add the raw version

Author: andrey-churkin

src/nncf/onnx/graph/metatypes/groups.py

@@ -139,7 +139,7 @@
 OPERATIONS_WITH_BIAS_REDUCED = [
     onnx_metatypes.ONNXConvolutionMetatype,
     onnx_metatypes.ONNXGemmMetatype,
-    # TODO: Need to add MatMul with the separate bias support (CVS-135433)
+    onnx_metatypes.ONNXMatMulMetatype,
 ]
 
 OPERATIONS_WITH_BIAS = [

src/nncf/onnx/graph/metatypes/onnx_metatypes.py

@@ -143,7 +143,6 @@ class ONNXMatMulMetatype(ONNXOpWithWeightsMetatype):
     op_names = ["MatMul"]
     hw_config_names = [HWConfigOpName.MATMUL]
     weight_channel_axis = -1  # For port_id=1
-    bias_port_id = 2
     possible_weight_ports = [0, 1]
     output_channel_axis = -1
 

src/nncf/onnx/graph/nncf_graph_builder.py

@@ -25,6 +25,7 @@
 from nncf.onnx.graph.metatypes.groups import CONSTANT_WEIGHT_LAYER_METATYPES
 from nncf.onnx.graph.metatypes.groups import OPERATIONS_WITH_BIAS
 from nncf.onnx.graph.metatypes.groups import POSSIBLE_WEIGHT_LAYER_METATYPES
+from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXMatMulMetatype
 from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXGemmMetatype
 from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXOpMetatype
 from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXOpWithWeightsMetatype
@@ -186,6 +187,7 @@ def _get_bias_attr(
     node: onnx.NodeProto,
     model: onnx.ModelProto,
     parents_node_mapping: dict[str, onnx.NodeProto],
+    children_node_mapping
 ) -> dict[str, str]:
     """
     Returns bias tensor attributes.
@@ -197,6 +199,25 @@ def _get_bias_attr(
     """
     bias_attrs = {}
     metatype = get_metatype(model, node)
+
+    if metatype == ONNXMatMulMetatype:
+        weight_port_ids = _get_weight_port_ids(node, model, parents_node_mapping)
+        if weight_port_ids:
+            y = node.output[0]  # only 1 output
+            consumers = children_node_mapping[y]
+            if len(consumers) == 1 and consumers[0].op_type == "Add":
+                add_node = consumers[0]
+                for port_id, input_name in enumerate(add_node.input):
+                    if input_name != y:
+                        bias_attrs["name"] = input_name
+                        bias_attrs["port_id"] = port_id
+
+                bias_attrs["add_node"] = add_node.name
+
+                # `name` should be or output from constant or `initializer`
+
+        return bias_attrs
+
     if _is_node_with_bias(node, model):
         bias_tensor_port_id = get_bias_tensor_port_id(metatype)
         bias_edge_name = get_tensor_edge_name(model, node, bias_tensor_port_id, parents_node_mapping)
@@ -358,7 +379,8 @@ def create_nncf_graph(onnx_model: onnx.ModelProto) -> NNCFGraph:
             is_shared = None
             weight_attrs = {}
             node_attrs = _get_node_attrs(node, onnx_model)
-            bias_attrs = _get_bias_attr(node, onnx_model, parents_node_mapping)
+            bias_attrs = _get_bias_attr(node, onnx_model, parents_node_mapping, children_node_mapping)
+
             if weight_port_ids:  # If node has weight
                 weight_edge_names = []
                 for weight_port_id in weight_port_ids:

src/nncf/onnx/graph/transformations/command_creation.py

@@ -29,8 +29,13 @@ def create_bias_correction_command(node: NNCFNode, bias_value: np.ndarray) -> ON
     :param bias_value: The new bias value that will be set.
     :return: The `ONNXInitializerUpdateCommand` command to update bias.
     """
-    bias_port_id = node.metatype.bias_port_id
-    target_point = ONNXTargetPoint(TargetType.LAYER, node.node_name, bias_port_id)
+    node_name = node.layer_attributes.bias_attrs.get("add_node")
+    if node_name:
+        port_id =  node.layer_attributes.bias_attrs["port_id"]
+        target_point = ONNXTargetPoint(TargetType.LAYER, node_name, port_id)
+    else:
+        bias_port_id = node.metatype.bias_port_id
+        target_point = ONNXTargetPoint(TargetType.LAYER, node.node_name, bias_port_id)
     return ONNXInitializerUpdateCommand(target_point, bias_value)
 
 

src/nncf/quantization/algorithms/fast_bias_correction/onnx_backend.py

@@ -104,6 +104,10 @@ def is_node_with_bias(node: NNCFNode, nncf_graph: NNCFGraph) -> bool:
 
     @staticmethod
     def get_node_names_for_input_output_statistics(node: NNCFNode, nncf_graph: NNCFGraph) -> tuple[str, str]:
+        output = node.layer_attributes.bias_attrs.get("add_node", None)
+        if output:
+            return node.node_name, output
+
         return node.node_name, node.node_name
 
     @staticmethod