[ONNX] Add eliminate_nop_cast pass (#3376)

### Changes

Apply the `eliminate_nop_cast` optimization pass before building the
NNCF graph. This pass aims to eliminate no-op cast nodes in the graph. A
no-op cast is a cast that doesn't change the value, such as casting a
tensor to its own type.

### Reason for changes

Statistics cannot be collected after the no-op Cast nodes because such
nodes are removed from the ONNX inference graph during the session.

### Related tickets

Ref: 164211

Author: andrey-churkin

nncf/onnx/graph/passes.py

@@ -0,0 +1,79 @@
+# Copyright (c) 2025 Intel Corporation
+# 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.
+
+import onnx
+
+from nncf.onnx.graph.onnx_helper import get_children
+from nncf.onnx.graph.onnx_helper import get_children_node_mapping
+
+
+def eliminate_nop_cast(model: onnx.ModelProto) -> onnx.ModelProto:
+    """
+    Inspects the provided ONNX model to identify and remove any 'No-op' (no-operation)
+    cast nodes, which are operations that do not change the data type of their input.
+
+    :param model: The ONNX model to be processed.
+    :return: The ONNX model with the redundant cast nodes removed.
+    """
+    tensor_name_to_info = {
+        tensor.name: tensor
+        for tensor in (*model.graph.value_info, *model.graph.input, *model.graph.output, *model.graph.initializer)
+    }
+    redundant_cast_nodes = []
+    for node in model.graph.node:
+        if node.op_type == "Cast":
+            to_attr = None
+            for attr in node.attribute:
+                if attr.name == "to":
+                    to_attr = onnx.helper.get_attribute_value(attr)
+
+            if to_attr is None:
+                continue
+
+            inp = node.input[0]
+            info = tensor_name_to_info[inp]
+            if info.type.tensor_type.elem_type == to_attr:
+                redundant_cast_nodes.append(node)
+
+    value_infos = {i.name: i for i in model.graph.value_info}
+    input_name_to_nodes_map = get_children_node_mapping(model)
+
+    for cast_node in redundant_cast_nodes:
+        # Unlink Cast node from the graph
+        children = get_children(cast_node, input_name_to_nodes_map)
+        for child in children:
+            for i, input_name in enumerate(child.input):
+                if input_name == cast_node.output[0]:
+                    child.input[i] = cast_node.input[0]
+
+        # Remove Cast node from the graph
+        model.graph.value_info.remove(value_infos[cast_node.output[0]])
+        model.graph.node.remove(cast_node)
+
+    return model
+
+
+def apply_preprocess_passes(model: onnx.ModelProto) -> onnx.ModelProto:
+    """
+    Preprocesses the provided ONNX model for quantization.
+
+    This method performs the following steps:
+        1. Infers shapes in the model.
+        2. Removes redundant 'No-op' cast nodes from the model.
+
+    :param model: The ONNX model to be preprocessed.
+    :return: A preprocessed ONNX model, ready for quantization.
+    """
+    preprocessed_model = onnx.shape_inference.infer_shapes(model)
+    # The `eliminate_nop_cast` pass should be applied after onnx.shape_inference.infer_shapes() call.
+    # Otherwise, not all no-op Cast nodes will be found.
+    preprocessed_model = eliminate_nop_cast(preprocessed_model)
+    return preprocessed_model

nncf/onnx/quantization/quantize_model.py

@@ -29,6 +29,7 @@
 from nncf.onnx.graph.model_metadata import remove_metadata
 from nncf.onnx.graph.model_metadata import set_metadata
 from nncf.onnx.graph.nncf_graph_builder import GraphConverter
+from nncf.onnx.graph.passes import apply_preprocess_passes
 from nncf.onnx.quantization.backend_parameters import get_external_data_dir
 from nncf.parameters import BackupMode
 from nncf.parameters import CompressionFormat
@@ -155,6 +156,7 @@ def quantize_impl(
     external_data_dir = check_external_data_location(model, external_data_dir)
     if external_data_dir:
         set_metadata(model, MetadataKey.EXTERNAL_DATA_DIR, external_data_dir)
+    model = apply_preprocess_passes(model)
 
     quantization_algorithm = PostTrainingQuantization(
         preset=preset,

tests/onnx/models.py

@@ -1897,3 +1897,22 @@ def build_matmul_model() -> onnx.ModelProto:
     graph = onnx.helper.make_graph([matmul], "matmul-model", [X], [A], [W_initializer])
     model = onnx.helper.make_model(graph)
     return model
+
+
+def build_matmul_model_with_nop_cast() -> onnx.ModelProto:
+    """
+    Builds an ONNX model that contains a MatMul operation with a no-op Cast applied to the input.
+    """
+    X = onnx.helper.make_tensor_value_info("X", onnx.TensorProto.FLOAT, [2, 3])
+    A = onnx.helper.make_tensor_value_info("A", onnx.TensorProto.FLOAT, [2, 2])
+    cast = onnx.helper.make_node("Cast", inputs=["X"], outputs=["X_cast"], name="cast", to=onnx.TensorProto.FLOAT)
+    matmul = onnx.helper.make_node("MatMul", inputs=["X_cast", "W"], outputs=["A"], name="matmul")
+
+    W_values = np.random.rand(3, 2).astype(np.float32)
+    W_initializer = onnx.helper.make_tensor(
+        name="W", data_type=onnx.TensorProto.FLOAT, dims=[3, 2], vals=W_values.tobytes(), raw=True
+    )
+
+    graph = onnx.helper.make_graph([cast, matmul], "matmul-model", [X], [A], [W_initializer])
+    model = onnx.helper.make_model(graph)
+    return model

tests/onnx/test_passes.py

@@ -0,0 +1,23 @@
+# Copyright (c) 2025 Intel Corporation
+# 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 nncf.onnx.graph.passes import apply_preprocess_passes
+from tests.onnx.models import build_matmul_model_with_nop_cast
+
+
+def test_apply_preprocess_passes():
+    model = build_matmul_model_with_nop_cast()
+    before_nodes = [node.name for node in model.graph.node]
+    preprocessed_model = apply_preprocess_passes(model)
+    after_nodes = [node.name for node in preprocessed_model.graph.node]
+
+    assert set(after_nodes) - set(before_nodes) == set()
+    assert set(before_nodes) - set(after_nodes) == set(["cast"])