mct-nightly 2.3.0.20250604.611__tar.gz → 2.4.0.20250605.606__tar.gz

{mct_nightly-2.3.0.20250604.611 → mct_nightly-2.4.0.20250605.606}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.3.0.20250604.611
+Version: 2.4.0.20250605.606
 Summary: A Model Compression Toolkit for neural networks
 Author-email: ssi-dnn-dev@sony.com
 Classifier: Programming Language :: Python :: 3

{mct_nightly-2.3.0.20250604.611 → mct_nightly-2.4.0.20250605.606}/mct_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.3.0.20250604.611
+Version: 2.4.0.20250605.606
 Summary: A Model Compression Toolkit for neural networks
 Author-email: ssi-dnn-dev@sony.com
 Classifier: Programming Language :: Python :: 3

{mct_nightly-2.3.0.20250604.611 → mct_nightly-2.4.0.20250605.606}/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__ = "2.3.0.20250604.000611"
+__version__ = "2.4.0.20250605.000606"

{mct_nightly-2.3.0.20250604.611 → mct_nightly-2.4.0.20250605.606}/model_compression_toolkit/core/common/graph/base_node.py

@@ -134,6 +134,12 @@ class BaseNode:
         """
         return self._is_single_quant_mode(ActivationQuantizationMode.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_quantization_preserving(self) -> bool:
         """
         Returns: Whether node activation quantization information is preserved from its inputs.

{mct_nightly-2.3.0.20250604.611 → mct_nightly-2.4.0.20250605.606}/model_compression_toolkit/core/common/mixed_precision/mixed_precision_quantization_config.py

@@ -22,12 +22,17 @@ from model_compression_toolkit.constants import MP_DEFAULT_NUM_SAMPLES, ACT_HESS
 
 class MpDistanceWeighting(Enum):
     """
+
     Defines interest points distances weighting methods.
 
     AVG - take the average distance over all interest points.
+
     LAST_LAYER - take only the distance of the last interest point.
+
     EXP - weighted average with weights based on exponent of negative distances between activations of the quantized and the float models.
+
     HESSIAN - weighted average with Hessians as weights.
+
     """
     AVG = auto()
     LAST_LAYER = auto()
@@ -37,9 +42,13 @@ class MpDistanceWeighting(Enum):
 
 class MpMetricNormalization(Enum):
     """
+
     MAXBIT: normalize sensitivity metrics of layer candidates by max-bitwidth candidate (of that layer).
+
     MINBIT: normalize sensitivity metrics of layer candidates by min-bitwidth candidate (of that layer).
+
     NONE: no normalization.
+
     """
     MAXBIT = 'MAXBIT'
     MINBIT = 'MINBIT'

{mct_nightly-2.3.0.20250604.611 → mct_nightly-2.4.0.20250605.606}/model_compression_toolkit/core/common/model_collector.py

@@ -44,7 +44,7 @@ def create_stats_collector_for_node(node: common.BaseNode,
         Statistics collector for statistics collection for the node.
     """
 
-    if node.is_activation_quantization_enabled():
+    if node.is_activation_quantization_enabled() or node.is_fln_quantization():
         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),

{mct_nightly-2.3.0.20250604.611 → mct_nightly-2.4.0.20250605.606}/model_compression_toolkit/exporter/model_exporter/pytorch/base_pytorch_exporter.py

@@ -18,7 +18,7 @@ from typing import Callable
 import torch.nn
 
 from mct_quantizers import PytorchQuantizationWrapper
-from mct_quantizers.common.constants import LAYER, WEIGHTS_QUANTIZERS
+from mct_quantizers.common.constants import LAYER, WEIGHTS_QUANTIZERS, QUANTIZED_POSITIONAL_WEIGHT
 from model_compression_toolkit.logger import Logger
 from model_compression_toolkit.exporter.model_exporter.fw_agonstic.exporter import Exporter
 
@@ -73,8 +73,27 @@ def _set_quantized_weights_in_wrapper(layer:PytorchQuantizationWrapper):
     for name in layer.weights_quantizers.keys():
         quantized_weight = torch.nn.Parameter(layer.get_quantized_weights()[name]).detach()
         linear_layer = getattr(layer, LAYER)
-        delattr(linear_layer, name)
-        setattr(linear_layer, name, torch.nn.Parameter(quantized_weight))
+
+        # If the name is a string, we assume it's a named attribute of the linear layer
+        if isinstance(name, str):
+            # Remove the existing attribute from the linear layer
+            delattr(linear_layer, name)
+            # Replace it with the quantized version as a new parameter
+            setattr(linear_layer, name, torch.nn.Parameter(quantized_weight))
+        else:
+            # If the name is not a string, it must be an integer representing a positional weight
+            assert isinstance(name, int)
+            attr_name = f'{QUANTIZED_POSITIONAL_WEIGHT}_{name}'
+
+            # Note: This naming scheme is used to mimic the behavior expected in
+            # the PytorchQuantizationWrapper's forward method, which looks for attributes
+            # like 'quantized_pos_weight_0', 'quantized_pos_weight_1', etc.
+
+            if hasattr(layer, attr_name):
+                delattr(layer, attr_name)
+
+            # Add the quantized weight as a new attribute directly on the parent layer
+            setattr(layer, attr_name, quantized_weight)
 
     # Clear the weights quantizers dictionary
     layer.weights_quantizers = {}