mct-nightly 2.4.0.20250617.613__py3-none-any.whl → 2.4.0.20250619.621__py3-none-any.whl

model_compression_toolkit/core/common/pruning/memory_calculator.py

@@ -34,18 +34,16 @@ class MemoryCalculator:
     which is crucial for deploying models on memory-constrained devices or optimizing for computational efficiency.
     """
 
-    def __init__(self, graph: Graph, fw_info: FrameworkInfo, fw_impl: PruningFrameworkImplementation):
+    def __init__(self, graph: Graph, fw_impl: PruningFrameworkImplementation):
         """
         Initializes the MemoryCalculator with necessary information about the model's graph,
         framework-specific details, and pruning implementation.
 
         Args:
             graph (Graph): Computational graph of the model.
-            fw_info (FrameworkInfo): Contains framework-specific information.
             fw_impl (PruningFrameworkImplementation): Implementation details for pruning.
         """
         self.graph = graph
-        self.fw_info = fw_info
         self.fw_impl = fw_impl
 
     def get_pruned_graph_memory(self,
@@ -204,19 +202,13 @@ class MemoryCalculator:
             if node == section.exit_node:
                 return masks.get(section.entry_node)
 
-        kernel_attr = self.fw_info.get_kernel_op_attributes(node.type)
-        # Ensure only one kernel attribute exists for the given node.
-        if len(kernel_attr) != 1:
-            Logger.critical(f"Expected a single attribute, but found {len(kernel_attr)} attributes for node '{node}'. Ensure the node configuration is correct.")
-        kernel_attr = kernel_attr[0]
-
         # Retrieve and validate the axis index for the output channels.
-        _, ic_axis = self.fw_info.kernel_channels_mapping.get(node.type)
+        ic_axis = node.channel_axis.input
         if ic_axis is None or int(ic_axis) != ic_axis:
             Logger.critical(f"Invalid input channel axis type for node '{node}': expected integer but got '{ic_axis}'.")
 
         # Get the number of output channels based on the kernel attribute and axis.
-        num_ic = node.get_weights_by_keys(kernel_attr).shape[ic_axis]
+        num_ic = node.get_weights_by_keys(node.kernel_attr).shape[ic_axis]
         mask = np.ones(num_ic, dtype=bool)
         return mask
 
@@ -289,7 +281,7 @@ class MemoryCalculator:
             int: The total number of parameters in the node after pruning.
         """
         total_params = 0
-        attributes_and_oc_axis = self.fw_impl.attrs_oi_channels_info_for_pruning(node, self.fw_info)
+        attributes_and_oc_axis = self.fw_impl.attrs_oi_channels_info_for_pruning(node)
 
         # Iterate over the node's weights and apply pruning based on the masks.
         for w_attr, w in node.weights.items():
@@ -311,7 +303,7 @@ class MemoryCalculator:
             num_oc = np.sum(output_mask)
         else:
             # Get the node channel axis from framework info
-            channel_axis = self.fw_info.out_channel_axis_mapping.get(node.type)
+            channel_axis = node.out_channel_axis
             if channel_axis is None:
                 Logger.critical(f"The channel axis is undefined. Please ensure the channel axis is explicitly defined for node {node.type} in the framework info.")
 

model_compression_toolkit/core/common/pruning/prune_graph.py

@@ -27,7 +27,6 @@ from model_compression_toolkit.logger import Logger
 
 def build_pruned_graph(graph: Graph,
                        masks: Dict[BaseNode, np.ndarray],
-                       fw_info: FrameworkInfo,
                        fw_impl: FrameworkImplementation) -> Graph:
     """
     Prunes the provided graph according to the given pruning output-channels masks.
@@ -35,7 +34,6 @@ def build_pruned_graph(graph: Graph,
     Args:
         graph: The original computational graph to be pruned.
         masks: A dictionary mapping each prunable node to its pruning mask.
-        fw_info: Framework-specific information object.
         fw_impl: Framework-specific implementation object.
 
     Returns:
@@ -66,8 +64,7 @@ def build_pruned_graph(graph: Graph,
             section_mask = PruningSectionMask(entry_node_oc_mask=mask,
                                               exit_node_ic_mask=mask)
             pruning_section.apply_inner_section_mask(section_mask,
-                                                     fw_impl,
-                                                     fw_info)
+                                                     fw_impl)
 
     return graph_to_prune
 

model_compression_toolkit/core/common/pruning/pruner.py

@@ -40,7 +40,6 @@ class Pruner:
     """
     def __init__(self,
                  float_graph: Graph,
-                 fw_info: FrameworkInfo,
                  fw_impl: PruningFrameworkImplementation,
                  target_resource_utilization: ResourceUtilization,
                  representative_data_gen: Callable,
@@ -49,7 +48,6 @@ class Pruner:
         """
         Args:
             float_graph (Graph): The floating-point representation of the model's computation graph.
-            fw_info (FrameworkInfo): Contains metadata and helper functions for the framework.
             fw_impl (PruningFrameworkImplementation): Implementation of specific framework methods required for pruning.
             target_resource_utilization (ResourceUtilization): The target resource utilization to be achieved after pruning.
             representative_data_gen (Callable): Generator function for representative dataset used in pruning analysis.
@@ -57,7 +55,6 @@ class Pruner:
             target_platform_capabilities (FrameworkQuantizationCapabilities): Object encapsulating the capabilities of the target hardware platform.
         """
         self.float_graph = float_graph
-        self.fw_info = fw_info
         self.fw_impl = fw_impl
         self.target_resource_utilization = target_resource_utilization
         self.representative_data_gen = representative_data_gen
@@ -84,7 +81,6 @@ class Pruner:
         # Apply Greedy strategy to compute masks based on importance scores.
         if self.pruning_config.channels_filtering_strategy == ChannelsFilteringStrategy.GREEDY:
             mask_calculator = GreedyMaskCalculator(entry_nodes,
-                                                   self.fw_info,
                                                    self.simd_scores,
                                                    self.target_resource_utilization,
                                                    self.float_graph,
@@ -99,7 +95,6 @@ class Pruner:
         Logger.info("Start pruning graph...")
         _pruned_graph = build_pruned_graph(self.float_graph,
                                            self.per_oc_mask,
-                                           self.fw_info,
                                            self.fw_impl)
         return _pruned_graph
 
@@ -116,7 +111,7 @@ class Pruner:
         # Retrieve and initialize the importance metric.
         im = get_importance_metric(self.pruning_config.importance_metric, graph=self.float_graph,
                                    representative_data_gen=self.representative_data_gen, fw_impl=self.fw_impl,
-                                   pruning_config=self.pruning_config, fw_info=self.fw_info)
+                                   pruning_config=self.pruning_config)
         entry_node_to_simd_score, simd_groups_indices = im.get_entry_node_to_simd_score(entry_nodes)
         return entry_node_to_simd_score, simd_groups_indices
 

model_compression_toolkit/core/common/pruning/pruning_framework_implementation.py

@@ -28,15 +28,13 @@ class PruningFrameworkImplementation(FrameworkImplementation):
     @abstractmethod
     def prune_entry_node(self,
                          node: BaseNode,
-                         output_mask: np.ndarray,
-                         fw_info: FrameworkInfo):
+                         output_mask: np.ndarray):
         """
         Abstract method to prune an entry node in the model.
 
         Args:
             node: The node to be pruned.
             output_mask: A numpy array representing the mask to be applied to the output channels.
-            fw_info: Framework-specific information.
 
         Raises:
             NotImplemented: If the method is not implemented in the subclass.
@@ -48,8 +46,7 @@ class PruningFrameworkImplementation(FrameworkImplementation):
     def prune_intermediate_node(self,
                                 node: BaseNode,
                                 input_mask: np.ndarray,
-                                output_mask: np.ndarray,
-                                fw_info: FrameworkInfo):
+                                output_mask: np.ndarray):
         """
         Abstract method to prune an intermediate node in the model.
 
@@ -57,7 +54,6 @@ class PruningFrameworkImplementation(FrameworkImplementation):
             node: The node to be pruned.
             input_mask: Mask to be applied to the input channels.
             output_mask: Mask to be applied to the output channels.
-            fw_info: Framework-specific information.
 
         Raises:
             NotImplemented: If the method is not implemented in the subclass.
@@ -68,15 +64,13 @@ class PruningFrameworkImplementation(FrameworkImplementation):
     @abstractmethod
     def prune_exit_node(self,
                         node: BaseNode,
-                        input_mask: np.ndarray,
-                        fw_info: FrameworkInfo):
+                        input_mask: np.ndarray):
         """
         Abstract method to prune an exit node in the model.
 
         Args:
             node: The node to be pruned.
             input_mask: Mask to be applied to the input channels.
-            fw_info: Framework-specific information.
 
         Raises:
             NotImplemented: If the method is not implemented in the subclass.
@@ -105,8 +99,7 @@ class PruningFrameworkImplementation(FrameworkImplementation):
     @abstractmethod
     def is_node_exit_node(self,
                           node: BaseNode,
-                          corresponding_entry_node: BaseNode,
-                          fw_info: FrameworkInfo) -> bool:
+                          corresponding_entry_node: BaseNode) -> bool:
 
         raise NotImplemented(f'{self.__class__.__name__} have to implement the '
                              f'framework\'s is_node_exit_node method.')  # pragma: no cover
@@ -129,7 +122,7 @@ class PruningFrameworkImplementation(FrameworkImplementation):
         raise NotImplemented(f'{self.__class__.__name__} have to implement the '
                              f'framework\'s is_node_intermediate_pruning_section method.')  # pragma: no cover
 
-    def attrs_oi_channels_info_for_pruning(self, node: BaseNode, fw_info: FrameworkInfo) -> Dict[str, Tuple[int, int]]:
+    def attrs_oi_channels_info_for_pruning(self, node: BaseNode) -> Dict[str, Tuple[int, int]]:
         """
         Retrieves the attributes of a given node along with the output/input (OI) channel axis
         for each attribute used to prune these attributes.
@@ -146,7 +139,6 @@ class PruningFrameworkImplementation(FrameworkImplementation):
 
         Args:
             node (BaseNode): The node from the computational graph.
-            fw_info (FrameworkInfo): Contains framework-specific information and utilities.
 
         Returns:
             Dict[str, Tuple[int, int]]: A dictionary where each key is an attribute name (like 'kernel' or 'bias')

model_compression_toolkit/core/common/pruning/pruning_section.py

@@ -76,34 +76,28 @@ class PruningSection:
 
     def apply_inner_section_mask(self,
                                  pruning_section_mask: PruningSectionMask,
-                                 fw_impl: Any,
-                                 fw_info: FrameworkInfo):
+                                 fw_impl: Any):
         """
         Apply the provided pruning section mask to all nodes within the pruning section.
 
         Args:
             pruning_section_mask (PruningSectionMask): The mask to be applied to the pruning section.
             fw_impl (PruningFrameworkImplementation): Framework-specific implementation for applying the mask.
-            fw_info (FrameworkInfo): Framework-specific information needed to apply the mask.
         """
         fw_impl.prune_entry_node(node=self.entry_node,
-                                 output_mask=pruning_section_mask.entry_node_oc_mask,
-                                 fw_info=fw_info)
+                                 output_mask=pruning_section_mask.entry_node_oc_mask)
 
         for inter_node in self.intermediate_nodes:
             fw_impl.prune_intermediate_node(node=inter_node,
                                             input_mask=pruning_section_mask.entry_node_oc_mask,
-                                            output_mask=pruning_section_mask.entry_node_oc_mask,
-                                            fw_info=fw_info)
+                                            output_mask=pruning_section_mask.entry_node_oc_mask)
 
         fw_impl.prune_exit_node(self.exit_node,
-                                input_mask=pruning_section_mask.exit_node_ic_mask,
-                                fw_info=fw_info)
+                                input_mask=pruning_section_mask.exit_node_ic_mask)
 
     @staticmethod
     def has_matching_channel_count(exit_node: BaseNode,
-                                   corresponding_entry_node: BaseNode,
-                                   fw_info: FrameworkInfo) -> bool:
+                                   corresponding_entry_node: BaseNode) -> bool:
         """
         Checks if the number of input channels of the exit node matches the number of output channels
         of its corresponding entry node.
@@ -115,13 +109,10 @@ class PruningSection:
         Returns:
             bool: True if the channel counts match, False otherwise.
         """
-        _, exit_input_channel_axis = fw_info.kernel_channels_mapping.get(exit_node.type)
-        entry_output_channel_axis, _ = fw_info.kernel_channels_mapping.get(corresponding_entry_node.type)
+        exit_input_channel_axis = exit_node.channel_axis.input
+        entry_output_channel_axis = corresponding_entry_node.channel_axis.output
 
-        exit_node_attr = fw_info.get_kernel_op_attributes(exit_node.type)[0]
-        entry_node_attr = fw_info.get_kernel_op_attributes(corresponding_entry_node.type)[0]
-
-        exit_input_channels = exit_node.get_weights_by_keys(exit_node_attr).shape[exit_input_channel_axis]
-        entry_output_channels = corresponding_entry_node.get_weights_by_keys(entry_node_attr).shape[entry_output_channel_axis]
+        exit_input_channels = exit_node.get_weights_by_keys(exit_node.kernel_attr).shape[exit_input_channel_axis]
+        entry_output_channels = corresponding_entry_node.get_weights_by_keys(corresponding_entry_node.kernel_attr).shape[entry_output_channel_axis]
 
         return exit_input_channels == entry_output_channels

model_compression_toolkit/core/common/quantization/candidate_node_quantization_config.py

@@ -15,6 +15,7 @@
 from typing import Callable, List, Tuple
 
 from model_compression_toolkit.core import QuantizationConfig
+from model_compression_toolkit.core.common.framework_info import ChannelAxisMapping
 from model_compression_toolkit.core.common.quantization.node_quantization_config import BaseNodeQuantizationConfig, \
     NodeWeightsQuantizationConfig, NodeActivationQuantizationConfig
 from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import AttributeQuantizationConfig, \
@@ -40,7 +41,7 @@ class CandidateNodeQuantizationConfig(BaseNodeQuantizationConfig):
                  activation_quantization_fn: Callable = None,
                  activation_quantization_params_fn: Callable = None,
                  weights_quantization_cfg: NodeWeightsQuantizationConfig = None,
-                 weights_channels_axis: Tuple[int, int] = None,
+                 weights_channels_axis: ChannelAxisMapping = None,
                  node_attrs_list: List[str] = None):
         """
 

model_compression_toolkit/core/common/quantization/filter_nodes_candidates.py

@@ -34,7 +34,7 @@ def filter_nodes_candidates(graph: Graph):
     """
     nodes = list(graph.nodes)
     for n in nodes:
-        n.candidates_quantization_cfg = filter_node_candidates(node=n, fw_info=graph.fw_info)
+        n.candidates_quantization_cfg = filter_node_candidates(node=n)
 
     return graph
 
@@ -71,7 +71,7 @@ def _filter_bit_method_dups(candidates: List[CandidateNodeQuantizationConfig],
     return final_candidates
 
 
-def filter_node_candidates(node: BaseNode, fw_info) -> List[CandidateNodeQuantizationConfig]:
+def filter_node_candidates(node: BaseNode) -> List[CandidateNodeQuantizationConfig]:
     """
     Updates a node's candidates configuration list.
     If the node's weights quantization is disabled (or it only has activations to quantize), then the updated list
@@ -81,15 +81,13 @@ def filter_node_candidates(node: BaseNode, fw_info) -> List[CandidateNodeQuantiz
 
     Args:
         node: Node to set its quantization configurations.
-        fw_info: FrameworkInfo object with information about the specific framework's model.
 
     """
 
     filtered_candidates = copy.deepcopy(node.candidates_quantization_cfg)
     final_candidates = copy.deepcopy(node.candidates_quantization_cfg)
-    kernel_attr = fw_info.get_kernel_op_attributes(node.type)[0]
 
-    if (kernel_attr is None or not node.is_weights_quantization_enabled(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 not node.is_activation_quantization_enabled():
         # 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.
@@ -97,8 +95,8 @@ def filter_node_candidates(node: BaseNode, fw_info) -> List[CandidateNodeQuantiz
         single_dummy_candidate.activation_quantization_cfg.activation_n_bits = FLOAT_BITWIDTH
         single_dummy_candidate.activation_quantization_cfg.activation_quantization_method = QuantizationMethod.POWER_OF_TWO
 
-        if kernel_attr is not None:
-            kernel_config = single_dummy_candidate.weights_quantization_cfg.get_attr_config(kernel_attr)
+        if node.kernel_attr is not None:
+            kernel_config = single_dummy_candidate.weights_quantization_cfg.get_attr_config(node.kernel_attr)
             kernel_config.weights_n_bits = FLOAT_BITWIDTH
             kernel_config.weights_quantization_method = QuantizationMethod.POWER_OF_TWO
 
@@ -116,9 +114,9 @@ def filter_node_candidates(node: BaseNode, fw_info) -> List[CandidateNodeQuantiz
             c.activation_quantization_cfg.activation_n_bits = FLOAT_BITWIDTH
             c.activation_quantization_cfg.activation_quantization_method = QuantizationMethod.POWER_OF_TWO
 
-        final_candidates = _filter_bit_method_dups(filtered_candidates, kernel_attr)
+        final_candidates = _filter_bit_method_dups(filtered_candidates, node.kernel_attr)
 
-    elif kernel_attr is None or not node.is_weights_quantization_enabled(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.
         # Remove candidates that have duplicated activation candidates for node with disabled weights quantization.
@@ -129,11 +127,11 @@ def filter_node_candidates(node: BaseNode, fw_info) -> List[CandidateNodeQuantiz
                                and not seen_candidates.add(candidate.activation_quantization_cfg)]
 
         for c in filtered_candidates:
-            if kernel_attr is not None:
-                kernel_config = c.weights_quantization_cfg.get_attr_config(kernel_attr)
+            if node.kernel_attr is not None:
+                kernel_config = c.weights_quantization_cfg.get_attr_config(node.kernel_attr)
                 kernel_config.weights_n_bits = FLOAT_BITWIDTH
                 kernel_config.weights_quantization_method = QuantizationMethod.POWER_OF_TWO
 
-        final_candidates = _filter_bit_method_dups(filtered_candidates, kernel_attr)
+        final_candidates = _filter_bit_method_dups(filtered_candidates, node.kernel_attr)
 
     return final_candidates

model_compression_toolkit/core/common/quantization/node_quantization_config.py

@@ -18,6 +18,7 @@ from typing import Callable, Any, List, Tuple, Union, Dict, TYPE_CHECKING
 from enum import Enum, auto
 import numpy as np
 
+from model_compression_toolkit.core.common.framework_info import ChannelAxisMapping
 from model_compression_toolkit.core.common.quantization.quantization_fn_selection import get_weights_quantization_fn
 from model_compression_toolkit.logger import Logger
 from model_compression_toolkit.core.common.quantization.quantization_params_fn_selection import \
@@ -262,7 +263,7 @@ class WeightsAttrQuantizationConfig:
     def __init__(self,
                  qc: QuantizationConfig,
                  weights_attr_cfg: AttributeQuantizationConfig,
-                 weights_channels_axis: Tuple[int, int] = None):
+                 weights_channels_axis: ChannelAxisMapping = None):
         """
 
         Args:
@@ -352,7 +353,7 @@ class WeightsAttrQuantizationConfig:
                                                     p=self.l_p_value,
                                                     n_bits=self.weights_n_bits,
                                                     per_channel=self.weights_per_channel_threshold and self.weights_channels_axis is not None,
-                                                    channel_axis=self.weights_channels_axis[0],  # output channel axis
+                                                    channel_axis=self.weights_channels_axis.output,  # output channel axis
                                                     min_threshold=min_threshold)[0]  # Take only first output, the q-params, as axis is already chosen.
             )
         else:
@@ -400,7 +401,7 @@ class NodeWeightsQuantizationConfig(BaseNodeQuantizationConfig):
     """
     def __init__(self, qc: QuantizationConfig,
                  op_cfg: OpQuantizationConfig,
-                 weights_channels_axis: Tuple[int, int],
+                 weights_channels_axis: ChannelAxisMapping,
                  node_attrs_list: List[str]):
         """
 

model_compression_toolkit/core/common/quantization/quantization_params_generation/qparams_computation.py

@@ -20,6 +20,7 @@ from typing import List, Callable, Generator
 from model_compression_toolkit.constants import NUM_QPARAM_HESSIAN_SAMPLES
 from model_compression_toolkit.core import QuantizationErrorMethod
 from model_compression_toolkit.core.common import Graph, BaseNode
+from model_compression_toolkit.core.common.framework_info import ChannelAxisMapping
 from model_compression_toolkit.core.common.framework_implementation import FrameworkImplementation
 from model_compression_toolkit.core.common.hessian import HessianInfoService, HessianScoresRequest, HessianMode, \
     HessianScoresGranularity
@@ -44,11 +45,8 @@ def _collect_nodes_for_hmse(nodes_list: List[BaseNode], graph: Graph) -> List[Ba
     """
     hmse_nodes = []
     for n in nodes_list:
-        kernel_attr_name = graph.fw_info.get_kernel_op_attributes(n.type)
-        kernel_attr_name = None if kernel_attr_name is None or len(kernel_attr_name) == 0 else kernel_attr_name[0]
-
-        if kernel_attr_name is not None and n.is_weights_quantization_enabled(kernel_attr_name) and \
-            all([c.weights_quantization_cfg.get_attr_config(kernel_attr_name).weights_error_method ==
+        if n.kernel_attr is not None and n.is_weights_quantization_enabled(n.kernel_attr) and \
+            all([c.weights_quantization_cfg.get_attr_config(n.kernel_attr).weights_error_method ==
                  QuantizationErrorMethod.HMSE for c in n.candidates_quantization_cfg]):
             hmse_nodes.append(n)
 
@@ -114,11 +112,7 @@ def calculate_quantization_params(graph: Graph,
                     if attr_cfg.weights_error_method == QuantizationErrorMethod.HMSE:
                         # Although we collected nodes for HMSE before running the loop, we keep this verification to
                         # notify the user in case of HMSE configured for node that is not compatible for this method
-                        kernel_attr_name = graph.fw_info.get_kernel_op_attributes(n.type)
-                        if len(kernel_attr_name) > 0:
-                            kernel_attr_name = kernel_attr_name[0]
-
-                        if kernel_attr_name is None or kernel_attr_name not in attr:
+                        if n.kernel_attr is None or n.kernel_attr not in attr:
                             Logger.warning(f"The HMSE error method for parameters selection is only supported for "
                                            f"kernel weights attributes. Running parameters selection for attribute "
                                            f"'{attr}' in node '{n.name}' with the default MSE error method instead.")
@@ -132,7 +126,7 @@ def calculate_quantization_params(graph: Graph,
                                                                                node=n,
                                                                                hessian_info_service=hessian_info_service,
                                                                                num_hessian_samples=num_hessian_samples)
-                    attr_cfg.weights_channels_axis = (output_channels_axis, attr_cfg.weights_channels_axis[1])
+                    attr_cfg.weights_channels_axis = ChannelAxisMapping(output_channels_axis, attr_cfg.weights_channels_axis.input)
                     attr_cfg.set_weights_quantization_param(weights_params)
 
             if n.is_activation_quantization_enabled():

model_compression_toolkit/core/common/quantization/set_node_quantization_config.py

@@ -20,7 +20,7 @@ from model_compression_toolkit.constants import WEIGHTS, ACTIVATION
 from model_compression_toolkit.core.common import BaseNode
 from model_compression_toolkit.core.common.quantization.bit_width_config import BitWidthConfig
 from model_compression_toolkit.logger import Logger
-from model_compression_toolkit.core.common.framework_info import FrameworkInfo
+from model_compression_toolkit.core.common.framework_info import get_fw_info, ChannelAxisMapping
 from model_compression_toolkit.core.common.graph.base_graph import Graph
 from model_compression_toolkit.core.common.quantization.candidate_node_quantization_config import \
     CandidateNodeQuantizationConfig
@@ -73,7 +73,6 @@ def set_quantization_configuration_to_graph(graph: Graph,
         set_quantization_configs_to_node(node=n,
                                          graph=graph,
                                          quant_config=quant_config,
-                                         fw_info=graph.fw_info,
                                          fqc=graph.fqc,
                                          mixed_precision_enable=mixed_precision_enable,
                                          manual_bit_width_override=manual_bit_width_override)
@@ -154,7 +153,6 @@ def filter_node_qco_by_graph(node: BaseNode,
 def set_quantization_configs_to_node(node: BaseNode,
                                      graph: Graph,
                                      quant_config: QuantizationConfig,
-                                     fw_info: FrameworkInfo,
                                      fqc: FrameworkQuantizationCapabilities,
                                      mixed_precision_enable: bool = False,
                                      manual_bit_width_override: Optional[Dict] = None):
@@ -165,7 +163,6 @@ def set_quantization_configs_to_node(node: BaseNode,
         node (BaseNode): Node to set its quantization configurations.
         graph (Graph): Model's internal representation graph.
         quant_config (QuantizationConfig): Quantization configuration to generate the node's configurations from.
-        fw_info (FrameworkInfo): Information needed for quantization about the specific framework.
         fqc (FrameworkQuantizationCapabilities): FrameworkQuantizationCapabilities to get default OpQuantizationConfig.
         mixed_precision_enable (bool): Whether mixed precision is enabled. Defaults to False.
         manual_bit_width_override (Optional[int]): Specifies a custom bit-width to override the node's activation bit-width. Defaults to None.
@@ -186,10 +183,8 @@ def set_quantization_configs_to_node(node: BaseNode,
         mixed_precision_enable=mixed_precision_enable)
 
     # Create QC candidates for weights and activation combined
-    weight_channel_axis = fw_info.kernel_channels_mapping.get(node.type)
     node.candidates_quantization_cfg = _create_node_candidates_qc(quant_config,
-                                                                  fw_info,
-                                                                  weight_channel_axis,
+                                                                  node.channel_axis,
                                                                   node_qc_options_list,
                                                                   base_config,
                                                                   node,
@@ -198,7 +193,7 @@ def set_quantization_configs_to_node(node: BaseNode,
     # sorting the candidates by kernel attribute weights number of bits first and then by activation number of bits
     # (in reversed order). since only kernel attribute is quantized in weights mixed precision,
     # if the node doesn't have a kernel attribute, we only sort by activation_n_bits.
-    node.sort_node_candidates(fw_info)
+    node.sort_node_candidates()
 
     for candidate_qc in node.candidates_quantization_cfg:
         if candidate_qc.activation_quantization_cfg.quant_mode == ActivationQuantizationMode.QUANT and \
@@ -217,14 +212,12 @@ def set_quantization_configs_to_node(node: BaseNode,
 
 
 def create_node_activation_qc(qc: QuantizationConfig,
-                              fw_info: FrameworkInfo,
                               op_cfg: OpQuantizationConfig) -> NodeActivationQuantizationConfig:
     """
     Create an activation quantization configuration from a QuantizationConfig object.
 
     Args:
         qc: QuantizationConfig to create the node's config from.
-        fw_info: Information about the specific framework the node was created from (e.g., whether or not its
         weights/activations should be quantized)
         op_cfg: OpQuantizationConfig with quantizers types to set in node quantization configuration.
 
@@ -232,7 +225,7 @@ def create_node_activation_qc(qc: QuantizationConfig,
         Activation quantization configuration of a node.
     """
 
-    activation_quantization_fn = fw_info.activation_quantizer_mapping.get(op_cfg.activation_quantization_method)
+    activation_quantization_fn = get_fw_info().activation_quantizer_mapping.get(op_cfg.activation_quantization_method)
     if activation_quantization_fn is None:
         Logger.critical('Unknown activation quantization method specified.')  # pragma: no cover
 
@@ -245,8 +238,7 @@ def create_node_activation_qc(qc: QuantizationConfig,
 
 
 def _create_node_single_candidate_qc(qc: QuantizationConfig,
-                                     fw_info: FrameworkInfo,
-                                     weight_channel_axis: Tuple[int, int],
+                                     weight_channel_axis: ChannelAxisMapping,
                                      op_cfg: OpQuantizationConfig,
                                      node_attrs_list: List[str]) -> CandidateNodeQuantizationConfig:
     """
@@ -256,8 +248,6 @@ def _create_node_single_candidate_qc(qc: QuantizationConfig,
 
     Args:
         qc: QuantizationConfig to create the node's config from.
-        fw_info: Information about the specific framework the node was created from (e.g., whether its
-            weights/activations should be quantized)
         weight_channel_axis: (Output, Input) channel index of the node's kernel.
         op_cfg: OpQuantizationConfig of the node with quantizers types to use when creating node quantization configuration.
         node_attrs_list: A list of the node's weights attributes names.
@@ -269,7 +259,7 @@ def _create_node_single_candidate_qc(qc: QuantizationConfig,
     # parameters for weights attributes quantization are set within  CandidateNodeQuantizationConfig initialization
 
     # get parameters for activation quantization
-    activation_quantization_fn = fw_info.activation_quantizer_mapping.get(op_cfg.activation_quantization_method)
+    activation_quantization_fn = get_fw_info().activation_quantizer_mapping.get(op_cfg.activation_quantization_method)
     if activation_quantization_fn is None:
         Logger.critical('Unknown activation quantization method specified.')  # pragma: no cover
 
@@ -293,8 +283,7 @@ def _create_node_single_candidate_qc(qc: QuantizationConfig,
 
 
 def _create_node_candidates_qc(qc: QuantizationConfig,
-                               fw_info: FrameworkInfo,
-                               weight_channel_axis: Tuple[int, int],
+                               weight_channel_axis: ChannelAxisMapping,
                                node_qc_options_list: List[OpQuantizationConfig],
                                base_config: OpQuantizationConfig,
                                node: BaseNode,
@@ -304,8 +293,7 @@ def _create_node_candidates_qc(qc: QuantizationConfig,
 
     Args:
         qc (QuantizationConfig): Quantization configuration the quantization process should follow.
-        fw_info (FrameworkInfo): Framework information (e.g., which layers should have their kernels quantized).
-        weight_channel_axis (Tuple[int, int]): (Output, Input) channel index of the node's kernel.
+        weight_channel_axis (ChannelAxisMapping): (Output, Input) channel index of the node's kernel.
         node_qc_options_list (List[OpQuantizationConfig]): List of quantization configs of node.
         base_config (OpQuantizationConfig): Base quantization config for node.
         node (BaseNode): A node to set quantization configuration candidates to.
@@ -322,14 +310,12 @@ def _create_node_candidates_qc(qc: QuantizationConfig,
         for op_cfg in node_qc_options_list:
             candidate_qc = copy.deepcopy(qc)
             candidates.append(_create_node_single_candidate_qc(candidate_qc,
-                                                               fw_info,
                                                                weight_channel_axis,
                                                                op_cfg,
                                                                node_attrs_list))
 
     else:
         candidates.append(_create_node_single_candidate_qc(qc,
-                                                           fw_info,
                                                            weight_channel_axis,
                                                            base_config,
                                                            node_attrs_list))

model_compression_toolkit/core/common/statistics_correction/apply_activation_bias_correction_to_graph.py

@@ -38,8 +38,7 @@ def apply_activation_bias_correction_to_graph(graph: Graph,
 
     for n in graph.nodes:
         # Activation bias correction is only relevant for nodes with kernel op
-        kernel_attr = graph.fw_info.get_kernel_op_attributes(n.type)[0]
-        if core_config.quantization_config.activation_bias_correction and kernel_attr is not None and \
+        if core_config.quantization_config.activation_bias_correction and n.kernel_attr is not None and \
                 n.final_activation_quantization_cfg.activation_bias_correction_term is not None:
             # If activation bias correction is enabled in n.quantization_cfg, an activation bias correction term was
             # calculated during model preparation, and is used now in the node's bias term.

model_compression_toolkit/core/common/statistics_correction/apply_bias_correction_to_graph.py

@@ -41,9 +41,8 @@ def apply_bias_correction_to_graph(graph_to_apply_bias_correction: Graph,
     graph = copy.deepcopy(graph_to_apply_bias_correction)
     for n in graph.nodes:
         # bias correction is only relevant for nodes with kernel op
-        kernel_attr = graph.fw_info.get_kernel_op_attributes(n.type)[0]
-        if core_config.quantization_config.weights_bias_correction and kernel_attr is not None and \
-            n.is_weights_quantization_enabled(kernel_attr) and \
+        if core_config.quantization_config.weights_bias_correction and n.kernel_attr is not None and \
+            n.is_weights_quantization_enabled(n.kernel_attr) and \
                 not n.final_weights_quantization_cfg.weights_second_moment_correction:
             # If a kernel was quantized and weights bias correction is enabled in n.quantization_cfg,
             # a bias correction term was calculated during model preparation, and is used now in the node's bias term.