr)) and not node.is_activation_quantization_enabled():
+    if (node.kernel_attr is None or not node.is_weights_quantization_enabled(node.kernel_attr)) and node.is_no_quantization():
         # If activation quantization is disabled and the node doesn't have a kernel or doesn't quantize the kernel,
         # but for some reason the node has multiple candidates then replace it with a single dummy candidate with
         # default bit-width values.
@@ -102,9 +101,10 @@ def filter_node_candidates(node: BaseNode) -> List[CandidateNodeQuantizationConf
         final_candidates = [single_dummy_candidate]
-    elif not node.is_activation_quantization_enabled():
+    elif node.is_no_quantization():
         # Remove candidates that have duplicated weights candidates for node with disabled activation quantization.
         # Replacing the activation n_bits in the remained configurations with default value to prevent confusion.
+        # Set the config of the non-quantized FLN node to POWER_OF_TWO.
         seen_candidates = set()
         filtered_candidates = [candidate for candidate in filtered_candidates if
                                candidate.weights_quantization_cfg not in seen_candidates
@@ -116,6 +116,14 @@ def filter_node_candidates(node: BaseNode) -> List[CandidateNodeQuantizationConf
         final_candidates = _filter_bit_method_dups(filtered_candidates, node.kernel_attr)
+    elif node.is_fln_no_quantization() or node.is_fln_quantization():
+        # Remove candidates that have duplicated weights candidates for node with disabled activation quantization.
+        seen_candidates = set()
+        filtered_candidates = [candidate for candidate in filtered_candidates if
+                               candidate.weights_quantization_cfg not in seen_candidates
+                               and not seen_candidates.add(candidate.weights_quantization_cfg)]
+        final_candidates = _filter_bit_method_dups(filtered_candidates, node.kernel_attr)
     elif node.kernel_attr is None or not node.is_weights_quantization_enabled(node.kernel_attr):
         # TODO:
         #  To allow MP on positional weights we need to modify this to consider all weights not only kernel.

@@ -47,7 +47,7 @@ class ActivationQuantizationMode(Enum):
     FLN_QUANT = auto()
     PRESERVE_QUANT = auto()
     NO_QUANT = auto()
+    FLN_NO_QUANT = auto()
 class BaseNodeQuantizationConfig(object):
     """

@@ -155,7 +155,7 @@ def get_finalized_graph(initial_graph: Graph,
     ######################################
     fusing_info = FusingInfoGenerator(fqc.get_fusing_patterns()).generate_fusing_info(transformed_graph)
     transformed_graph.fusing_info = fusing_info
-    transformed_graph.disable_fused_nodes_activation_quantization()
+    transformed_graph.override_fused_node_activation_quantization_candidates()
     ######################################
     # Channel equalization

mct-nightly 2.4.0.20250626.617__tar.gz → 2.4.0.20250628.555__tar.gz

mct-nightly 2.4.0.20250626.617tar.gz → 2.4.0.20250628.555tar.gz