PyPI - mct-nightly - Versions diffs - 2.3.0.20250405.519__tar.gz → 2.3.0.20250407.1120__tar.gz - Mend

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.3.0.20250405.519
+Version: 2.3.0.20250407.1120
 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: 2.3.0.20250405.519
+Version: 2.3.0.20250407.1120
 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__ = "2.3.0.20250405.000519"
+__version__ = "2.3.0.20250407.001120"

@@ -119,6 +119,8 @@ def filter_node_candidates(node: BaseNode, fw_info) -> List[CandidateNodeQuantiz
         final_candidates = _filter_bit_method_dups(filtered_candidates, kernel_attr)
     elif kernel_attr is None or not node.is_weights_quantization_enabled(kernel_attr):
+        # TODO:
+        #  To allow MP on positional weights we need to modify this to consider all weights not only kernel.
         # Remove candidates that have duplicated activation candidates for node with disabled weights quantization.
         # Replacing the weights n_bits in the remained configurations with default value to prevent confusion.
         seen_candidates = set()

@@ -25,7 +25,9 @@ from model_compression_toolkit.core.common.quantization.quantization_params_fn_s
 from model_compression_toolkit.core.common.quantization.quantization_config import QuantizationConfig, \
     QuantizationErrorMethod
-from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import AttributeQuantizationConfig, \
+from model_compression_toolkit.target_platform_capabilities.constants import POS_ATTR
+from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import \
+    AttributeQuantizationConfig, \
     OpQuantizationConfig
 if TYPE_CHECKING:
@@ -209,7 +211,7 @@ class NodeActivationQuantizationConfig(BaseNodeQuantizationConfig):
                self.shift_negative_activation_correction == other.shift_negative_activation_correction and \
                self.z_threshold == other.z_threshold and \
                self.shift_negative_ratio == other.shift_negative_ratio and \
-               self.shift_negative_threshold_recalculation == other.shift_negative_threshold_recalculation
+               self.shift_negative_threshold_recalculation == other.shift_negative_threshold_recalculation
     def __hash__(self):
         return hash((self.activation_quantization_fn,
@@ -395,9 +397,35 @@ class NodeWeightsQuantizationConfig(BaseNodeQuantizationConfig):
         for attr in node_attrs_list:
             if isinstance(attr, int):
                 # this is a positional attribute, so it needs to be handled separately.
-                self.pos_attributes_config_mapping[attr] = WeightsAttrQuantizationConfig(qc=qc,
-                                                                                         weights_attr_cfg=op_cfg.default_weight_attr_config,
-                                                                                         weights_channels_axis=weights_channels_axis)
+                # Search for any keys in the op config's attribute weight config mapping that contain the
+                # POS_ATTR string. If none are found, it indicates that no specific quantization config is defined for
+                # positional weights, so the default config will be used instead.
+                attrs_included_in_name = {k: v for k, v in op_cfg.attr_weights_configs_mapping.items() if
+                                          POS_ATTR in k}
+                if len(attrs_included_in_name) > 1:  # pragma: no cover
+                    raise ValueError(f"Found multiple attribute in FQC OpConfig that are contained "
+                                     f"in the attribute name '{attr}'."
+                                     f"Please fix the FQC attribute names mapping such that each operator's attribute"
+                                     f" would have a unique matching name.")
+                # If no specific positional attribute config is found, fall back to the default weight attribute config.
+                if len(attrs_included_in_name) == 0:
+                    attr_cfg = op_cfg.default_weight_attr_config
+                    # Register this attribute under the positional attributes config mapping.
+                    self.pos_attributes_config_mapping[attr] = WeightsAttrQuantizationConfig(qc=qc,
+                                                                                             weights_attr_cfg=attr_cfg,
+                                                                                             weights_channels_axis=
+                                                                                             weights_channels_axis)
+                else:
+                    # If a specific config was found using POS_ATTR, use it.
+                    attr_cfg = list(attrs_included_in_name.values())[0]
+                    # Register this attribute under the regular attributes config mapping.
+                    self.attributes_config_mapping[attr] = WeightsAttrQuantizationConfig(qc=qc,
+                                                                                         weights_attr_cfg=attr_cfg,
+                                                                                         weights_channels_axis=
+                                                                                         weights_channels_axis)
             else:
                 # In Tensorflow, the attribute name is composed of the framework attribute name and the layer name,
                 # therefore, we need to look for the attribute in the op_cfg that is contained in the node attribute's name.

@@ -73,6 +73,11 @@ if FOUND_ONNX:
             """
             for layer in self.model.children():
                 self.is_layer_exportable_fn(layer)
+                # Set reuse for weight quantizers if quantizer is reused
+                if isinstance(layer, PytorchQuantizationWrapper):
+                    for _, quantizer in layer.weights_quantizers.items():
+                        if quantizer.reuse:
+                            quantizer.enable_reuse_quantizer()
             # Set forward that is used during onnx export.
             # If _use_onnx_custom_quantizer_ops is set to True, the quantizer forward function will use
@@ -116,6 +121,13 @@ if FOUND_ONNX:
                                   dynamic_axes={'input': {0: 'batch_size'},
                                                 'output': {0: 'batch_size'}})
+            for layer in self.model.children():
+                # Set disable for reuse for weight quantizers if quantizer was reused
+                if isinstance(layer, PytorchQuantizationWrapper):
+                    for _, quantizer in layer.weights_quantizers.items():
+                        if quantizer.reuse:
+                            quantizer.disable_reuse_quantizer()
         def _enable_onnx_custom_ops_export(self):
             """
             Enable the custom implementation forward in quantizers, so it is exported

@@ -48,7 +48,7 @@ if FOUND_TORCH:
             # Set reuse for weight quantizers if node is reused
             for _, quantizer in weight_quantizers.items():
                 if node.reuse_group:
-                    quantizer.enable_reuse_quantizer()
+                    quantizer.reuse = True
             # for positional weights we need to extract the weight's value.
             weights_values = {attr: fw_impl.to_tensor(node.get_weights_by_keys(attr))
                               for attr in weight_quantizers if isinstance(attr, int)}

@@ -29,6 +29,7 @@ QNNPACK_TP_MODEL = 'qnnpack'
 # TP Attributes
 KERNEL_ATTR = "kernel_attr"
 BIAS_ATTR = "bias_attr"
+POS_ATTR = "pos_attr"
 # TODO: this is duplicated from the core frameworks constants files, because the original consts can't be used here
 #  duo to circular dependency. It might be best to extract the constants from the core file and put them here (in a

mct-nightly 2.3.0.20250405.519__tar.gz → 2.3.0.20250407.1120__tar.gz

mct-nightly 2.3.0.20250405.519tar.gz → 2.3.0.20250407.1120tar.gz