mct-nightly 2.3.0.20250604.611__py3-none-any.whl → 2.4.0.20250605.606__py3-none-any.whl

{mct_nightly-2.3.0.20250604.611.dist-info → mct_nightly-2.4.0.20250605.606.dist-info}/METADATA

@@ -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.dist-info → mct_nightly-2.4.0.20250605.606.dist-info}/RECORD

@@ -1,5 +1,5 @@
-mct_nightly-2.3.0.20250604.611.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-model_compression_toolkit/__init__.py,sha256=MNjhJRHWg20p0VjHTzI6AhnDetVQ7Oy8YQrAOLRA-4Y,1557
+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
 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
@@ -16,7 +16,7 @@ model_compression_toolkit/core/common/framework_implementation.py,sha256=JQI_eoZ
 model_compression_toolkit/core/common/framework_info.py,sha256=5tderHT-7Cd21QrRFIJj3hH_gAcnlivOzwZ5m1ldJOs,6526
 model_compression_toolkit/core/common/memory_computation.py,sha256=ixoSpV5ZYZGyzhre3kQcvR2sNA8KBsPZ3lgbkDnw9Cs,1205
 model_compression_toolkit/core/common/model_builder_mode.py,sha256=jll9-59OPaE3ug7Y9-lLyV99_FoNHxkGZMgcm0Vkpss,1324
-model_compression_toolkit/core/common/model_collector.py,sha256=Tno3-qx9jmPZAZyLYgbPlMLHakVfuEH5deuToZNuCb0,13195
+model_compression_toolkit/core/common/model_collector.py,sha256=ssYLdo2FPEzXoFKQonUA3ZbrIaDp-UmLhv0mWMCHH8U,13225
 model_compression_toolkit/core/common/model_validation.py,sha256=LaG8wd6aZl0OJgieE3SeiVDEPxtk8IHq9-3wSnmWhY4,1214
 model_compression_toolkit/core/common/node_prior_info.py,sha256=WXX_PrGVG9M9I_REG5ZzFBohwmV4yf356sZnrja_FLo,2832
 model_compression_toolkit/core/common/similarity_analyzer.py,sha256=S3f6WgHyw62dGcxpX51FGKyfebe2zv9ABKbjtGyKRvY,9215
@@ -35,7 +35,7 @@ model_compression_toolkit/core/common/fusion/fusing_info.py,sha256=uDxF0awrjn3Sb
 model_compression_toolkit/core/common/fusion/graph_fuser.py,sha256=yxxxuwrmQ4wLW-PlTu0MEW59LmNJEh1OWy9Li15YH-8,7520
 model_compression_toolkit/core/common/graph/__init__.py,sha256=Xr-Lt_qXMdrCnnOaUS_OJP_3iTTGfPCLf8_vSrQgCs0,773
 model_compression_toolkit/core/common/graph/base_graph.py,sha256=YA0c8ucaaZu9eRO-xruLqDT3QFOpxq24ViG6ILS2jqA,41403
-model_compression_toolkit/core/common/graph/base_node.py,sha256=AbUadAT581zelVcGcK9_--6CAGiht9qwkeWahwT3RzE,33389
+model_compression_toolkit/core/common/graph/base_node.py,sha256=8GEqZ8VMtVvJZuiSUVdokCq6NkFosOssetcod21DwDM,33604
 model_compression_toolkit/core/common/graph/edge.py,sha256=buoSEUZwilWBK3WeBKpJ-GeDaUA1SDdOHxDpxU_bGpk,3784
 model_compression_toolkit/core/common/graph/functional_node.py,sha256=GH5wStmw8SoAj5IdT_-ItN1Meo_P5NUTt_5bgJC4fak,3935
 model_compression_toolkit/core/common/graph/graph_matchers.py,sha256=CrDoHYq4iPaflgJWmoJ1K4ziLrRogJvFTVWg8P0UcDU,4744
@@ -65,7 +65,7 @@ model_compression_toolkit/core/common/mixed_precision/bit_width_setter.py,sha256
 model_compression_toolkit/core/common/mixed_precision/configurable_quant_id.py,sha256=LLDguK7afsbN742ucLpmJr5TUfTyFpK1vbf2bpVr1v0,882
 model_compression_toolkit/core/common/mixed_precision/configurable_quantizer_utils.py,sha256=7dKMi5S0zQZ16m8NWn1XIuoXsKuZUg64G4-uK8-j1PQ,5177
 model_compression_toolkit/core/common/mixed_precision/mixed_precision_candidates_filter.py,sha256=6pLUEEIqRTVIlCYQC4JIvY55KAvuBHEX8uTOQ-1Ac4Q,3859
-model_compression_toolkit/core/common/mixed_precision/mixed_precision_quantization_config.py,sha256=3vuhlpGvUdeGpEMfNIYONMN9NQ_VFIJykC6Gwlo-gaI,6728
+model_compression_toolkit/core/common/mixed_precision/mixed_precision_quantization_config.py,sha256=qsFW_H3HiN3Mr1lwSg15CQb4cUBtGVfewdGzZoJVijo,6737
 model_compression_toolkit/core/common/mixed_precision/mixed_precision_ru_helper.py,sha256=axgAypzsiCOw04ZOtOEjK4riuNsaEU2qU6KkWnEXtMo,4951
 model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_facade.py,sha256=1877xOUdgpWrXWyhdX1pJOePuopq43L71WqBFMqzyR4,6418
 model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_manager.py,sha256=3D_gHdcxsuINiTZqEAJXbxwYDg5qqXD51k4_smmWI9M,28553
@@ -334,7 +334,7 @@ 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=kmddLPXN47b4UOlip8u7zfILcw7rvKHp5UZ_-nII_9U,5369
+model_compression_toolkit/exporter/model_exporter/pytorch/base_pytorch_exporter.py,sha256=sCfK6wXWFgMW0dUaRgWhLc0GUiCHTU-H16I17-XNJu0,6360
 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_torchscript_pytorch_exporter.py,sha256=ksWV2A-Njo-wAxQ_Ye2sLIZXBWJ_WNyjT7-qFFwvV2o,2897
@@ -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-2.3.0.20250604.611.dist-info/METADATA,sha256=vtgFVscrMOf1owGEtgvDKYIE2oCQ2ewOg0-JbeDPI0k,25087
-mct_nightly-2.3.0.20250604.611.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-mct_nightly-2.3.0.20250604.611.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
-mct_nightly-2.3.0.20250604.611.dist-info/RECORD,,
+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,,

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"

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.

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'

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),

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 = {}