mct-nightly 2.4.0.20250605.606__py3-none-any.whl → 2.4.0.20250605.151156__py3-none-any.whl

{mct_nightly-2.4.0.20250605.606.dist-info → mct_nightly-2.4.0.20250605.151156.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.4.0.20250605.606
+Version: 2.4.0.20250605.151156
 Summary: A Model Compression Toolkit for neural networks
 Author-email: ssi-dnn-dev@sony.com
 Classifier: Programming Language :: Python :: 3

{mct_nightly-2.4.0.20250605.606.dist-info → mct_nightly-2.4.0.20250605.151156.dist-info}/RECORD

@@ -1,5 +1,5 @@
-mct_nightly-2.4.0.20250605.606.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-model_compression_toolkit/__init__.py,sha256=lCWC3Mfi8O78xOPtfaV5XQKWjw8T4hxVdcOiUD-CGkU,1557
+mct_nightly-2.4.0.20250605.151156.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
+model_compression_toolkit/__init__.py,sha256=G6fQ6rBpGb4hP9184dUezrop85VHb7DUM4hqOjRCOss,1557
 model_compression_toolkit/constants.py,sha256=KNgiNLpsMgSYyXMNEbHXd4bFNerQc1D6HH3vpbUq_Gs,4086
 model_compression_toolkit/defaultdict.py,sha256=LSc-sbZYXENMCw3U9F4GiXuv67IKpdn0Qm7Fr11jy-4,2277
 model_compression_toolkit/logger.py,sha256=L3q7tn3Uht0i_7phnlOWMR2Te2zvzrt2HOz9vYEInts,4529
@@ -334,9 +334,9 @@ model_compression_toolkit/exporter/model_exporter/keras/int8_tflite_exporter.py,
 model_compression_toolkit/exporter/model_exporter/keras/keras_export_facade.py,sha256=NzcX7rxLk__Kpuim_VXaOHS4tyiRtRBoERPE00GbdfA,5862
 model_compression_toolkit/exporter/model_exporter/keras/mctq_keras_exporter.py,sha256=qXXkv3X_wb7t622EOHwXIxfGLGaDqh0T0y4UxREi4Bo,1976
 model_compression_toolkit/exporter/model_exporter/pytorch/__init__.py,sha256=uZ2RigbY9O2PJ0Il8wPpS_s7frgg9WUGd_SHeKGyl1A,699
-model_compression_toolkit/exporter/model_exporter/pytorch/base_pytorch_exporter.py,sha256=sCfK6wXWFgMW0dUaRgWhLc0GUiCHTU-H16I17-XNJu0,6360
+model_compression_toolkit/exporter/model_exporter/pytorch/base_pytorch_exporter.py,sha256=9adOGG1nyviNzuL-1aJXyL0c_VQllSZWiG2gR-puywo,6420
 model_compression_toolkit/exporter/model_exporter/pytorch/export_serialization_format.py,sha256=bPevy6OBqng41PqytBR55e6cBEuyrUS0H8dWX4zgjQ4,967
-model_compression_toolkit/exporter/model_exporter/pytorch/fakely_quant_onnx_pytorch_exporter.py,sha256=v6DwLNZX5RcWHUfYXfJZXNFXPfHsMm5Wl0N29o0diSc,8856
+model_compression_toolkit/exporter/model_exporter/pytorch/fakely_quant_onnx_pytorch_exporter.py,sha256=5S3lyNVc3F62mvS8Q-RTmgQXWI6GWZ5YRvjG4qFy2MM,9520
 model_compression_toolkit/exporter/model_exporter/pytorch/fakely_quant_torchscript_pytorch_exporter.py,sha256=ksWV2A-Njo-wAxQ_Ye2sLIZXBWJ_WNyjT7-qFFwvV2o,2897
 model_compression_toolkit/exporter/model_exporter/pytorch/pytorch_export_facade.py,sha256=7xuUrHPMiifn23sWfeiqR9wkYhm8EweDRL_vF-JSxMY,6642
 model_compression_toolkit/exporter/model_wrapper/__init__.py,sha256=7CF2zvpTrIEm8qnbuHnLZyTZkwBBxV24V8QA0oxGbh0,1187
@@ -529,7 +529,7 @@ model_compression_toolkit/xquant/pytorch/model_analyzer.py,sha256=b93o800yVB3Z-i
 model_compression_toolkit/xquant/pytorch/pytorch_report_utils.py,sha256=UVN_S9ULHBEldBpShCOt8-soT8YTQ5oE362y96qF_FA,3950
 model_compression_toolkit/xquant/pytorch/similarity_functions.py,sha256=CERxq5K8rqaiE-DlwhZBTUd9x69dtYJlkHOPLB54vm8,2354
 model_compression_toolkit/xquant/pytorch/tensorboard_utils.py,sha256=mkoEktLFFHtEKzzFRn_jCnxjhJolK12TZ5AQeDHzUO8,9767
-mct_nightly-2.4.0.20250605.606.dist-info/METADATA,sha256=lpXR9ECa1ziVfo4rUtBZsGmZPS--wjUOfVw-1emCUqg,25087
-mct_nightly-2.4.0.20250605.606.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-mct_nightly-2.4.0.20250605.606.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
-mct_nightly-2.4.0.20250605.606.dist-info/RECORD,,
+mct_nightly-2.4.0.20250605.151156.dist-info/METADATA,sha256=mw_6NFD665X2EECv7pw8PTTnwnhNNyVh5skSnLTlmtw,25090
+mct_nightly-2.4.0.20250605.151156.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+mct_nightly-2.4.0.20250605.151156.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
+mct_nightly-2.4.0.20250605.151156.dist-info/RECORD,,

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__ = "2.4.0.20250605.000606"
+__version__ = "2.4.0.20250605.151156"

model_compression_toolkit/exporter/model_exporter/pytorch/base_pytorch_exporter.py

@@ -56,7 +56,7 @@ def find_and_assign_metadata_attr(model: torch.nn.Module, attr_name: str = 'meta
                 f"Only the first one was assigned to 'model.metadata'.")
 
 
-def _set_quantized_weights_in_wrapper(layer:PytorchQuantizationWrapper):
+def _set_quantized_weights_in_wrapper(layer: PytorchQuantizationWrapper):
     """
        Sets the quantized weights in the provided PytorchQuantizationWrapper layer.
        Replaces the original weights in the layer with the quantized weights.
@@ -124,7 +124,7 @@ class BasePyTorchExporter(Exporter):
         self.model = copy.deepcopy(self.model)
         self.repr_dataset = repr_dataset
 
-    def _substitute_fully_quantized_model(self):
+    def _substitute_fully_quantized_model(self, replace_wrapped=True):
         """
         Substitution for pytorch "fully-quantized" models. It first uses the weight quantizers
         in PytorchQuantizationWrapper layers to quantize the weights and set them in the layer.
@@ -136,8 +136,9 @@ class BasePyTorchExporter(Exporter):
             if isinstance(layer, PytorchQuantizationWrapper):
                 _set_quantized_weights_in_wrapper(layer)
 
-        # Replace PytorchQuantizationWrapper layers with their internal layers
-        self._replace_wrapped_with_unwrapped()
+        if replace_wrapped:
+            # Replace PytorchQuantizationWrapper layers with their internal layers
+            self._replace_wrapped_with_unwrapped()
 
     def _replace_wrapped_with_unwrapped(self):
         """
@@ -148,3 +149,4 @@ class BasePyTorchExporter(Exporter):
         for name, module in self.model.named_children():
             if isinstance(module, PytorchQuantizationWrapper):
                 setattr(self.model, name, module.layer)
+

model_compression_toolkit/exporter/model_exporter/pytorch/fakely_quant_onnx_pytorch_exporter.py

@@ -73,6 +73,15 @@ if FOUND_ONNX:
             Returns:
                 Fake-quant PyTorch model.
             """
+            # List all activation quantization holders with num_bits>8 and replace them with Identity, because
+            # ONNX doesn't support quantization of more than 8 bits for torch.fake_quantize_per_tensor_affine.
+            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 layer in self.model.children():
                 self.is_layer_exportable_fn(layer)
                 # Set reuse for weight quantizers if quantizer is reused
@@ -89,7 +98,7 @@ if FOUND_ONNX:
             if self._use_onnx_custom_quantizer_ops:
                 self._enable_onnx_custom_ops_export()
             else:
-                self._substitute_fully_quantized_model()
+                self._substitute_fully_quantized_model(replace_wrapped=False)
 
             if self._use_onnx_custom_quantizer_ops:
                 Logger.info(f"Exporting onnx model with MCTQ quantizers: {self.save_model_path}")
@@ -166,6 +175,6 @@ if FOUND_ONNX:
                         wq.enable_custom_impl()
 
 else:
-    def FakelyQuantONNXPyTorchExporter(*args, **kwargs):
+    def FakelyQuantONNXPyTorchExporter(*args, **kwargs):  # pragma: no cover
         Logger.critical("ONNX must be installed to use 'FakelyQuantONNXPyTorchExporter'. "
-                        "The 'onnx' package is missing.")  # pragma: no cover
+                        "The 'onnx' package is missing.")