mct-nightly 2.4.0.20250626.617__py3-none-any.whl → 2.4.0.20250628.555__py3-none-any.whl

{mct_nightly-2.4.0.20250626.617.dist-info → mct_nightly-2.4.0.20250628.555.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.4.0.20250626.617
+Version: 2.4.0.20250628.555
 Summary: A Model Compression Toolkit for neural networks
 Author-email: ssi-dnn-dev@sony.com
 Classifier: Programming Language :: Python :: 3

{mct_nightly-2.4.0.20250626.617.dist-info → mct_nightly-2.4.0.20250628.555.dist-info}/RECORD

@@ -1,5 +1,5 @@
-mct_nightly-2.4.0.20250626.617.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-model_compression_toolkit/__init__.py,sha256=7rL9mYdsLggQYsvr-mE5212sKK0euktVF9gCAbTK6a4,1557
+mct_nightly-2.4.0.20250628.555.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
+model_compression_toolkit/__init__.py,sha256=i97j--Rm-5M5jfoHTDDtJSsIJ2swWaHPeoSNq5P8zY4,1557
 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
@@ -7,7 +7,7 @@ model_compression_toolkit/metadata.py,sha256=x_Bk4VpzILdsFax6--CZ3X18qUTP28sbF_A
 model_compression_toolkit/verify_packages.py,sha256=l0neIRr8q_QwxmuiTI4vyCMDISDedK0EihjEQUe66tE,1319
 model_compression_toolkit/core/__init__.py,sha256=HNverPpoqEyFKTa7iEdOqqY2P0Gq-7GMejNOi6ZPcQs,2042
 model_compression_toolkit/core/analyzer.py,sha256=5P03LbkFy-mu31TMAiQoIKcsA1-DNz7cTzkGvRaXtbw,3505
-model_compression_toolkit/core/graph_prep_runner.py,sha256=d-NrEUQETiY23-n9rLcf73PEnfaW-DbF5FSgWgRRE2w,10515
+model_compression_toolkit/core/graph_prep_runner.py,sha256=8K5JWOiCJxd54nlyYQS7g5UJl30sE7VWuscxForJ3ZA,10526
 model_compression_toolkit/core/quantization_prep_runner.py,sha256=tz91E1BaNc_K0lvVZGB8oS6ya5N4Z5TJLG4pSM3hx30,6229
 model_compression_toolkit/core/runner.py,sha256=pmRJeIqB0dKnyNsNSoaBgAkHv_RhQZylknWRFmnoStM,12423
 model_compression_toolkit/core/common/__init__.py,sha256=Wh127PbXcETZX_d1PQqZ71ETK3J9XO5A-HpadGUbj6o,1447
@@ -34,8 +34,8 @@ model_compression_toolkit/core/common/fusion/__init__.py,sha256=Rf1RcYmelmdZmBV5
 model_compression_toolkit/core/common/fusion/fusing_info.py,sha256=Z-O03-DlM4XyllVg5FaQlYeIgk5UqoC8dSA6IlRODNI,22693
 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=mlQ-y3sX7C6ebJk43OdQKgQuyfVCpG6zuCQiRGvm3Co,39738
-model_compression_toolkit/core/common/graph/base_node.py,sha256=n_BEQWiKxhpB_vz-ILlKDsllt7E_F4lXhdhnDZSSOYg,35124
+model_compression_toolkit/core/common/graph/base_graph.py,sha256=_bv_XeENRoJJLvro6raa-Cn57awgnl70TvMM4ZNjkyA,41449
+model_compression_toolkit/core/common/graph/base_node.py,sha256=wmhNY6bSM8920HAhl7hRZsMBKIi5h7g6DrytbEbuz18,35536
 model_compression_toolkit/core/common/graph/edge.py,sha256=buoSEUZwilWBK3WeBKpJ-GeDaUA1SDdOHxDpxU_bGpk,3784
 model_compression_toolkit/core/common/graph/functional_node.py,sha256=wtX6CYzAwgzkt_gp3wByfYydIUZV4vjIHt9TU3Ps9nw,4731
 model_compression_toolkit/core/common/graph/graph_matchers.py,sha256=CrDoHYq4iPaflgJWmoJ1K4ziLrRogJvFTVWg8P0UcDU,4744
@@ -106,8 +106,8 @@ model_compression_toolkit/core/common/quantization/bit_width_config.py,sha256=03
 model_compression_toolkit/core/common/quantization/candidate_node_quantization_config.py,sha256=X9W_az8RQtOi4aFDLNzHxES_r6qs0kixB8OQ7seOVe8,4992
 model_compression_toolkit/core/common/quantization/core_config.py,sha256=yxCzWqldcHoe8GGxrH0tp99bhrc5jDT7SgZftnMUUBE,2374
 model_compression_toolkit/core/common/quantization/debug_config.py,sha256=uH45Uq3Tp9FIyMynex_WY2_y-Kv8LuPw2XXZydnpW5A,1649
-model_compression_toolkit/core/common/quantization/filter_nodes_candidates.py,sha256=iYr3hBkAa4rQ5gCDhKQhrW8xK9luEASrngwdZiLAY_Q,7064
-model_compression_toolkit/core/common/quantization/node_quantization_config.py,sha256=VhM8Wp0_cpG7ukU9trkoAaxhX-Jb4ogGtr1ODfxzNYA,30225
+model_compression_toolkit/core/common/quantization/filter_nodes_candidates.py,sha256=AFabyE57oL-XmI8jUhMA5p6ZVZGJW8dkI1BClN8Gu3w,7670
+model_compression_toolkit/core/common/quantization/node_quantization_config.py,sha256=Zy2Cz3vhRcjV_J7fqJf79J5dyBQHQy9AMGxD9RFuV9I,30250
 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
@@ -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=Y0oBl8qPFsdNrK49XczwmVacInJcOPHslVnFBs-iTCc,3742
 model_compression_toolkit/xquant/pytorch/similarity_functions.py,sha256=CERxq5K8rqaiE-DlwhZBTUd9x69dtYJlkHOPLB54vm8,2354
 model_compression_toolkit/xquant/pytorch/tensorboard_utils.py,sha256=n0HvWBzkBkUJZlS3WeynhpsRTps2qQkjlq7luliBHNU,9627
-mct_nightly-2.4.0.20250626.617.dist-info/METADATA,sha256=lbqZdnDqrpeyBbBGOvnhU3sN00dShfqBziSBw61N9Yo,25087
-mct_nightly-2.4.0.20250626.617.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-mct_nightly-2.4.0.20250626.617.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
-mct_nightly-2.4.0.20250626.617.dist-info/RECORD,,
+mct_nightly-2.4.0.20250628.555.dist-info/METADATA,sha256=FY4d6D9wbWtWWvn7sXzMdVLz_t4R3iVfbrYz_YQBPPA,25087
+mct_nightly-2.4.0.20250628.555.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+mct_nightly-2.4.0.20250628.555.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
+mct_nightly-2.4.0.20250628.555.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__ = "2.4.0.20250626.000617"
+__version__ = "2.4.0.20250628.000555"

model_compression_toolkit/core/common/graph/base_graph.py

@@ -32,13 +32,13 @@ from model_compression_toolkit.core.common.collectors.statistics_collector impor
 from model_compression_toolkit.core.common.collectors.statistics_collector import scale_statistics, shift_statistics
 from model_compression_toolkit.core.common.pruning.pruning_section import PruningSection
 from model_compression_toolkit.core.common.user_info import UserInformation
-from model_compression_toolkit.core.common.quantization.node_quantization_config import ActivationQuantizationMode
+from model_compression_toolkit.core.common.quantization.node_quantization_config import \
+    NodeActivationQuantizationConfig, ActivationQuantizationMode
 from model_compression_toolkit.logger import Logger
 from model_compression_toolkit.target_platform_capabilities.targetplatform2framework import LayerFilterParams
 from model_compression_toolkit.target_platform_capabilities.targetplatform2framework.framework_quantization_capabilities import \
     FrameworkQuantizationCapabilities
 
-
 def validate_graph_after_change(method: Callable) -> Callable:
     """
     Decorator for graph-mutating methods. After the decorated method executes,
@@ -876,15 +876,32 @@ class Graph(nx.MultiDiGraph, GraphSearches):
 
         return intermediate_nodes, next_node
 
-    def disable_fused_nodes_activation_quantization(self):
+    def override_fused_node_activation_quantization_candidates(self):
         """
-        Disable activation quantization for all nodes in fused operations,
+        Override fused node activation quantization candidates for all nodes in fused operations,
         except for the last node in each fused group.
-        """
-        nodes_to_disable = self.fusing_info.get_inner_fln_nodes()
-        for node in nodes_to_disable:
-            for qc in node.candidates_quantization_cfg:
-                qc.activation_quantization_cfg.quant_mode = ActivationQuantizationMode.FLN_QUANT
+        Update the value of quantization_config with the value of op_quaitization_cfg from FusingInfo.
+        """
+        from model_compression_toolkit.core.common.quantization.candidate_node_quantization_config import CandidateNodeQuantizationConfig
+
+        nodes_in_fln = self.fusing_info.get_inner_fln_nodes()
+        for node in nodes_in_fln:
+            fused_node_op_id = self.fusing_info.get_fused_op_id_for_node(node.name)
+            fusiong_op_quaitization_cfg = self.fusing_info.get_fused_op_quantization_config(fused_node_op_id)             
+            org_candidate = node.candidates_quantization_cfg[0]
+            if fusiong_op_quaitization_cfg is not None and fusiong_op_quaitization_cfg.enable_activation_quantization:
+                # Set ActivationQuantizationMode to FLN_QUANT and update the value of quantization_config
+                activation_quantization_cfg = NodeActivationQuantizationConfig(qc=org_candidate,
+                                                                               op_cfg=fusiong_op_quaitization_cfg,
+                                                                               activation_quantization_fn=org_candidate.activation_quantization_cfg.activation_quantization_fn,
+                                                                               activation_quantization_params_fn=org_candidate.activation_quantization_cfg.activation_quantization_params_fn)
+                activation_quantization_cfg.quant_mode = ActivationQuantizationMode.FLN_QUANT
+                for qc in node.candidates_quantization_cfg:
+                    qc.activation_quantization_cfg = activation_quantization_cfg
+            else:
+                # Set ActivationQuantizationMode to FLN_NO_QUANT
+                for qc in node.candidates_quantization_cfg:
+                    qc.activation_quantization_cfg.quant_mode = ActivationQuantizationMode.FLN_NO_QUANT
 
     def validate(self):
         """
@@ -908,4 +925,4 @@ class Graph(nx.MultiDiGraph, GraphSearches):
         """
         Wrap networkx functions (that modifies the graph) with our validate decorator.
         """
-        return super().remove_edge(*args, **kwargs)
+        return super().remove_edge(*args, **kwargs)

model_compression_toolkit/core/common/graph/base_node.py

@@ -216,19 +216,31 @@ class BaseNode:
         Returns: Whether node activation quantization is enabled or not.
         """
         return self._is_single_quant_mode(ActivationQuantizationMode.QUANT)
-
-    def is_fln_quantization(self) -> bool:
+    
+    def is_fln_no_quantization(self) -> bool:
         """
-        Returns: Whether the node's activation quantization is FLN
+        Returns: Whether node is FLN no quantization.
         """
-        return self._is_single_quant_mode(ActivationQuantizationMode.FLN_QUANT)
-
+        return self._is_single_quant_mode(ActivationQuantizationMode.FLN_NO_QUANT)
+    
     def is_quantization_preserving(self) -> bool:
         """
         Returns: Whether node activation quantization information is preserved from its inputs.
         """
         return self._is_single_quant_mode(ActivationQuantizationMode.PRESERVE_QUANT)
 
+    def is_no_quantization(self) -> bool:
+        """
+        Returns: Whether node is no quantization.
+        """
+        return self._is_single_quant_mode(ActivationQuantizationMode.NO_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_weights_quantization_enabled(self, attr_name: str) -> bool:
         """
         Checks whether a node's weights attribute quantization is enabled.

model_compression_toolkit/core/common/quantization/filter_nodes_candidates.py

@@ -21,7 +21,6 @@ from model_compression_toolkit.constants import FLOAT_BITWIDTH
 from model_compression_toolkit.core.common.quantization.candidate_node_quantization_config import \
     CandidateNodeQuantizationConfig
 
-
 def filter_nodes_candidates(graph: Graph):
     """
     Filters the graph's nodes candidates configuration list.
@@ -87,7 +86,7 @@ def filter_node_candidates(node: BaseNode) -> List[CandidateNodeQuantizationConf
     filtered_candidates = copy.deepcopy(node.candidates_quantization_cfg)
     final_candidates = copy.deepcopy(node.candidates_quantization_cfg)
 
-    if (node.kernel_attr is None or not node.is_weights_quantization_enabled(node.kernel_attr)) and not node.is_activation_quantization_enabled():
+    if (node.kernel_attr is None or not node.is_weights_quantization_enabled(node.kernel_attr)) and node.is_no_quantization():
         # If activation quantization is disabled and the node doesn't have a kernel or doesn't quantize the kernel,
         # but for some reason the node has multiple candidates then replace it with a single dummy candidate with
         # default bit-width values.
@@ -102,9 +101,10 @@ def filter_node_candidates(node: BaseNode) -> List[CandidateNodeQuantizationConf
 
         final_candidates = [single_dummy_candidate]
 
-    elif not node.is_activation_quantization_enabled():
+    elif node.is_no_quantization():
         # Remove candidates that have duplicated weights candidates for node with disabled activation quantization.
         # Replacing the activation n_bits in the remained configurations with default value to prevent confusion.
+        # Set the config of the non-quantized FLN node to POWER_OF_TWO.
         seen_candidates = set()
         filtered_candidates = [candidate for candidate in filtered_candidates if
                                candidate.weights_quantization_cfg not in seen_candidates
@@ -116,6 +116,14 @@ def filter_node_candidates(node: BaseNode) -> List[CandidateNodeQuantizationConf
 
         final_candidates = _filter_bit_method_dups(filtered_candidates, node.kernel_attr)
 
+    elif node.is_fln_no_quantization() or node.is_fln_quantization():
+        # Remove candidates that have duplicated weights candidates for node with disabled activation quantization.
+        seen_candidates = set()
+        filtered_candidates = [candidate for candidate in filtered_candidates if
+                               candidate.weights_quantization_cfg not in seen_candidates
+                               and not seen_candidates.add(candidate.weights_quantization_cfg)]
+        final_candidates = _filter_bit_method_dups(filtered_candidates, node.kernel_attr)
+
     elif node.kernel_attr is None or not node.is_weights_quantization_enabled(node.kernel_attr):
         # TODO:
         #  To allow MP on positional weights we need to modify this to consider all weights not only kernel.

model_compression_toolkit/core/common/quantization/node_quantization_config.py

@@ -47,7 +47,7 @@ class ActivationQuantizationMode(Enum):
     FLN_QUANT = auto()
     PRESERVE_QUANT = auto()
     NO_QUANT = auto()
-
+    FLN_NO_QUANT = auto()
 
 class BaseNodeQuantizationConfig(object):
     """

model_compression_toolkit/core/graph_prep_runner.py

@@ -155,7 +155,7 @@ def get_finalized_graph(initial_graph: Graph,
     ######################################
     fusing_info = FusingInfoGenerator(fqc.get_fusing_patterns()).generate_fusing_info(transformed_graph)
     transformed_graph.fusing_info = fusing_info
-    transformed_graph.disable_fused_nodes_activation_quantization()
+    transformed_graph.override_fused_node_activation_quantization_candidates()
 
     ######################################
     # Channel equalization