PyPI - mct-nightly - Versions diffs - 2.4.0.20250618.606__tar.gz → 2.4.0.20250619.621__tar.gz - Mend

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.4.0.20250618.606
+Version: 2.4.0.20250619.621
 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: 2.4.0.20250618.606
+Version: 2.4.0.20250619.621
 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__ = "2.4.0.20250618.000606"
+__version__ = "2.4.0.20250619.000621"

@@ -28,12 +28,14 @@ if FOUND_TF:
         AttachTpcToKeras
     from model_compression_toolkit.target_platform_capabilities.constants import DEFAULT_TP_MODEL
     from model_compression_toolkit.core.keras.keras_implementation import KerasImplementation
+    from model_compression_toolkit.core.keras.default_framework_info import set_keras_info
     from tensorflow.keras.models import Model
     from model_compression_toolkit import get_target_platform_capabilities
     KERAS_DEFAULT_TPC = get_target_platform_capabilities(TENSORFLOW, DEFAULT_TP_MODEL)
+    @set_keras_info
     def keras_resource_utilization_data(in_model: Model,
                                         representative_data_gen: Callable,
                                         core_config: CoreConfig = CoreConfig(

@@ -21,6 +21,7 @@ from model_compression_toolkit.logger import Logger
 if FOUND_TF:
     import keras
+    from model_compression_toolkit.core.keras.default_framework_info import set_keras_info
     from model_compression_toolkit.exporter.model_wrapper.keras.validate_layer import is_keras_layer_exportable
     from model_compression_toolkit.exporter.model_exporter.keras.fakely_quant_keras_exporter import \
         FakelyQuantKerasExporter
@@ -36,6 +37,7 @@ if FOUND_TF:
         KerasExportSerializationFormat.TFLITE: [QuantizationFormat.FAKELY_QUANT, QuantizationFormat.INT8]
     }
+    @set_keras_info
     def keras_export_model(model: keras.models.Model,
                            save_model_path: str,
                            is_layer_exportable_fn: Callable = is_keras_layer_exportable,

@@ -73,23 +73,25 @@ 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
-                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.")
+            # When exporting using Fakely Quant Quantization Format 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.
+            if not self._use_onnx_custom_quantizer_ops:
+                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
+                    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"
+                                    f"FQ export, deletion of activation holder '{act_holder}' failed — could not locate"
+                                    f"one or more intermediate attributes in the path.")
             for layer in self.model.children():
                 self.is_layer_exportable_fn(layer)

@@ -27,6 +27,7 @@ DEFAULT_ONNX_OPSET_VERSION = 15
 if FOUND_TORCH:
     import torch.nn
+    from model_compression_toolkit.core.pytorch.default_framework_info import set_pytorch_info
     from model_compression_toolkit.exporter.model_exporter.pytorch.fakely_quant_onnx_pytorch_exporter import FakelyQuantONNXPyTorchExporter
     from model_compression_toolkit.exporter.model_exporter.pytorch.fakely_quant_torchscript_pytorch_exporter import FakelyQuantTorchScriptPyTorchExporter
     from model_compression_toolkit.exporter.model_wrapper.pytorch.validate_layer import is_pytorch_layer_exportable
@@ -41,6 +42,7 @@ if FOUND_TORCH:
         PytorchExportSerializationFormat.ONNX: [QuantizationFormat.FAKELY_QUANT, QuantizationFormat.MCTQ]
     }
+    @set_pytorch_info
     def pytorch_export_model(model: torch.nn.Module,
                              save_model_path: str,
                              repr_dataset: Callable,

mct-nightly 2.4.0.20250618.606__tar.gz → 2.4.0.20250619.621__tar.gz

mct-nightly 2.4.0.20250618.606tar.gz → 2.4.0.20250619.621tar.gz