mct-nightly 2.3.0.20250513.611__py3-none-any.whl → 2.3.0.20250515.544__py3-none-any.whl

{mct_nightly-2.3.0.20250513.611.dist-info → mct_nightly-2.3.0.20250515.544.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.3.0.20250513.611
+Version: 2.3.0.20250515.544
 Summary: A Model Compression Toolkit for neural networks
 Author-email: ssi-dnn-dev@sony.com
 Classifier: Programming Language :: Python :: 3

{mct_nightly-2.3.0.20250513.611.dist-info → mct_nightly-2.3.0.20250515.544.dist-info}/RECORD

@@ -1,6 +1,6 @@
-mct_nightly-2.3.0.20250513.611.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-model_compression_toolkit/__init__.py,sha256=p_G6GkwHl_GiPtc0E2qL6iUBG-UpYcgFx1HDi073s0Q,1557
-model_compression_toolkit/constants.py,sha256=iJ6vfTjC2oFIZWt8wvHoxEw5YJi3yl0Hd4q30_8q0Zc,3958
+mct_nightly-2.3.0.20250515.544.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
+model_compression_toolkit/__init__.py,sha256=ZuiC7LBUZRbxQhR-vJI5NKeCIc9cX-tIpkHCw_Ynb0o,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
 model_compression_toolkit/metadata.py,sha256=x_Bk4VpzILdsFax6--CZ3X18qUTP28sbF_AhoQW8dNc,4003
@@ -31,7 +31,7 @@ model_compression_toolkit/core/common/collectors/min_max_per_channel_collector.p
 model_compression_toolkit/core/common/collectors/statistics_collector.py,sha256=psijsQZefwjMDH8SU5E18n65HiGtQilPhKr1hhzZX-I,8268
 model_compression_toolkit/core/common/collectors/weighted_histogram_collector.py,sha256=zp3dE7YTqWmkD5QWdRhsl9zD8W6Lr96G1Wjw1g2D3T0,4894
 model_compression_toolkit/core/common/fusion/__init__.py,sha256=Rf1RcYmelmdZmBV5qOKvKWF575ofc06JFQSq83Jz99A,696
-model_compression_toolkit/core/common/fusion/fusing_info.py,sha256=W8qZejLwbm-lkvNF3GepNL3ypO10vFRxOxbq-o_rt_I,15479
+model_compression_toolkit/core/common/fusion/fusing_info.py,sha256=S7hBbUJxL52Z8uJ9_upLdFyoSEJvgmVX0OmneqDIj-c,18656
 model_compression_toolkit/core/common/fusion/graph_fuser.py,sha256=F0AaAUBpJ9JjHMB5H2LD9pdwTSWJK-Kqm9dQmGHX1Jo,7368
 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=BSQpKy0BXoGX0G0bySTo72n2isTqvtpkbRYYa8-hPO4,41435
@@ -446,7 +446,7 @@ model_compression_toolkit/target_platform_capabilities/targetplatform2framework/
 model_compression_toolkit/target_platform_capabilities/targetplatform2framework/attach2pytorch.py,sha256=NCwuvnByeexLL987h67XhU8vQvCgq63bt0hFSiSSxvE,6400
 model_compression_toolkit/target_platform_capabilities/targetplatform2framework/attribute_filter.py,sha256=jfhszvuD2Fyy6W2KjlLzXBQKFzTqGAaDZeFVr4-ONQw,8776
 model_compression_toolkit/target_platform_capabilities/targetplatform2framework/current_tpc.py,sha256=_kFG0USYa6yzvLsi82_Vusv_KR8Hi7J1u680pPXECuo,2192
-model_compression_toolkit/target_platform_capabilities/targetplatform2framework/framework_quantization_capabilities.py,sha256=UKzckLYLdBcFAptyKnVMwpPpfRkmF0SK1Kl0g0eGjQA,9710
+model_compression_toolkit/target_platform_capabilities/targetplatform2framework/framework_quantization_capabilities.py,sha256=1jkj0ZO3t9M0SRpe9ZcSucraSoB4raezIbpcO_lZcP4,10084
 model_compression_toolkit/target_platform_capabilities/targetplatform2framework/framework_quantization_capabilities_component.py,sha256=9Hg6AMCzTdDsKKgivRd61UjxGT5SWvKsc3mIUPPsYDQ,1021
 model_compression_toolkit/target_platform_capabilities/targetplatform2framework/layer_filter_params.py,sha256=dIu6k1xvGKLtk_47wq1eKYvrS4lYAknAXTeJfFstW0Y,3878
 model_compression_toolkit/target_platform_capabilities/targetplatform2framework/operations_to_layers.py,sha256=vZ7I2XDr_YDgU8oQt8gKkcuUOJf28DCzCPunPK2h_Xw,6563
@@ -528,7 +528,7 @@ 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-2.3.0.20250513.611.dist-info/METADATA,sha256=dx0fsYTzsB_Y1IVuSNMaJPgPO4lhotb3TlDZ-dq2JF8,25136
-mct_nightly-2.3.0.20250513.611.dist-info/WHEEL,sha256=DnLRTWE75wApRYVsjgc6wsVswC54sMSJhAEd4xhDpBk,91
-mct_nightly-2.3.0.20250513.611.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
-mct_nightly-2.3.0.20250513.611.dist-info/RECORD,,
+mct_nightly-2.3.0.20250515.544.dist-info/METADATA,sha256=dV9aRBw1JVkuZDXyGl4aFtA91lLC_NtYTDquO5yA8rY,25136
+mct_nightly-2.3.0.20250515.544.dist-info/WHEEL,sha256=QZxptf4Y1BKFRCEDxD4h2V0mBFQOVFLFEpvxHmIs52A,91
+mct_nightly-2.3.0.20250515.544.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
+mct_nightly-2.3.0.20250515.544.dist-info/RECORD,,

{mct_nightly-2.3.0.20250513.611.dist-info → mct_nightly-2.3.0.20250515.544.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (80.4.0)
+Generator: setuptools (80.6.0)
 Root-Is-Purelib: true
 Tag: py3-none-any
 

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__ = "2.3.0.20250513.000611"
+__version__ = "2.3.0.20250515.000544"

model_compression_toolkit/constants.py

@@ -138,3 +138,8 @@ SHAPE = 'shape'
 NODE_NAME = 'node_name'
 TOTAL_SIZE = 'total_size'
 NODE_OUTPUT_INDEX = 'node_output_index'
+
+
+# Fusing Patterns constants
+FUSED_LAYER_PATTERN = 'fused_layer_pattern'
+FUSED_OP_QUANT_CONFIG = 'fused_op_quantization_config'

model_compression_toolkit/core/common/fusion/fusing_info.py

@@ -14,6 +14,8 @@
 #  ==============================================================================
 
 from model_compression_toolkit.target_platform_capabilities import LayerFilterParams
+from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import OpQuantizationConfig
+from model_compression_toolkit.constants import FUSED_LAYER_PATTERN, FUSED_OP_QUANT_CONFIG
 from dataclasses import dataclass, field
 
 from typing import Optional, List, Dict, Any, Tuple
@@ -41,6 +43,7 @@ class FusingInfo:
     fusing_patterns: any = None
     fusing_data: Dict[str, Tuple['BaseNode']] = field(default_factory=dict)
     node_to_fused_node_map: Dict[str, str] = field(init=False, default_factory=dict)
+    fused_op_id_to_quant_config: Dict[str, OpQuantizationConfig] = field(default_factory=dict)
 
     def __post_init__(self):
         """Validates and initializes mappings after dataclass instantiation."""
@@ -49,6 +52,7 @@ class FusingInfo:
             assert isinstance(op_nodes, tuple) and len(op_nodes) > 1, f"Found invalid fused op nodes: {op_nodes}"
 
         self._init_node_mapping()
+        self._init_quantization_config_map()
 
     def _init_node_mapping(self) -> None:
         """
@@ -59,6 +63,15 @@ class FusingInfo:
             for node in nodes:
                 self.node_to_fused_node_map[node.name] = op_id
 
+    def _init_quantization_config_map(self) -> None:
+        """
+        Init the mapping between fused operation IDs and their quantization configurations.
+        """
+        self.fused_op_id_to_quant_config.clear()
+        if self.fusing_patterns is not None:
+            for op_id, nodes in self.fusing_data.items():
+                self.set_fused_op_quantization_config(op_id, nodes)
+
     def add_fused_operation(self, op_id: str, nodes: Tuple['BaseNode']) -> None:
         """
         Add a new fused operation with the given ID and set of nodes.
@@ -78,6 +91,22 @@ class FusingInfo:
         for node in nodes:
             self.node_to_fused_node_map[node.name] = op_id
 
+        # Update the quantization config mapping for this operation
+        if self.fusing_patterns is not None:
+            self.set_fused_op_quantization_config(op_id, nodes)
+
+    def set_fused_op_quantization_config(self, op_id: str, nodes: Tuple['BaseNode']) -> None:
+        """
+        Set the quantization configuration for a given fused operation ID.
+
+        Args:
+            op_id (str): The identifier for the fused operation.
+            nodes (Tuple[BaseNode]): The tuple of nodes that form the fused operation.
+        """
+        fusing_pattern = next((fp for fp in self.fusing_patterns if is_valid_fusion([fp.get(FUSED_LAYER_PATTERN)], nodes)), None)
+        if fusing_pattern is not None:
+            self.fused_op_id_to_quant_config[op_id] = fusing_pattern.get(FUSED_OP_QUANT_CONFIG)
+
     def remove_fused_operation(self, op_id: str) -> None:
         """
         Remove a fused operation by its ID.
@@ -95,6 +124,7 @@ class FusingInfo:
         for node in nodes:
             self.node_to_fused_node_map.pop(node.name, None)
         del self.fusing_data[op_id]
+        self.fused_op_id_to_quant_config.pop(op_id, None)
 
     def get_fused_node_name(self, node_name: str) -> Optional[str]:
         """
@@ -117,6 +147,15 @@ class FusingInfo:
         """
         return self.node_to_fused_node_map.copy()
 
+    def get_fusing_quantization_config_map(self) -> Dict[str, OpQuantizationConfig]:
+        """
+        Retrieve a copy of the mapping from fused operation IDs to their quantization configurations.
+
+        Returns:
+            A dictionary mapping each fused operation ID to its quantization configuration.
+        """
+        return self.fused_op_id_to_quant_config.copy()
+
     def get_fused_nodes(self, op_id: str) -> Optional[List['BaseNode']]:
         """
         Retrieve the list of nodes for a given fused operation ID.
@@ -129,6 +168,18 @@ class FusingInfo:
         """
         return self.fusing_data.get(op_id)
 
+    def get_fused_op_quantization_config(self, op_id: str) -> OpQuantizationConfig:
+        """
+        Retrieve the quantization configuration for a given fused operation ID.
+
+        Args:
+            op_id (str): The identifier for the fused operation.
+
+        Returns:
+            OpQuantizationConfig: The quantization configuration for the operation, or None if not found.
+        """
+        return self.fused_op_id_to_quant_config.get(op_id)
+
     def is_node_in_fused_op(self, node: 'BaseNode') -> bool:
         """
         Check if a node is part of any fused operation.
@@ -216,10 +267,11 @@ class FusingInfo:
             all_fused_nodes.update(node_set)
 
             # Check 4: Ensure the sequence matches a valid fusing pattern
-            if not is_valid_fusion(self.fusing_patterns, nodes):
+            valid_fusing_patterns = _get_fusing_layer_patterns(self.fusing_patterns)
+            if not is_valid_fusion(valid_fusing_patterns, nodes):
                 raise ValueError(
                     f"Fused operation {op_id} does not match any valid fusing pattern "
-                    f"from {self.fusing_patterns}."
+                    f"from {valid_fusing_patterns}."
                 )
 
     def is_nodes_eligible_to_be_fused(self, nodes: List['BaseNode']) -> bool:
@@ -240,7 +292,8 @@ class FusingInfo:
             return False
 
         # Check if the provided nodes match a valid fusion pattern
-        return is_valid_fusion(fusing_patterns=self.fusing_patterns, nodes=nodes)
+        valid_fusing_patterns = _get_fusing_layer_patterns(self.fusing_patterns)
+        return is_valid_fusion(fusing_patterns=valid_fusing_patterns, nodes=nodes)
 
     def __repr__(self) -> str:
         """
@@ -287,8 +340,11 @@ class FusingInfoGenerator:
         if not self._fusing_patterns:
             return FusingInfo(fusing_patterns=self._fusing_patterns)
 
+        # Extract fusing layer patterns
+        fusing_layer_patterns = _get_fusing_layer_patterns(self._fusing_patterns)
+
         # Find max fusion
-        max_layers_fusing = max([len(fusing_pattern) for fusing_pattern in self._fusing_patterns])
+        max_layer_patterns = max([len(fusing_layer_pattern) for fusing_layer_pattern in fusing_layer_patterns])
 
         # Travel along the graph to find layers for fusing
         nodes = graph.get_topo_sorted_nodes()
@@ -302,9 +358,9 @@ class FusingInfoGenerator:
                 continue
             # Start fusing search
             fusing_nodes = []  # nodes that are candidates for participating in fusing
-            patterns = copy.deepcopy(self._fusing_patterns)
+            patterns = copy.deepcopy(fusing_layer_patterns)
             next_nodes = [node]
-            for i in range(max_layers_fusing):
+            for i in range(max_layer_patterns):
                 patterns = get_valid_fusing_patterns_for_node(patterns, next_nodes[0], i)
                 if len(patterns) == 0:  # Give up if no more fusion pattern
                     break
@@ -314,7 +370,7 @@ class FusingInfoGenerator:
                     break
 
             # New fusion
-            if is_valid_fusion(self._fusing_patterns, fusing_nodes):
+            if is_valid_fusion(fusing_layer_patterns, fusing_nodes):
                 fused_op_id = FusingInfo.generate_fused_op_id(fusing_nodes)
                 assert fused_op_id not in fusing_info, f"{fused_op_id} is already in fusing info: {fusing_info}"
                 fusing_info[fused_op_id] = tuple(fusing_nodes)
@@ -371,3 +427,15 @@ def is_valid_fusion(fusing_patterns: List[List[Any]], nodes: List['BaseNode']) -
         if counter == fusion_depth:
             return True
     return False
+
+
+def _get_fusing_layer_patterns(fusing_patterns: List[Dict[Any, OpQuantizationConfig]]) -> List[List[Any]]:
+    """
+    Extracts the fusing layer patterns from the provided fusing patterns.
+    Args:
+        fusing_patterns: List of patterns of layers/LayerFilterParams to fuse and their mapping quantization config.
+
+    Returns:
+        supported fusing layer patterns
+    """
+    return [f.get(FUSED_LAYER_PATTERN) for f in fusing_patterns]

model_compression_toolkit/target_platform_capabilities/targetplatform2framework/framework_quantization_capabilities.py

@@ -31,6 +31,9 @@ from model_compression_toolkit.target_platform_capabilities.schema.mct_current_s
     OpQuantizationConfig, QuantizationConfigOptions
 from model_compression_toolkit.target_platform_capabilities.targetplatform2framework.current_tpc import  _current_tpc
 
+from model_compression_toolkit.constants import FUSED_LAYER_PATTERN, FUSED_OP_QUANT_CONFIG
+
+
 class FrameworkQuantizationCapabilities(ImmutableClass):
     """
     Attach framework information to a modeled hardware.
@@ -94,20 +97,26 @@ class FrameworkQuantizationCapabilities(ImmutableClass):
         """
         return self.op_sets_to_layers.get_layers_by_op(op)
 
-    def get_fusing_patterns(self) -> List[List[Any]]:
+    def get_fusing_patterns(self) -> List[Dict[List[Any], OpQuantizationConfig]]:
         """
 
-        Returns: List of patterns of layers/LayerFilterParams to fuse.
+        Returns: List of patterns of layers/LayerFilterParams to fuse and their mapping quantization config.
 
         """
-        res = []
+
+        patterns = []
         if self.tpc.fusing_patterns is None:
-            return res
+            return patterns
+
         for p in self.tpc.fusing_patterns:
+            res = []
             ops = [self.get_layers_by_opset(x) for x in p.operator_groups]
             res.extend(itertools.product(*ops))
-        return [list(x) for x in res]
 
+            fused_op_quant_config = getattr(p, FUSED_OP_QUANT_CONFIG, None)
+            patterns.extend({FUSED_LAYER_PATTERN: list(x), FUSED_OP_QUANT_CONFIG: fused_op_quant_config} for x in res)
+
+        return patterns
 
     def get_info(self) -> Dict[str, Any]:
         """