[ONNX] Add MatMul for FBC/BC (#3657)

andrey-churkin · web-flow · commit 9faf5592d11c · 2025-10-27T06:47:08.000Z
### Changes Add support for the `MatMul` operation in the Fast Bias/Bias Correction algorithm. ### Reason for changes The Fast Bias/Bias Correction algorithm is not applied to the `MatMul`->`Add` subgraph in which one of the inputs to the Add operation is a constant. In this case, the `MatMul` operation is not considered as having a bias, so the algorithm is not applied. ### Related tickets Ref: 135433 ### Tests - test_update_bias_in_matmul_add - text_examples: [build](https://github.com/openvinotoolkit/nncf/actions/runs/18529108467)
diff --git a/src/nncf/onnx/graph/metatypes/groups.py b/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 = [
diff --git a/src/nncf/onnx/graph/metatypes/onnx_metatypes.py b/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
 
diff --git a/src/nncf/onnx/graph/nncf_graph_builder.py b/src/nncf/onnx/graph/nncf_graph_builder.py
@@ -26,6 +26,7 @@
 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 ONNXGemmMetatype
+from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXMatMulMetatype
 from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXOpMetatype
 from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXOpWithWeightsMetatype
 from nncf.onnx.graph.metatypes.onnx_metatypes import get_metatype
@@ -186,17 +187,54 @@ def _get_bias_attr(
     node: onnx.NodeProto,
     model: onnx.ModelProto,
     parents_node_mapping: dict[str, onnx.NodeProto],
+    children_node_mapping: dict[str, onnx.NodeProto],
 ) -> dict[str, str]:
     """
     Returns bias tensor attributes.
 
     :param node: ONNX node.
     :param model: ONNX model.
     :param parents_node_mapping: Mapping from edge name to node which outputs this edge.
+    :param children_node_mapping: mapping from edge name to nodes which consume this edge as an input.
     :return: Bias tensor attributes.
     """
-    bias_attrs = {}
     metatype = get_metatype(model, node)
+
+    if metatype == ONNXMatMulMetatype:
+        weight_port_ids = _get_weight_port_ids(node, model, parents_node_mapping)
+
+        if not weight_port_ids:
+            # `node` is a MatMul without weights, so return empty attributes
+            return {}
+
+        # Retrieve all nodes that consume the output of the MatMul operation.
+        # The MatMul operation has only one output.
+        y = node.output[0]
+        consumers = children_node_mapping[y]
+
+        if len(consumers) != 1 or consumers[0].op_type != "Add":
+            return {}
+
+        # Here, we are certain that after a `MatMul` operation, there is only
+        # the `Add` operation.
+        add_node = consumers[0]
+
+        # Find the input of `add_node` that is not equal to `y`.
+        tensor_name = None
+        port_id = None
+        for i, name in enumerate(add_node.input):
+            if name != y:
+                tensor_name = name
+                port_id = i
+                break
+
+        # Ensure that `tensor_name` is the output of a `Constant` node or an initializer.
+        initializer = {x.name: x for x in model.graph.initializer}
+        if tensor_name in initializer or parents_node_mapping[tensor_name].op_type == "Constant":
+            return {"node": add_node.name, "name": tensor_name, "port_id": port_id}
+        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)
@@ -348,6 +386,7 @@ def create_nncf_graph(onnx_model: onnx.ModelProto) -> NNCFGraph:
         """
         onnx_model = GraphConverter._replace_empty_node_name(onnx_model)
         onnx_model = onnx.shape_inference.infer_shapes(onnx_model)
+
         edge_info_mapping = get_edge_info_mapping(onnx_model)
         children_node_mapping = get_children_node_mapping(onnx_model)
         parents_node_mapping = get_parents_node_mapping(onnx_model)
@@ -358,7 +397,8 @@ 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:
diff --git a/src/nncf/onnx/graph/node_utils.py b/src/nncf/onnx/graph/node_utils.py
@@ -45,6 +45,7 @@ def get_bias_value(node_with_bias: NNCFNode, model: onnx.ModelProto) -> np.ndarr
     :return: The bias value that is applied to the output tensor of the node's operation.
     """
     assert node_with_bias.layer_attributes.has_bias()
+    # TODO(andrey-churkin): Support Add + Constant case
     bias_name = node_with_bias.layer_attributes.bias_attrs["name"]
     return get_tensor_value(model, bias_name)
 
diff --git a/src/nncf/onnx/graph/transformations/command_creation.py b/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("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)
 
 
diff --git a/src/nncf/quantization/algorithms/fast_bias_correction/algorithm.py b/src/nncf/quantization/algorithms/fast_bias_correction/algorithm.py
@@ -175,13 +175,11 @@ def apply(
 
             output_channel_axis = node.metatype.output_channel_axis
             input_channel_axis = self._backend_entity.get_activation_channel_axis(node, input_port_id, input_shape)
-            if bias_value.ndim > 1:
-                # Make index positive
-                output_channel_axis = range(bias_value.ndim)[output_channel_axis]
-                input_channel_axis = range(bias_value.ndim)[input_channel_axis]
+
             input_blob = self._backend_entity.create_input_data(
                 input_shape, input_fp, sub_input_name, input_channel_axis
             )
+
             bias_shift = self._get_bias_shift(
                 model=extracted_model,
                 input_blob=input_blob,
diff --git a/src/nncf/quantization/algorithms/fast_bias_correction/onnx_backend.py b/src/nncf/quantization/algorithms/fast_bias_correction/onnx_backend.py
@@ -67,6 +67,7 @@ def get_sub_input_output_names(subgraph: onnx.ModelProto) -> tuple[str, str]:
     def create_input_data(
         shape: tuple[int], data: list[Tensor], input_name: str, channel_axis: int
     ) -> dict[str, np.array]:
+        channel_axis = range(len(shape))[channel_axis]
         blob = np.zeros(shape, dtype=data[0].data.dtype)
         for j, idx in enumerate(np.ndindex(blob.shape[channel_axis])):
             index = tuple(slice(None) if i != channel_axis else idx for i in range(blob.ndim))
diff --git a/src/nncf/quantization/algorithms/fast_bias_correction/openvino_backend.py b/src/nncf/quantization/algorithms/fast_bias_correction/openvino_backend.py
@@ -72,6 +72,7 @@ def get_sub_input_output_names(subgraph: ov.Model) -> tuple[str, str]:
     def create_input_data(
         shape: tuple[int], data: list[Tensor], input_name: str, channel_axis: int
     ) -> dict[str, np.ndarray]:
+        channel_axis = range(len(shape))[channel_axis]
         blob = np.zeros(shape, dtype=data[0].data.dtype)
         for j, idx in enumerate(np.ndindex(blob.shape[channel_axis])):
             index = tuple(slice(None) if i != channel_axis else idx for i in range(blob.ndim))
diff --git a/src/nncf/quantization/algorithms/fast_bias_correction/torch_backend.py b/src/nncf/quantization/algorithms/fast_bias_correction/torch_backend.py
@@ -78,6 +78,7 @@ def get_sub_input_output_names(subgraph: NNCFNetwork) -> tuple[Optional[str], Op
 
     @staticmethod
     def create_input_data(shape: tuple[int], data: list[Tensor], input_name: str, channel_axis: int) -> torch.Tensor:
+        channel_axis = range(len(shape))[channel_axis]
         blob = torch.zeros(shape, dtype=data[0].data.dtype, device=data[0].data.device)
         for j, idx in enumerate(np.ndindex(blob.shape[channel_axis])):
             index = tuple(slice(None) if i != channel_axis else idx for i in range(blob.ndim))
diff --git a/src/nncf/quantization/algorithms/fast_bias_correction/torch_fx_backend.py b/src/nncf/quantization/algorithms/fast_bias_correction/torch_fx_backend.py
@@ -67,6 +67,7 @@ def get_sub_input_output_names(subgraph: torch.fx.GraphModule) -> tuple[Optional
 
     @staticmethod
     def create_input_data(shape: tuple[int], data: list[Tensor], input_name: str, channel_axis: int) -> torch.Tensor:
+        channel_axis = range(len(shape))[channel_axis]
         blob = torch.zeros(shape, dtype=data[0].data.dtype, device=data[0].data.device)
         for j, idx in enumerate(np.ndindex(blob.shape[channel_axis])):
             index = tuple(slice(None) if i != channel_axis else idx for i in range(blob.ndim))
diff --git a/tests/onnx/common.py b/tests/onnx/common.py
@@ -155,6 +155,15 @@ def add_mul(self, input_a: str, input_b: str, output: Optional[str] = None) -> s
         )
         return output
 
+    def add_add(self, input_a: str, input_b: str, output: Optional[str] = None) -> str:
+        i = len(self._nodes)
+
+        output = f"Add_{i}_output" if output is None else output
+        self._nodes.append(
+            onnx.helper.make_node(op_type="Add", inputs=[input_a, input_b], outputs=[output], name=f"Add_{i}")
+        )
+        return output
+
     def add_relu(self, input: str, output: Optional[str] = None) -> str:
         i = len(self._nodes)
 
diff --git a/tests/onnx/quantization/test_fast_bias_correction.py b/tests/onnx/quantization/test_fast_bias_correction.py
@@ -14,11 +14,15 @@
 import onnx
 import torch
 
+from nncf.common.factory import ModelTransformerFactory
 from nncf.common.factory import NNCFGraphFactory
+from nncf.common.graph.transformations.layout import TransformationLayout
 from nncf.onnx.graph.node_utils import get_bias_value
 from nncf.onnx.graph.node_utils import is_node_with_bias
+from nncf.onnx.graph.transformations.command_creation import ONNXCommandCreator
 from nncf.quantization.algorithms.fast_bias_correction.onnx_backend import ONNXFastBiasCorrectionAlgoBackend
 from tests.cross_fw.test_templates.test_fast_bias_correction import TemplateTestFBCAlgorithm
+from tests.onnx.common import ModelBuilder
 
 
 def get_data_from_node(model: onnx.ModelProto, node_name: str):
@@ -71,3 +75,47 @@ def check_bias(model: onnx.ModelProto, ref_bias: list):
             return
         msg = "Not found node with bias"
         raise ValueError(msg)
+
+
+def _build_matmul_add_model() -> onnx.ModelProto:
+    mb = ModelBuilder()
+
+    x = mb.add_input("X", (2, 3))
+
+    x = mb.add_matmul(x, (3, 3))
+    x = mb.add_add(x, mb.add_initializer(np.array([1, 1, 1], dtype=np.float32)))
+
+    x = mb.add_matmul(x, (3, 3))
+    x = mb.add_add(x, mb.add_initializer(np.array([2, 2, 2], dtype=np.float32)))
+
+    mb.add_output(x, (2, 3))
+
+    return mb.build(opset_version=19, ir_version=9)
+
+
+def test_update_bias_in_matmul_add():
+    """
+    Tests the ability to retrieve and update the value of the bias constant in a MatMul->Add subgraph,
+    where the second input to the Add operation is a constant.
+    """
+    model = _build_matmul_add_model()
+    graph = NNCFGraphFactory.create(model)
+
+    nodes = [node for node in graph.get_all_nodes() if is_node_with_bias(node)]
+    assert [x.node_name for x in nodes] == ["MatMul_0", "MatMul_2"]
+
+    for matmul, data in zip(nodes, [[1, 1, 1], [2, 2, 2]]):
+        bias = get_bias_value(matmul, model)
+        bias_ref = np.array(data, dtype=np.float32)
+        assert np.all(np.isclose(bias, bias_ref, atol=0.0001)), f"{bias} != {bias_ref}"
+
+    layout = TransformationLayout()
+    for matmul, data in zip(nodes, [[2, 2, 2], [1, 1, 1]]):
+        new_bias = np.array(data, dtype=np.float32)
+        layout.register(ONNXCommandCreator.create_command_to_update_bias(matmul, new_bias, graph))
+    model = ModelTransformerFactory.create(model).transform(layout)
+
+    for matmul, data in zip(nodes, [[2, 2, 2], [1, 1, 1]]):
+        bias = get_bias_value(matmul, model)
+        bias_ref = np.array(data, dtype=np.float32)
+        assert np.all(np.isclose(bias, bias_ref, atol=0.0001)), f"{bias} != {bias_ref}"
diff --git a/tests/post_training/data/ptq_reference_data.yaml b/tests/post_training/data/ptq_reference_data.yaml
@@ -139,7 +139,7 @@ timm/deit3_small_patch16_224_backend_CUDA_TORCH:
 timm/deit3_small_patch16_224_backend_FP32:
   metric_value: 0.81358
 timm/deit3_small_patch16_224_backend_ONNX:
-  metric_value: 0.81116
+  metric_value: 0.81156
 timm/deit3_small_patch16_224_backend_OV:
   metric_value: 0.81276
 timm/deit3_small_patch16_224_backend_TORCH:

Original file line number	Diff line number	Diff line change
`@@ -139,7 +139,7 @@`
`139`	`139`	`OPERATIONS_WITH_BIAS_REDUCED = [`
`140`	`140`	`onnx_metatypes.ONNXConvolutionMetatype,`
`141`	`141`	`onnx_metatypes.ONNXGemmMetatype,`
`142`		`- # TODO: Need to add MatMul with the separate bias support (CVS-135433)`
	`142`	`+ onnx_metatypes.ONNXMatMulMetatype,`
`143`	`143`	`]`
`144`	`144`
`145`	`145`	`OPERATIONS_WITH_BIAS = [`