mct-nightly 2.3.0.20250405.519__py3-none-any.whl → 2.3.0.20250407.1120__py3-none-any.whl

{mct_nightly-2.3.0.20250405.519.dist-info → mct_nightly-2.3.0.20250407.1120.dist-info}/METADATA

@@ -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

{mct_nightly-2.3.0.20250405.519.dist-info → mct_nightly-2.3.0.20250407.1120.dist-info}/RECORD

@@ -1,5 +1,5 @@
-mct_nightly-2.3.0.20250405.519.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-model_compression_toolkit/__init__.py,sha256=q-6vPTQu-LuGF_ciYO6xH7ip2LG_WgfTyo51-zRt4so,1557
+mct_nightly-2.3.0.20250407.1120.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
+model_compression_toolkit/__init__.py,sha256=rbCx-x__4DNITXM-m79qge3vAJYMLHqI9ym62xr6BTE,1557
 model_compression_toolkit/constants.py,sha256=2ltuH-gdaLZoZV4CPUgKjC3S9ojz2z4OTVdenyVEypU,3912
 model_compression_toolkit/defaultdict.py,sha256=LSc-sbZYXENMCw3U9F4GiXuv67IKpdn0Qm7Fr11jy-4,2277
 model_compression_toolkit/logger.py,sha256=L3q7tn3Uht0i_7phnlOWMR2Te2zvzrt2HOz9vYEInts,4529
@@ -105,8 +105,8 @@ model_compression_toolkit/core/common/quantization/bit_width_config.py,sha256=0H
 model_compression_toolkit/core/common/quantization/candidate_node_quantization_config.py,sha256=lyWPvnoX8BmulhLKR20r5gT2_Yan7P40d8EcgDhErPk,4905
 model_compression_toolkit/core/common/quantization/core_config.py,sha256=yxCzWqldcHoe8GGxrH0tp99bhrc5jDT7SgZftnMUUBE,2374
 model_compression_toolkit/core/common/quantization/debug_config.py,sha256=zJP2W9apUPX9RstpPWWK71wr9xJsg7j-s7lGV4_bQdc,1510
-model_compression_toolkit/core/common/quantization/filter_nodes_candidates.py,sha256=IHVX-Gdekru4xLuDTgcsp_JCnRtuVWnbYsDBQuSXTKc,7079
-model_compression_toolkit/core/common/quantization/node_quantization_config.py,sha256=teDclY8WmuVqqa9Fgr6WY-7ILDep0QKzKxoZCKzBG2k,26960
+model_compression_toolkit/core/common/quantization/filter_nodes_candidates.py,sha256=n2A8pO7_DMMae4o69U0I00iW6mzeRlRfKHDxlQUBBuI,7204
+model_compression_toolkit/core/common/quantization/node_quantization_config.py,sha256=bRK8c3iH64RGBZ8jZb0i4vVjcEvCxgKX5CdTEmIMtZ4,29077
 model_compression_toolkit/core/common/quantization/quantization_config.py,sha256=UkSVW7d1OF_Px9gAjsqqK65aYhIBFWaBO-_IH6_AFfg,4403
 model_compression_toolkit/core/common/quantization/quantization_fn_selection.py,sha256=HfBkSiRTOf9mNF-TNQHTCCs3xSg66F20no0O6vl5v1Y,2154
 model_compression_toolkit/core/common/quantization/quantization_params_fn_selection.py,sha256=7eG7dl1TcbdnHwgmvyjarxLs0o6Lw_9VAjXAm4rsiBk,3791
@@ -335,7 +335,7 @@ model_compression_toolkit/exporter/model_exporter/keras/mctq_keras_exporter.py,s
 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=UPVkEUQCMZ4Lld6CRnEOPEmlfe5vcQZG0Q3FwRBodD4,4021
 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=07yCCypQaQ3FfhrxKhWwa03bnEgJ8lIknafqeYfvzCQ,6803
+model_compression_toolkit/exporter/model_exporter/pytorch/fakely_quant_onnx_pytorch_exporter.py,sha256=G2X_lDx6u12U4ErEuEHCdNczh0qSGWObySw3upEys6Q,7506
 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=8vYGKa58BkasvoHejYaPwubOJPcW0s-RY79_Kkw0Hy8,6236
 model_compression_toolkit/exporter/model_wrapper/__init__.py,sha256=7CF2zvpTrIEm8qnbuHnLZyTZkwBBxV24V8QA0oxGbh0,1187
@@ -349,7 +349,7 @@ model_compression_toolkit/exporter/model_wrapper/keras/builder/node_to_quantizer
 model_compression_toolkit/exporter/model_wrapper/pytorch/__init__.py,sha256=Rf1RcYmelmdZmBV5qOKvKWF575ofc06JFQSq83Jz99A,696
 model_compression_toolkit/exporter/model_wrapper/pytorch/validate_layer.py,sha256=vQUGbCi8_pGoN8DwQ0IblSeN6L9t6Cr0reZNuCbBpkM,3469
 model_compression_toolkit/exporter/model_wrapper/pytorch/builder/__init__.py,sha256=cco4TmeIDIh32nj9ZZXVkws4dd9F2UDrmjKzTN8G0V0,697
-model_compression_toolkit/exporter/model_wrapper/pytorch/builder/fully_quantized_model_builder.py,sha256=dpN2Hyb56Wt4INEtBJAOxZeFdhIwdx__WFTmOVkxMLc,6470
+model_compression_toolkit/exporter/model_wrapper/pytorch/builder/fully_quantized_model_builder.py,sha256=bkcuNcbPwdB6pAvQhaBaCfk0qGaiJHMJdKYG5A1XdHE,6458
 model_compression_toolkit/exporter/model_wrapper/pytorch/builder/node_to_quantizer.py,sha256=Pl8a8MSZMzNbm5vngujFjCt_iSMbSmKjlcL1DvN9nTM,9292
 model_compression_toolkit/gptq/__init__.py,sha256=pEgkJvmf05KSw70iLDTz_6LI_2Oi5L8sTN0JsEUpnpk,1445
 model_compression_toolkit/gptq/runner.py,sha256=La12JTYjWyJW0YW4Al4TP1_Xi4JWBCEKw6FR_JQsxe0,5982
@@ -431,7 +431,7 @@ model_compression_toolkit/qat/pytorch/quantizer/ste_rounding/__init__.py,sha256=
 model_compression_toolkit/qat/pytorch/quantizer/ste_rounding/symmetric_ste.py,sha256=p1JqtBZZVHTV5caR1U0d1t2UcTz0ACNyLcJTBFUEq98,6173
 model_compression_toolkit/qat/pytorch/quantizer/ste_rounding/uniform_ste.py,sha256=wWehe5R0xVHSm3ruMrUc8RzW5UVAVCMgUTUMPDsvy9g,5487
 model_compression_toolkit/target_platform_capabilities/__init__.py,sha256=8RVOriZg-XNjSt53h_4Yum0oRgOe2gp5H45dfG_lZxE,1415
-model_compression_toolkit/target_platform_capabilities/constants.py,sha256=BFSgDwYWU1sZShjoW2S7eH3AI0D4SqDOeOu_sQ971LE,1518
+model_compression_toolkit/target_platform_capabilities/constants.py,sha256=JRz9DoxLRpkqvu532TFkIvv0595Bfb9NtU4pRp4urDY,1540
 model_compression_toolkit/target_platform_capabilities/immutable.py,sha256=YhROBiXEIB3TU-bAFrnL3qbAsb1yuWPBAQ_CLOJbYUU,1827
 model_compression_toolkit/target_platform_capabilities/tpc_io_handler.py,sha256=4ydTWWKv_PEOAFok2JtxFNj8rav-0IlqcXKF6lnhHNE,4157
 model_compression_toolkit/target_platform_capabilities/schema/__init__.py,sha256=pKAdbTCFM_2BrZXUtTIw0ouKotrWwUDF_hP3rPwCM2k,696
@@ -527,7 +527,7 @@ 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-2.3.0.20250405.519.dist-info/METADATA,sha256=PHbbtoK9OgtSBmT1-3MirDEL_DgGZjTARyZ99aB_VUM,27098
-mct_nightly-2.3.0.20250405.519.dist-info/WHEEL,sha256=CmyFI0kx5cdEMTLiONQRbGQwjIoR1aIYB7eCAQ4KPJ0,91
-mct_nightly-2.3.0.20250405.519.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
-mct_nightly-2.3.0.20250405.519.dist-info/RECORD,,
+mct_nightly-2.3.0.20250407.1120.dist-info/METADATA,sha256=3FbL_LGv5kbK8httRgRpIX64Ff6HElqC7NSH2BU5WDQ,27099
+mct_nightly-2.3.0.20250407.1120.dist-info/WHEEL,sha256=CmyFI0kx5cdEMTLiONQRbGQwjIoR1aIYB7eCAQ4KPJ0,91
+mct_nightly-2.3.0.20250407.1120.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
+mct_nightly-2.3.0.20250407.1120.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__ = "2.3.0.20250405.000519"
+__version__ = "2.3.0.20250407.001120"

model_compression_toolkit/core/common/quantization/filter_nodes_candidates.py

@@ -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()

model_compression_toolkit/core/common/quantization/node_quantization_config.py

@@ -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.

model_compression_toolkit/exporter/model_exporter/pytorch/fakely_quant_onnx_pytorch_exporter.py

@@ -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

model_compression_toolkit/exporter/model_wrapper/pytorch/builder/fully_quantized_model_builder.py

@@ -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)}

model_compression_toolkit/target_platform_capabilities/constants.py

@@ -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