mct-nightly 2.4.0.20250924.535__py3-none-any.whl → 2.4.2.20250926.532__py3-none-any.whl

model_compression_toolkit/core/common/graph/base_node.py

@@ -14,37 +14,31 @@
 # ==============================================================================
 
 import copy
-from typing import Dict, Any, Tuple, List, Type, Union, NamedTuple
+from typing import Dict, Any, Tuple, List, Type, Union
 
 import numpy as np
 
-from model_compression_toolkit.core.common.framework_info import get_fw_info, ChannelAxisMapping
 from model_compression_toolkit.constants import WEIGHTS_NBITS_ATTRIBUTE, CORRECTED_BIAS_ATTRIBUTE, \
     ACTIVATION_N_BITS_ATTRIBUTE, FP32_BYTES_PER_PARAMETER
-from model_compression_toolkit.core.common.quantization.candidate_node_quantization_config import NodeQuantizationConfig
 from model_compression_toolkit.core.common.quantization.node_quantization_config import WeightsAttrQuantizationConfig, \
     ActivationQuantizationMode
 from model_compression_toolkit.logger import Logger
+from model_compression_toolkit.target_platform_capabilities.schema.mct_current_schema import QuantizationConfigOptions, \
+    OpQuantizationConfig
+from model_compression_toolkit.target_platform_capabilities.schema.schema_functions import max_input_activation_n_bits
 from model_compression_toolkit.target_platform_capabilities.targetplatform2framework import LayerFilterParams
+from model_compression_toolkit.target_platform_capabilities.targetplatform2framework.framework_quantization_capabilities import \
+    FrameworkQuantizationCapabilities
 
 
 WeightAttrT = Union[str, int]
 
 
-class NodeFrameworkInfo(NamedTuple):
-    """
-    Node's specific framework information.
-    """
-    channel_axis: ChannelAxisMapping
-    out_channel_axis: int
-    minmax: Tuple[float, float]
-    kernel_attr: str
-
-
 class BaseNode:
     """
     Class to represent a node in a graph that represents the model.
     """
+
     def __init__(self,
                  name: str,
                  framework_attr: Dict[str, Any],
@@ -90,84 +84,28 @@ class BaseNode:
         self.inputs_as_list = inputs_as_list
         self.final_weights_quantization_cfg = None
         self.final_activation_quantization_cfg = None
-        self.quantization_cfg: NodeQuantizationConfig = None
+        self.candidates_quantization_cfg = None
         self.prior_info = None
         self.has_activation = has_activation
         self.is_custom = is_custom
-        self.node_fw_info = self._get_fw_node_attrs(layer_class, framework_attr)
-
-    def _get_fw_node_attrs(self, node_type, framework_attr):
-        fw_info = get_fw_info()
-        return None if fw_info is None else NodeFrameworkInfo(
-            fw_info.get_kernel_channels(node_type),
-            fw_info.get_out_channel_axis(node_type),
-            fw_info.get_layer_min_max(node_type, framework_attr),
-            fw_info.get_kernel_op_attribute(node_type),
-        )
-
-    def _assert_fw_info_exists(self):
-        """
-        Verify NodeFrameworkInfo was initialized.
-        """
-        assert self.node_fw_info is not None, f"NodeFrameworkInfo not initialized for node {self.name}"  # pragma: no cover
-
-    @property
-    def channel_axis(self) -> ChannelAxisMapping:
-        """
-        Extract channels axis from node's NodeFrameworkInfo.
-
-        Returns:
-            Channels axis named tuple.
-        """
-        self._assert_fw_info_exists()
-        return self.node_fw_info.channel_axis
-
-    @property
-    def out_channel_axis(self) -> int:
-        """
-        Extract output channel axis from node's NodeFrameworkInfo.
-
-        Returns:
-            Output channel axis.
-        """
-        self._assert_fw_info_exists()
-        return self.node_fw_info.out_channel_axis
 
     @property
-    def minmax(self) -> Tuple[float, float]:
+    def type(self):
         """
-        Extract expected min-max activation values from node's NodeFrameworkInfo.
-
+        A function to get the node's layer_class op for convenient comparison
         Returns:
-            A tuple of min-max values.
-        """
-        self._assert_fw_info_exists()
-        return self.node_fw_info.minmax
-
-    @property
-    def kernel_attr(self) -> str:
+            the node's layer_class
         """
-        Extract kernel name from node's NodeFrameworkInfo.
+        return self.layer_class
 
-        Returns:
-            Kernel name.
+    def get_has_activation(self):
         """
-        self._assert_fw_info_exists()
-        return self.node_fw_info.kernel_attr
+        Returns has_activation attribute.
 
-    @property
-    def candidates_quantization_cfg(self):
-        assert self.quantization_cfg
-        return self.quantization_cfg.candidates_quantization_cfg
+        Returns: Whether the node has activation to quantize.
 
-    @property
-    def type(self):
-        """
-        A function to get the node's layer_class op for convenient comparison
-        Returns:
-            the node's layer_class
         """
-        return self.layer_class
+        return self.has_activation
 
     @property
     def has_positional_weights(self):
@@ -195,31 +133,19 @@ class BaseNode:
         Returns: Whether node activation quantization is enabled or not.
         """
         return self._is_single_quant_mode(ActivationQuantizationMode.QUANT)
-    
-    def is_fln_no_quantization(self) -> bool:
+
+    def is_fln_quantization(self) -> bool:
         """
-        Returns: Whether node is FLN no quantization.
+        Returns: Whether the node's activation quantization is FLN
         """
-        return self._is_single_quant_mode(ActivationQuantizationMode.FLN_NO_QUANT)
-    
+        return self._is_single_quant_mode(ActivationQuantizationMode.FLN_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.
@@ -372,11 +298,14 @@ class BaseNode:
 
         return input_tensors
 
-    def get_num_parameters(self) -> Tuple[int,int]:
+    def get_num_parameters(self, fw_info) -> Tuple[int,int]:
         """
         Compute the number of parameters the node holds.
         It returns a tuple: Number of quantized parameters, number of float parameters.
 
+        Args:
+            fw_info: Framework info to decide which attributes should be quantized.
+
         Returns:
             A tuple of (Number of quantized parameters, number of float parameters).
 
@@ -385,10 +314,11 @@ class BaseNode:
 
         q_node_num_params = 0
 
-        if self.kernel_attr is not None:
-            w = self.get_weights_by_keys(self.kernel_attr)
-            if w is not None:
-                q_node_num_params += w.flatten().shape[0]
+        for attr in fw_info.get_kernel_op_attributes(self.type):
+            if attr is not None:
+                w = self.get_weights_by_keys(attr)
+                if w is not None:
+                    q_node_num_params += w.flatten().shape[0]
 
         f_node_num_params = total_node_params - q_node_num_params
 
@@ -396,19 +326,22 @@ class BaseNode:
         assert int(f_node_num_params) == f_node_num_params
         return int(q_node_num_params), int(f_node_num_params)
 
-    def get_memory_bytes(self) -> float:
+    def get_memory_bytes(self, fw_info) -> float:
         """
         Compute the number of bytes the node's memory requires.
 
+        Args:
+            fw_info: Framework info to decide which attributes should be quantized.
+
         Returns: Number of bytes the node's memory requires.
 
         """
         # TODO: this method is used for tensorboard only. If we want to enable logging of other attributes memory
         #  then it needs to be modified. But, it might be better to remove this method from the BaseNode completely.
-        kernel_attr = self.kernel_attr
+        kernel_attr = fw_info.get_kernel_op_attributes(self.type)[0]
         if kernel_attr is None:
             return 0
-        q_params, f_params = self.get_num_parameters()
+        q_params, f_params = self.get_num_parameters(fw_info)
         if self.final_weights_quantization_cfg is None:  # float coefficients
             memory = (f_params+q_params) * FP32_BYTES_PER_PARAMETER
         else:
@@ -418,12 +351,15 @@ class BaseNode:
 
         return memory
 
-    def get_unified_weights_candidates_dict(self) -> Dict[str, Any]:
+    def get_unified_weights_candidates_dict(self, fw_info) -> Dict[str, Any]:
         """
         In Mixed-Precision, a node's kernel can have multiple candidates for weights quantization configuration.
         In order to display a single view of a node (for example, for logging in TensorBoard) we need a way
         to create a single dictionary from all candidates.
-        This method is aimed to build such a unified dictionary for a node.
+        This method is aimed to build such an unified dictionary for a node.
+
+        Args:
+            fw_info: FrameworkInfo object about the specific framework (e.g., attributes of different layers' weights to quantize).
 
         Returns: A dictionary containing information from node's weight quantization configuration candidates.
 
@@ -433,7 +369,7 @@ class BaseNode:
         # We assume that only the kernel attribute have more than one candidate, since we only allow to
         # quantize the kernel using mixed precision
         # TODO: need to modify if we want to present a unified config for other attributes
-        kernel_attr = self.kernel_attr
+        kernel_attr = fw_info.get_kernel_op_attributes(self.type)[0]
         if kernel_attr is None:
             # This node doesn't have a kernel attribute
             return {}
@@ -501,13 +437,20 @@ class BaseNode:
         candidates = self.get_all_weights_attr_candidates(attr)
         return all(candidate == candidates[0] for candidate in candidates[1:])
 
-    def has_kernel_weight_to_quantize(self):
+    def has_kernel_weight_to_quantize(self, fw_info):
         """
-        Checks whether the node has kernel attribute that need to be quantized according to the node's framework info.
+        Checks whether the node has kernel attribute that need to be quantized according to the framework info.
+
+        Args:
+            fw_info: FrameworkInfo object about the specific framework (e.g., attributes of different layers' weights to quantize).
 
-        Returns: Whether the node's kernel need to be quantized.
+        Returns: Whether the node has weights that need to be quantized.
         """
-        return self.kernel_attr and self.get_weights_by_keys(self.kernel_attr) is not None
+        attrs = fw_info.get_kernel_op_attributes(self.type)
+        for attr in attrs:
+            if attr and self.get_weights_by_keys(attr) is not None:
+                return True
+        return False
 
     def has_any_weight_attr_to_quantize(self) -> bool:
         """
@@ -625,9 +568,8 @@ class BaseNode:
         Returns: True if the node has at list one quantization configuration candidate with activation quantization enabled.
         """
 
-        return (len(self.candidates_quantization_cfg) > 0 and
-                any([c.activation_quantization_cfg.enable_activation_quantization
-                     for c in self.candidates_quantization_cfg]))
+        return len(self.candidates_quantization_cfg) > 0 and \
+            any([c.activation_quantization_cfg.enable_activation_quantization for c in self.candidates_quantization_cfg])
 
     def get_all_weights_attr_candidates(self, attr: str) -> List[WeightsAttrQuantizationConfig]:
         """
@@ -643,6 +585,79 @@ class BaseNode:
         # the inner method would log an exception.
         return [c.weights_quantization_cfg.get_attr_config(attr) for c in self.candidates_quantization_cfg]
 
+    def get_qco(self, fqc: FrameworkQuantizationCapabilities) -> QuantizationConfigOptions:
+        """
+        Get the QuantizationConfigOptions of the node according
+        to the mappings from layers/LayerFilterParams to the OperatorsSet in the TargetPlatformCapabilities.
+
+        Args:
+            fqc: FQC to extract the QuantizationConfigOptions for the node.
+
+        Returns:
+            QuantizationConfigOptions of the node.
+        """
+
+        if fqc is None:
+            Logger.critical(f'Can not retrieve QC options for None FQC')  # pragma: no cover
+
+        for fl, qco in fqc.filterlayer2qco.items():
+            if self.is_match_filter_params(fl):
+                return qco
+        # Extract qco with is_match_type to overcome mismatch of function types in TF 2.15
+        matching_qcos = [_qco for _type, _qco in fqc.layer2qco.items() if self.is_match_type(_type)]
+        if matching_qcos:
+            if all([_qco == matching_qcos[0] for _qco in matching_qcos]):
+                return matching_qcos[0]
+            else:
+                Logger.critical(f"Found duplicate qco types for node '{self.name}' of type '{self.type}'!")  # pragma: no cover
+        return fqc.tpc.default_qco
+
+    def filter_node_qco_by_graph(self, fqc: FrameworkQuantizationCapabilities,
+                                 next_nodes: List, node_qc_options: QuantizationConfigOptions
+                                 ) -> Tuple[OpQuantizationConfig, List[OpQuantizationConfig]]:
+        """
+        Filter quantization config options that don't match the graph.
+        A node may have several quantization config options with 'activation_n_bits' values, and
+        the next nodes in the graph may support different bit-width as input activation. This function
+        filters out quantization config that don't comply to these attributes.
+
+        Args:
+            fqc: FQC to extract the QuantizationConfigOptions for the next nodes.
+            next_nodes: Output nodes of current node.
+            node_qc_options: Node's QuantizationConfigOptions.
+
+        Returns:
+
+        """
+        # Filter quantization config options that don't match the graph.
+        _base_config = node_qc_options.base_config
+        _node_qc_options = node_qc_options.quantization_configurations
+        if len(next_nodes):
+            next_nodes_qc_options = [_node.get_qco(fqc) for _node in next_nodes]
+            next_nodes_supported_input_bitwidth = min([max_input_activation_n_bits(op_cfg)
+                                                       for qc_opts in next_nodes_qc_options
+                                                       for op_cfg in qc_opts.quantization_configurations])
+
+            # Filter node's QC options that match next nodes input bit-width.
+            _node_qc_options = [_option for _option in _node_qc_options
+                                if _option.activation_n_bits <= next_nodes_supported_input_bitwidth]
+            if len(_node_qc_options) == 0:
+                Logger.critical(f"Graph doesn't match FQC bit configurations: {self} -> {next_nodes}.")  # pragma: no cover
+
+            # Verify base config match
+            if any([node_qc_options.base_config.activation_n_bits > max_input_activation_n_bits(qc_opt.base_config)
+                    for qc_opt in next_nodes_qc_options]):
+                # base_config activation bits doesn't match next node supported input bit-width -> replace with
+                # a qco from quantization_configurations with maximum activation bit-width.
+                if len(_node_qc_options) > 0:
+                    output_act_bitwidth = {qco.activation_n_bits: i for i, qco in enumerate(_node_qc_options)}
+                    _base_config = _node_qc_options[output_act_bitwidth[max(output_act_bitwidth)]]
+                    Logger.warning(f"Node {self} base quantization config changed to match Graph and FQC configuration.\nCause: {self} -> {next_nodes}.")
+                else:
+                    Logger.critical(f"Graph doesn't match FQC bit configurations: {self} -> {next_nodes}.")  # pragma: no cover
+
+        return _base_config, _node_qc_options
+
     def is_match_type(self, _type: Type) -> bool:
         """
         Check if input type matches the node type, either in instance type or in type name.
@@ -675,7 +690,7 @@ class BaseNode:
             return False
 
         # Get attributes from node to filter
-        layer_config = self.framework_attr.copy()
+        layer_config = self.framework_attr
         if hasattr(self, "op_call_kwargs"):
             layer_config.update(self.op_call_kwargs)
 
@@ -709,7 +724,7 @@ class BaseNode:
             Logger.critical(f"SIMD is expected to be a non-positive integer but found: {_simd}")
         return _simd
 
-    def sort_node_candidates(self):
+    def sort_node_candidates(self, fw_info):
         """
         Sorts the node candidates.
         We assume that the candidates are ordered in the following way (for mixed precision purposes):
@@ -718,12 +733,17 @@ class BaseNode:
             - If the node doesn't have a kernel we only consider the candidate activation number of bits to sort
             the candidates in descending order.
         The operation is done inplace.
+
+        Args:
+            fw_info: FrameworkInfo object about the specific framework (e.g., attributes of different layers' weights to quantize).
+
         """
-        if self.quantization_cfg.candidates_quantization_cfg is not None:
-            if self.kernel_attr is not None:
-                self.quantization_cfg.candidates_quantization_cfg.sort(
-                    key=lambda c: (c.weights_quantization_cfg.get_attr_config(self.kernel_attr).weights_n_bits,
+        if self.candidates_quantization_cfg is not None:
+            kernel_attr = fw_info.get_kernel_op_attributes(self.type)[0]
+            if kernel_attr is not None:
+                self.candidates_quantization_cfg.sort(
+                    key=lambda c: (c.weights_quantization_cfg.get_attr_config(kernel_attr).weights_n_bits,
                                    c.activation_quantization_cfg.activation_n_bits), reverse=True)
             else:
-                self.quantization_cfg.candidates_quantization_cfg.sort(
-                    key=lambda c: c.activation_quantization_cfg.activation_n_bits, reverse=True)
+                self.candidates_quantization_cfg.sort(key=lambda c: c.activation_quantization_cfg.activation_n_bits,
+                                                      reverse=True)

model_compression_toolkit/core/common/graph/functional_node.py

@@ -1,21 +1,6 @@
-# Copyright 2021 Sony Semiconductor Israel, Inc. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-# ==============================================================================
-
 from typing import Dict, Any, Tuple, Type, List, Union
 
-from model_compression_toolkit.core.common.framework_info import get_fw_info
+from model_compression_toolkit.verify_packages import FOUND_TF
 from model_compression_toolkit.core.common.graph.base_node import BaseNode
 import numpy as np
 
@@ -60,7 +45,6 @@ class FunctionalNode(BaseNode):
             inputs_as_list: Whether to pass the node its input tensors as a list or not when calling the layer.
             has_activation: Whether the node has activations that we might want to quantize.
             tensor_input_allocs: A list of indices and strings for allocations input tensors in the node's args and kwargs.
-
         """
 
         super().__init__(name,
@@ -79,7 +63,6 @@ class FunctionalNode(BaseNode):
         self.op_call_args = list(op_call_args)
         self.functional_op = functional_op
         self.tensor_input_allocs = [] if tensor_input_allocs is None else tensor_input_allocs
-        self.node_fw_info = self._get_fw_node_attrs(functional_op, framework_attr)
 
     @property
     def type(self):
@@ -103,4 +86,4 @@ class FunctionalNode(BaseNode):
 
         """
         names_match = _type.__name__ == self.type.__name__
-        return names_match or super().is_match_type(_type)
+        return super().is_match_type(_type) or names_match

model_compression_toolkit/core/common/graph/virtual_activation_weights_node.py

@@ -15,11 +15,13 @@
 import abc
 import uuid
 
+from model_compression_toolkit.core import FrameworkInfo
 from model_compression_toolkit.constants import VIRTUAL_ACTIVATION_WEIGHTS_NODE_PREFIX, \
     VIRTUAL_WEIGHTS_SUFFIX, VIRTUAL_ACTIVATION_SUFFIX, FLOAT_BITWIDTH
+from model_compression_toolkit.core.common.framework_info import DEFAULT_KERNEL_ATTRIBUTES
 from model_compression_toolkit.core.common.graph.base_node import BaseNode
 from model_compression_toolkit.core.common.quantization.candidate_node_quantization_config import \
-    CandidateNodeQuantizationConfig, NodeQuantizationConfig
+    CandidateNodeQuantizationConfig
 from model_compression_toolkit.core.common.quantization.node_quantization_config import ActivationQuantizationMode
 
 
@@ -75,11 +77,8 @@ class VirtualSplitWeightsNode(VirtualSplitNode):
 
         self.name = origin_node.name + VIRTUAL_WEIGHTS_SUFFIX
 
-        self.quantization_cfg = NodeQuantizationConfig(
-            candidates_quantization_cfg=origin_node.get_unique_weights_candidates(kernel_attr),
-            base_quantization_cfg=None, validate=False
-        )
-        for c in self.quantization_cfg.candidates_quantization_cfg:
+        self.candidates_quantization_cfg = origin_node.get_unique_weights_candidates(kernel_attr)
+        for c in self.candidates_quantization_cfg:
             c.activation_quantization_cfg.quant_mode = ActivationQuantizationMode.NO_QUANT
             c.activation_quantization_cfg.activation_n_bits = FLOAT_BITWIDTH
 
@@ -108,9 +107,10 @@ class VirtualSplitActivationNode(VirtualSplitNode):
         self.weights = {}
         self.layer_class = activation_class
 
-        self.quantization_cfg = NodeQuantizationConfig(candidates_quantization_cfg=origin_node.get_unique_activation_candidates(),
-                                                       base_quantization_cfg=None, validate=False)
-        self.quantization_cfg.disable_weights_quantization()
+        self.candidates_quantization_cfg = origin_node.get_unique_activation_candidates()
+        for c in self.candidates_quantization_cfg:
+            c.weights_quantization_cfg.enable_weights_quantization = False
+            c.weights_quantization_cfg.weights_n_bits = FLOAT_BITWIDTH
 
 
 class VirtualActivationWeightsNode(VirtualNode):
@@ -128,23 +128,28 @@ class VirtualActivationWeightsNode(VirtualNode):
 
     def __init__(self,
                  act_node: BaseNode,
-                 weights_node: BaseNode):
+                 weights_node: BaseNode,
+                 fw_info: FrameworkInfo):
         """
         Init a VirtualActivationWeightsNode object.
 
         Args:
             act_node: The original activation node.
             weights_node: The original weights node.
+            fw_info: A FrameworkInfo object with framework specific information.
         """
         # Validate weights node
+        kernel_attrs = fw_info.get_kernel_op_attributes(weights_node.type)
+        assert len(kernel_attrs) == 1 and kernel_attrs[0] is not None, f'Expected exactly one kernel attr, {kernel_attrs}'
+        kernel_attr = kernel_attrs[0]
         conf_weights = [attr for attr in weights_node.weights if weights_node.is_configurable_weight(attr)]
-        if len(conf_weights) > 1 or len(conf_weights) == 1 and not weights_node.is_configurable_weight(weights_node.kernel_attr):
+        if len(conf_weights) > 1 or len(conf_weights) == 1 and not weights_node.is_configurable_weight(kernel_attr):
             raise NotImplementedError(f'Only kernel weight can be configurable. Got configurable {conf_weights}.')
 
         weights = weights_node.weights.copy()
         act_node_w_rename = {}
         if act_node.weights:
-            if act_node.kernel_attr:
+            if fw_info.get_kernel_op_attributes(act_node) != DEFAULT_KERNEL_ATTRIBUTES:
                 raise NotImplementedError(f'Node {act_node} with kernel cannot be used as activation for '
                                           f'VirtualActivationWeightsNode.')
             if act_node.has_any_configurable_weight():
@@ -152,7 +157,7 @@ class VirtualActivationWeightsNode(VirtualNode):
                                           'VirtualActivationWeightsNode.')
             # combine weights from activation and weights
             for w_id, w in act_node.weights.items():
-                if w_id not in weights and not (isinstance(w_id, str) and weights_node.kernel_attr in w_id):
+                if w_id not in weights and not (isinstance(w_id, str) and kernel_attr in w_id):
                     weights[w_id] = w
                     continue
                 # if same identifier is used as in weight nodes (or contains the kernel substring), generate a new
@@ -180,7 +185,7 @@ class VirtualActivationWeightsNode(VirtualNode):
         self.original_weights_node = weights_node
 
         v_candidates = []
-        weights_candidates_quantization_cfg = weights_node.get_unique_weights_candidates(weights_node.kernel_attr)
+        weights_candidates_quantization_cfg = weights_node.get_unique_weights_candidates(kernel_attr)
         for c_a in act_node.candidates_quantization_cfg:
             for c_w in weights_candidates_quantization_cfg:
                 composed_candidate = CandidateNodeQuantizationConfig(activation_quantization_cfg=c_a.activation_quantization_cfg,
@@ -198,8 +203,7 @@ class VirtualActivationWeightsNode(VirtualNode):
                 v_candidates.append(composed_candidate)
 
         # sorting the candidates by weights number of bits first and then by activation number of bits (reversed order)
-        v_candidates.sort(key=lambda c: (c.weights_quantization_cfg.get_attr_config(weights_node.kernel_attr).weights_n_bits,
+        v_candidates.sort(key=lambda c: (c.weights_quantization_cfg.get_attr_config(kernel_attr).weights_n_bits,
                                          c.activation_quantization_cfg.activation_n_bits), reverse=True)
 
-        self.quantization_cfg = NodeQuantizationConfig(candidates_quantization_cfg=v_candidates,
-                                                       base_quantization_cfg=None, validate=False)
+        self.candidates_quantization_cfg = v_candidates

model_compression_toolkit/core/common/mixed_precision/bit_width_setter.py

@@ -37,18 +37,20 @@ def set_bit_widths(mixed_precision_enable: bool,
     """
     if mixed_precision_enable:
         assert all([len(n.candidates_quantization_cfg) > 0
-                    for n in graph.get_configurable_sorted_nodes()]), \
+                    for n in graph.get_configurable_sorted_nodes(graph.fw_info)]), \
             "All configurable nodes in graph should have at least one candidate configuration in mixed precision mode"
 
         # Get a list of nodes' names we need to finalize (that they have at least one weight qc candidate).
-        sorted_nodes_names = graph.get_configurable_sorted_nodes_names()
+        sorted_nodes_names = graph.get_configurable_sorted_nodes_names(graph.fw_info)
 
         for node in graph.nodes:  # set a specific node qc for each node final qc
             # If it's reused, take the configuration that the base node has
             node_name = node.name if not node.reuse else '_'.join(node.name.split('_')[:-2])
             if node_name in sorted_nodes_names:  # only configurable nodes are in this list
                 node_index_in_graph = sorted_nodes_names.index(node_name)
-                _set_node_final_qc(bit_widths_config[node_index_in_graph], node)
+                _set_node_final_qc(bit_widths_config[node_index_in_graph],
+                                   node,
+                                   graph.fw_info)
             else:
                 if node.is_activation_quantization_enabled():
                     # If we are here, this means that we are in weights-only mixed-precision
@@ -81,7 +83,8 @@ def set_bit_widths(mixed_precision_enable: bool,
 
 
 def _get_node_qc_by_bit_widths(node: BaseNode,
-                               node_bit_width_cfg: int) -> Any:
+                               node_bit_width_cfg: int,
+                               fw_info) -> Any:
     """
     Get the node's quantization configuration that
     matches to the bit width index as in the MP configuration bit_width_cfg.
@@ -90,18 +93,21 @@ def _get_node_qc_by_bit_widths(node: BaseNode,
     Args:
         node: Node to get its quantization configuration candidate.
         node_bit_width_cfg: Configuration which determines the node's desired bit width.
+        fw_info: Information relevant to a specific framework about how layers should be quantized.
 
     Returns:
         Node quantization configuration if it was found, or None otherwise.
     """
     # only the weights kernel attribute is quantized in weights mixed precision at the moment
+    kernel_attr = fw_info.get_kernel_op_attributes(node.type)
+
     if node.is_activation_quantization_enabled():
         qc = node.candidates_quantization_cfg[node_bit_width_cfg]
 
         return qc
 
-    elif node.kernel_attr is not None:
-        if node.is_weights_quantization_enabled(node.kernel_attr):
+    elif kernel_attr is not None:
+        if node.is_weights_quantization_enabled(kernel_attr[0]):
             qc = node.candidates_quantization_cfg[node_bit_width_cfg]
 
             return qc
@@ -110,7 +116,8 @@ def _get_node_qc_by_bit_widths(node: BaseNode,
 
 
 def _set_node_final_qc(node_bit_width_cfg: int,
-                       node: BaseNode):
+                       node: BaseNode,
+                       fw_info):
     """
     Get the node's quantization configuration that
     matches to the bit width index as in the MP configuration bit_width_cfg, and use it to finalize the node's
@@ -120,9 +127,12 @@ def _set_node_final_qc(node_bit_width_cfg: int,
     Args:
         node_bit_width_cfg: Configuration which determines the node's desired bit width.
         node: Node to set its node quantization configuration.
+        fw_info: Information relevant to a specific framework about how layers should be quantized.
 
     """
-    node_qc = _get_node_qc_by_bit_widths(node, node_bit_width_cfg)
+    node_qc = _get_node_qc_by_bit_widths(node,
+                                         node_bit_width_cfg,
+                                         fw_info)
 
     if node_qc is None:
         Logger.critical(f'Node {node.name} quantization configuration from configuration file'  # pragma: no cover