Add functionality to FusingInfo class to preserve quantization config

model_compression_toolkit/constants.py

@@ -138,3 +138,8 @@
 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)
+    fusing_data_to_quantization_config_map: Dict[str, OpQuantizationConfig] = field(default_factory=dict)
 
     def __post_init__(self):
         """Validates and initializes mappings after dataclass instantiation."""
@@ -49,6 +52,7 @@ def __post_init__(self):
             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 +63,17 @@ def _init_node_mapping(self) -> None:
             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.fusing_data_to_quantization_config_map.clear()
+        if self.fusing_patterns is not None:
+            for op_id, nodes in self.fusing_data.items():
+                for fusing_pattern in self.fusing_patterns:
+                    if is_valid_fusion([fusing_pattern], nodes):
+                        self.fusing_data_to_quantization_config_map[op_id] = fusing_pattern.get(FUSED_OP_QUANT_CONFIG)
+
     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 +93,12 @@ def add_fused_operation(self, op_id: str, nodes: Tuple['BaseNode']) -> None:
         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:
+            for fusing_pattern in self.fusing_patterns:
+                if is_valid_fusion([fusing_pattern], nodes):
+                    self.fusing_data_to_quantization_config_map[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 +116,7 @@ def remove_fused_operation(self, op_id: str) -> None:
         for node in nodes:
             self.node_to_fused_node_map.pop(node.name, None)
         del self.fusing_data[op_id]
+        del self.fusing_data_to_quantization_config_map[op_id]
 
     def get_fused_node_name(self, node_name: str) -> Optional[str]:
         """
@@ -117,6 +139,15 @@ def get_node_to_fused_node_map(self) -> Dict[str, str]:
         """
         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.fusing_data_to_quantization_config_map.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 +160,18 @@ def get_fused_nodes(self, op_id: str) -> Optional[List['BaseNode']]:
         """
         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.fusing_data_to_quantization_config_map.get(op_id)
+
     def is_node_in_fused_op(self, node: 'BaseNode') -> bool:
         """
         Check if a node is part of any fused operation.
@@ -284,11 +327,13 @@ def generate_fusing_info(self, graph: 'Graph') -> FusingInfo:
             - Fusions are linear sequences (each node has exactly one successor).
             - Each node belongs to at most one fused operation.
         """
-        if not self._fusing_patterns:
+        fusing_layer_patterns = [f.get(FUSED_LAYER_PATTERN) for f in self._fusing_patterns]
+
+        if not fusing_layer_patterns:
             return FusingInfo(fusing_patterns=self._fusing_patterns)
 
         # Find max fusion
-        max_layers_fusing = max([len(fusing_pattern) for fusing_pattern in self._fusing_patterns])
+        max_fusing_layer_patterns = max([len(fusing_pattern.get(FUSED_LAYER_PATTERN)) for fusing_pattern in self._fusing_patterns])
 
         # Travel along the graph to find layers for fusing
         nodes = graph.get_topo_sorted_nodes()
@@ -302,9 +347,9 @@ def generate_fusing_info(self, graph: 'Graph') -> FusingInfo:
                 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_fusing_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
@@ -361,10 +406,11 @@ def is_valid_fusion(fusing_patterns: List[List[Any]], nodes: List['BaseNode']) -
     if fusion_depth <= 1:
         return False
     for fusing_pattern in fusing_patterns:
-        if fusion_depth != len(fusing_pattern):
+        fusing_layer_pattern = fusing_pattern.get(FUSED_LAYER_PATTERN)
+        if fusion_depth != len(fusing_layer_pattern):
             continue
         counter = 0
-        for i, layer in enumerate(fusing_pattern):
+        for i, layer in enumerate(fusing_layer_pattern):
             if (type(layer) == LayerFilterParams and nodes[i].is_match_filter_params(layer)) or \
                     nodes[i].is_match_type(layer):
                 counter += 1

model_compression_toolkit/target_platform_capabilities/targetplatform2framework/framework_quantization_capabilities.py

@@ -31,6 +31,9 @@
     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.
@@ -97,17 +100,30 @@ def get_layers_by_opset(self, op: OperatorsSetBase) -> List[Any]:
     def get_fusing_patterns(self) -> List[List[Any]]:
         """
 
-        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]
 
+            if hasattr(p, 'fused_op_quantization_config'):
+                fused_op_quantization_config = p.fused_op_quantization_config
+            else:
+                fused_op_quantization_config = None
+
+            for x in res:
+                pattern = {FUSED_LAYER_PATTERN: list(x),
+                           FUSED_OP_QUANT_CONFIG: fused_op_quantization_config}
+                patterns.append(pattern)
+
+        return patterns
 
     def get_info(self) -> Dict[str, Any]:
         """

tests/keras_tests/non_parallel_tests/test_keras_tpc.py

@@ -42,7 +42,7 @@
     from keras import Input
 
 import model_compression_toolkit as mct
-from model_compression_toolkit.constants import TENSORFLOW
+from model_compression_toolkit.constants import TENSORFLOW, FUSED_LAYER_PATTERN, FUSED_OP_QUANT_CONFIG
 from model_compression_toolkit.target_platform_capabilities.constants import DEFAULT_TP_MODEL, IMX500_TP_MODEL, \
     QNNPACK_TP_MODEL, TFLITE_TP_MODEL, KERNEL_ATTR, BIAS_ATTR, KERAS_KERNEL, BIAS, WEIGHTS_N_BITS
 from model_compression_toolkit.core.keras.keras_implementation import KerasImplementation
@@ -250,7 +250,7 @@ def test_keras_fusing_patterns(self):
 
         fusings = hm_keras.get_fusing_patterns()
         self.assertEqual(len(fusings), 2)
-        p0, p1 = fusings[0], fusings[1]
+        p0, p1 = fusings[0].get(FUSED_LAYER_PATTERN), fusings[1].get(FUSED_LAYER_PATTERN)
 
         self.assertEqual(len(p0), 3)
         self.assertEqual(p0[0], Conv2D)

tests/pytorch_tests/function_tests/test_pytorch_tpc.py

@@ -26,7 +26,7 @@
 import model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema as schema
 from model_compression_toolkit.core import MixedPrecisionQuantizationConfig
 from model_compression_toolkit.defaultdict import DefaultDict
-from model_compression_toolkit.constants import PYTORCH
+from model_compression_toolkit.constants import PYTORCH, FUSED_LAYER_PATTERN, FUSED_OP_QUANT_CONFIG
 from model_compression_toolkit.core.common import BaseNode
 from model_compression_toolkit.target_platform_capabilities.constants import DEFAULT_TP_MODEL, IMX500_TP_MODEL, \
     TFLITE_TP_MODEL, QNNPACK_TP_MODEL, KERNEL_ATTR, WEIGHTS_N_BITS, PYTORCH_KERNEL, BIAS_ATTR, BIAS
@@ -236,15 +236,15 @@ def test_pytorch_fusing_patterns(self):
                                                fusing_patterns=tuple(fusing_patterns),
                                                add_metadata=False)
 
-        hm_keras = FrameworkQuantizationCapabilities(hm)
-        with hm_keras:
+        hm_torch = FrameworkQuantizationCapabilities(hm)
+        with hm_torch:
             OperationsSetToLayers("opA", [torch.conv2d])
             OperationsSetToLayers("opB", [torch.tanh])
             OperationsSetToLayers("opC", [LayerFilterParams(torch.relu, Greater("max_value", 7), negative_slope=0)])
 
-        fusings = hm_keras.get_fusing_patterns()
+        fusings = hm_torch.get_fusing_patterns()
         self.assertEqual(len(fusings), 2)
-        p0, p1 = fusings[0], fusings[1]
+        p0, p1 = fusings[0].get(FUSED_LAYER_PATTERN), fusings[1].get(FUSED_LAYER_PATTERN)
 
         self.assertEqual(len(p0), 3)
         self.assertEqual(p0[0], torch.conv2d)