mct-nightly 2.3.0.20250331.610__py3-none-any.whl → 2.3.0.20250401.618__py3-none-any.whl

{mct_nightly-2.3.0.20250331.610.dist-info → mct_nightly-2.3.0.20250401.618.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.3.0.20250331.610
+Version: 2.3.0.20250401.618
 Summary: A Model Compression Toolkit for neural networks
 Classifier: Programming Language :: Python :: 3
 Classifier: License :: OSI Approved :: Apache Software License

{mct_nightly-2.3.0.20250331.610.dist-info → mct_nightly-2.3.0.20250401.618.dist-info}/RECORD

@@ -1,5 +1,5 @@
-mct_nightly-2.3.0.20250331.610.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-model_compression_toolkit/__init__.py,sha256=tulTPjSjSee0ySdxccC46EZ_AbQkEIr3RvtUps-6IME,1557
+mct_nightly-2.3.0.20250401.618.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
+model_compression_toolkit/__init__.py,sha256=sgIxUWX9jDSvUZnwMqs3nHNjXfhgFSfniDDr2vvRTuQ,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
@@ -34,8 +34,8 @@ model_compression_toolkit/core/common/fusion/__init__.py,sha256=Rf1RcYmelmdZmBV5
 model_compression_toolkit/core/common/fusion/graph_fuser.py,sha256=b41_4rL_Adiza4vpWlmmqgvkpUmWVdfdx0nEIB0p2n8,6195
 model_compression_toolkit/core/common/fusion/layer_fusing.py,sha256=-2fnjyC9q2RPw9st6RxROW-gdtT2mSRz0QZ_Gz1KDz4,5579
 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=VhniLTiMqL7i1Vqg2UBQuFFTvw2cYeJayssUJwabp3E,38112
-model_compression_toolkit/core/common/graph/base_node.py,sha256=kZbmAMh5cPAwYzlY8KYa8w0ipL58yApB09-WXQ8plrE,33763
+model_compression_toolkit/core/common/graph/base_graph.py,sha256=cSwHUqwZEiR1t2DaBfc7_qSJbtX8crpqerN4ol9v3H8,38859
+model_compression_toolkit/core/common/graph/base_node.py,sha256=CJu8_r80MGVnYmlAUGOGKGRsD9xShMyaRNb3VMeRC0s,34523
 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
@@ -75,10 +75,10 @@ model_compression_toolkit/core/common/mixed_precision/set_layer_to_bitwidth.py,s
 model_compression_toolkit/core/common/mixed_precision/solution_refinement_procedure.py,sha256=fk7PWiZ6Na5O_Z_dymk_UfDCTqW_X_4EROU7DZknQnc,9444
 model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/__init__.py,sha256=Rf1RcYmelmdZmBV5qOKvKWF575ofc06JFQSq83Jz99A,696
 model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization.py,sha256=PKkhc5q8pEPnNLXwo3U56EOCfYnPXIvPs0LlCGZOoKU,4426
-model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_calculator.py,sha256=xCYL36K0nK41VSsLcy52uDA7zVfoLxhubmOrtXbqw7s,39140
+model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_calculator.py,sha256=MP4Q5lThvEIhfa1iBajQQM3nCUNgK-2yseqQQ8Rgiog,40624
 model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_data.py,sha256=QQwtl08DiDxUOQGpYPnek_RlZjWm1Ky7tL2ESHXMK78,4050
 model_compression_toolkit/core/common/mixed_precision/search_methods/__init__.py,sha256=sw7LOPN1bM82o3SkMaklyH0jw-TLGK0-fl2Wq73rffI,697
-model_compression_toolkit/core/common/mixed_precision/search_methods/linear_programming.py,sha256=rSSN5MhH5BO5b58d8pe2pY9wc5HbfescoUStfg-nWfk,7263
+model_compression_toolkit/core/common/mixed_precision/search_methods/linear_programming.py,sha256=TaK5NqVdmygsHw9_x5JsJ-BPvlbKA9cRyTno1R8gbnU,7269
 model_compression_toolkit/core/common/network_editors/__init__.py,sha256=vZmu55bYqiaOQs3AjfwWDXHmuKZcLHt-wm7uR5fPEqg,1307
 model_compression_toolkit/core/common/network_editors/actions.py,sha256=nid0_j-Cn10xvmztT8yCKW_6uA7JEnom9SW9syx7wc0,19594
 model_compression_toolkit/core/common/network_editors/edit_network.py,sha256=dfgawi-nB0ocAJ0xcGn9E-Zv203oUnQLuMiXpX8vTgA,1748
@@ -526,7 +526,7 @@ 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-2.3.0.20250331.610.dist-info/METADATA,sha256=BSWkSPN58Xfzc8eImgBxCG3v_wXYt8c7l8oFNOOoyGw,27098
-mct_nightly-2.3.0.20250331.610.dist-info/WHEEL,sha256=CmyFI0kx5cdEMTLiONQRbGQwjIoR1aIYB7eCAQ4KPJ0,91
-mct_nightly-2.3.0.20250331.610.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
-mct_nightly-2.3.0.20250331.610.dist-info/RECORD,,
+mct_nightly-2.3.0.20250401.618.dist-info/METADATA,sha256=53LoSDV2ox7X64SeEb9OwP4UsuLi75QeSwyFLponCrQ,27098
+mct_nightly-2.3.0.20250401.618.dist-info/WHEEL,sha256=CmyFI0kx5cdEMTLiONQRbGQwjIoR1aIYB7eCAQ4KPJ0,91
+mct_nightly-2.3.0.20250401.618.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
+mct_nightly-2.3.0.20250401.618.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__ = "2.3.0.20250331.000610"
+__version__ = "2.3.0.20250401.000618"

model_compression_toolkit/core/common/graph/base_graph.py

@@ -696,6 +696,23 @@ class Graph(nx.MultiDiGraph, GraphSearches):
         sorted_conf_activation = self.get_sorted_activation_configurable_nodes()
         return [(n, n.final_activation_quantization_cfg.activation_n_bits) for n in sorted_conf_activation]
 
+    def retrieve_preserved_quantization_node(self, node: BaseNode) -> BaseNode:
+        """
+        For a node with quantization_preserving == True, get the previous non-quantization_preserving node
+        to get activation quantization config from. If quantization_preserving is False return node.
+        Args:
+            node: quantization preserving node.
+
+        Returns:
+            The node that the quantization preserving node should get the activation quantization from.
+
+        """
+        while node.is_quantization_preserving():
+            prev_nodes = self.get_prev_nodes(node)
+            assert len(prev_nodes) == 1, "Activation preserving node should have only 1 input."
+            node = prev_nodes[0]
+        return node
+
     def update_fused_nodes(self, fusion: List[Any]):
         """
         Updates the graphs fusions list with a new list of nodes that have been fused.

model_compression_toolkit/core/common/graph/base_node.py

@@ -131,6 +131,19 @@ class BaseNode:
                    qc.activation_quantization_cfg.enable_activation_quantization
         return self.candidates_quantization_cfg[0].activation_quantization_cfg.enable_activation_quantization
 
+    def is_quantization_preserving(self) -> bool:
+        """
+        Returns: Whether node activation quantization information is preserved from its inputs.
+        """
+        if self.final_activation_quantization_cfg:
+            # if we have a final configuration, then we only care to check if it enables activation quantization.
+            return self.final_activation_quantization_cfg.quantization_preserving
+
+        for qc in self.candidates_quantization_cfg:
+            assert self.candidates_quantization_cfg[0].activation_quantization_cfg.quantization_preserving == \
+                   qc.activation_quantization_cfg.quantization_preserving
+        return self.candidates_quantization_cfg[0].activation_quantization_cfg.quantization_preserving
+
     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/mixed_precision/resource_utilization_tools/resource_utilization_calculator.py

@@ -335,13 +335,35 @@ class ResourceUtilizationCalculator:
         """
         return self.compute_activation_utilization_by_cut(target_criterion, bitwidth_mode, act_qcs)
 
+    def _extract_qc(self, n: BaseNode, act_qcs: Optional[ActivationQCfgPerNode] = None
+                    ) -> Union[NodeActivationQuantizationConfig, None]:
+        """
+        Extract quantization config the activation configs dictionary is provided. If node is quantization
+        preserving, extract the quantization config from the preceding activation quantized node (i.e.
+        the Quantization the original node preserves).
+
+        Args:
+            n: Node to extract qc for.
+            act_qcs: custom activations quantization configuration. If not provided, the default
+             configuration will be extracted from the node.
+
+        Returns:
+            The relevant quantization config.
+        """
+        if act_qcs:
+            assert not (n.is_quantization_preserving() and act_qcs.get(n.name) is not None), \
+                f"Quantization preserving node {n.name} should not have a qc for this computation."
+            return act_qcs.get(self.graph.retrieve_preserved_quantization_node(n).name)
+        return None
+
     def compute_activation_utilization_by_cut(self,
                                               target_criterion: TargetInclusionCriterion,
                                               bitwidth_mode: BitwidthMode,
                                               act_qcs: Optional[ActivationQCfgPerNode] = None) \
             -> Tuple[float, Dict[Cut, Utilization], Dict[Cut, Dict[BaseNode, Utilization]]]:
         """
-        Compute graph activation cuts utilization.
+        Compute graph activation cuts utilization. If activation quantization configs are provided, then for
+        quantization preserving nodes, get the previous quantized activation node bit-width.
 
         Args:
             target_criterion: criterion to include weights for computation.
@@ -369,7 +391,7 @@ class ResourceUtilizationCalculator:
             if not cut_target_nodes:
                 continue
             for n in cut_target_nodes:
-                qc = act_qcs.get(n.name) if act_qcs else None
+                qc = self._extract_qc(n, act_qcs)
                 util_per_cut_per_node[cut][n.name] = self.compute_node_activation_tensor_utilization(n, target_criterion,
                                                                                                      bitwidth_mode, qc)
             util_per_cut[cut] = sum(util_per_cut_per_node[cut].values())    # type: ignore
@@ -384,7 +406,8 @@ class ResourceUtilizationCalculator:
                                                include_reused=False) \
             -> Tuple[float, Dict[NodeName, Utilization]]:
         """
-        Compute resource utilization for graph's activations tensors.
+        Compute resource utilization for graph's activations tensors. If activation quantization configs are provided, then for
+        quantization preserving nodes, get the previous quantized activation node bit-width.
 
         Args:
             target_criterion: criterion to include weights for computation.
@@ -405,7 +428,7 @@ class ResourceUtilizationCalculator:
 
         util_per_node: Dict[NodeName, Utilization] = {}
         for n in self._topo_sort(nodes):
-            qc = act_qcs.get(n.name) if act_qcs else None
+            qc = self._extract_qc(n, act_qcs)
             util = self.compute_node_activation_tensor_utilization(n, None, bitwidth_mode, qc)
             util_per_node[n.name] = util
 
@@ -659,7 +682,7 @@ class ResourceUtilizationCalculator:
         if target_criterion == TargetInclusionCriterion.QConfigurable:
             nodes = [n for n in nodes if n.has_configurable_activation()]
         elif target_criterion == TargetInclusionCriterion.AnyQuantized:
-            nodes = [n for n in nodes if n.is_activation_quantization_enabled()]
+            nodes = [n for n in nodes if n.is_activation_quantization_enabled() or n.is_quantization_preserving()]
         elif target_criterion == TargetInclusionCriterion.QNonConfigurable:
             nodes = [n for n in nodes if n.is_activation_quantization_enabled() and not n.has_configurable_activation()]
         elif target_criterion != TargetInclusionCriterion.Any:    # pragma: no cover
@@ -668,8 +691,7 @@ class ResourceUtilizationCalculator:
             nodes = [n for n in nodes if not n.reuse]
         return nodes
 
-    @classmethod
-    def _get_activation_nbits(cls,
+    def _get_activation_nbits(self,
                               n: BaseNode,
                               bitwidth_mode: BitwidthMode,
                               act_qc: Optional[NodeActivationQuantizationConfig]) -> int:
@@ -690,21 +712,22 @@ class ResourceUtilizationCalculator:
             assert bitwidth_mode == BitwidthMode.QCustom
             return act_qc.activation_n_bits if act_qc.enable_activation_quantization else FLOAT_BITWIDTH
 
-        if bitwidth_mode == BitwidthMode.Float or not n.is_activation_quantization_enabled():
+        if bitwidth_mode == BitwidthMode.Float or not (n.is_activation_quantization_enabled() or
+                                                       n.is_quantization_preserving()):
             return FLOAT_BITWIDTH
 
         if bitwidth_mode == BitwidthMode.Q8Bit:
             return 8
 
-        if bitwidth_mode in cls._bitwidth_mode_fn:
+        if bitwidth_mode in self._bitwidth_mode_fn:
             candidates_nbits = [c.activation_quantization_cfg.activation_n_bits for c in n.candidates_quantization_cfg]
-            return cls._bitwidth_mode_fn[bitwidth_mode](candidates_nbits)
+            return self._bitwidth_mode_fn[bitwidth_mode](candidates_nbits)
 
         if bitwidth_mode in [BitwidthMode.QCustom, BitwidthMode.QDefaultSP]:
-            qcs = n.get_unique_activation_candidates()
+            qcs = self.graph.retrieve_preserved_quantization_node(n).get_unique_activation_candidates()
             if len(qcs) != 1:
                 raise ValueError(f'Could not retrieve the activation quantization candidate for node {n} '
-                                 f'as it has {len(qcs)}!=1 unique candidates .')
+                                 f'as it has {len(qcs)}!=1 unique candidates.')
             return qcs[0].activation_quantization_cfg.activation_n_bits
 
         raise ValueError(f'Unknown mode {bitwidth_mode}')    # pragma: no cover

model_compression_toolkit/core/common/mixed_precision/search_methods/linear_programming.py

@@ -14,7 +14,7 @@
 # ==============================================================================
 import numpy as np
 from pulp import *
-from typing import Dict, Tuple
+from typing import Dict, Tuple, List
 
 from model_compression_toolkit.core.common.mixed_precision.resource_utilization_tools.resource_utilization import RUTarget