OVWeightUpdateCommand copies a constant (#2277)

### Changes

- Update an implementation of `get_const_value()` method. This method
was initially implemented as a workaround in [PR
1654](#1654), but it is no
longer needed and can be replaced with `const_op.data`.
- Replace `const_op.get_data()` with `const_op.data` because the
`const_op.get_data()` legacy method and `.data` property should be used.
- Set the `shared_memory` parameter to `True` during constant creation
to avoid copying the constant.

### Reason for changes

- The `const_od.get_data()` method copies constant.
- Do not copy the constant value during constant creation in
`opset.constant(...)`.

### Related tickets

Ref: 122922

### Tests

N/A

---------

Co-authored-by: Nikita Malinin <[email protected]>

nncf/openvino/graph/model_transformer.py

@@ -354,9 +354,12 @@ def _set_const_value(node_with_const: ov.Node, const_port_id: int, const_value:
         if const_node is None:
             raise RuntimeError("Constant node was expected but could not find it.")
 
-        const_shape = const_node.get_data().shape
-        const_value = np.reshape(const_value, const_shape)
-        new_const_node = opset.constant(const_value, dtype=const_node.get_element_type())
+        const_shape = const_node.data.shape
+        const_dtype = const_node.data.dtype
+        const_value = np.reshape(const_value, const_shape).astype(const_dtype)
+
+        # TODO(andrey-churkin): Replace on opset13.constant() in a future release
+        new_const_node = ov.op.Constant(const_value, shared_memory=True)
         new_const_node.set_friendly_name(const_node.get_friendly_name())
         const_port.replace_source_output(new_const_node.output(0))
 

nncf/openvino/graph/node_utils.py

@@ -78,7 +78,7 @@ def get_const_value(const_node: ov.Node) -> np.ndarray:
     :param const_node: OpenVINO node.
     :return: The constant value.
     """
-    return const_node.get_vector().reshape(const_node.get_output_shape(0))
+    return const_node.data
 
 
 def get_bias_value(node_with_bias: NNCFNode, nncf_graph: NNCFGraph, model: ov.Model) -> np.ndarray:

tests/openvino/native/quantization/test_fq_params_calculation.py

@@ -42,10 +42,10 @@ def get_fq_nodes_stats_algo(model):
     nodes = {}
     for op in model.get_ops():
         if op.get_type_name() == "FakeQuantize":
-            input_low = op.input_value(1).get_node().get_data()
-            input_high = op.input_value(2).get_node().get_data()
-            output_low = op.input_value(3).get_node().get_data()
-            output_high = op.input_value(4).get_node().get_data()
+            input_low = op.input_value(1).get_node().data
+            input_high = op.input_value(2).get_node().data
+            output_low = op.input_value(3).get_node().data
+            output_high = op.input_value(4).get_node().data
 
             nodes[op.get_friendly_name()] = {
                 "input_low": input_low,

tests/openvino/native/test_model_transformer.py

@@ -182,13 +182,13 @@ def check_inplace_op(target_node, ref_types, ref_vals, inplace_branches_num, out
         if ref_val is not None:
             const = get_prev_node(node, 1)
             if ref_val == []:
-                assert const.get_data().shape == (0,)
+                assert const.data.shape == (0,)
             elif not isinstance(ref_val, tuple):
-                assert const.get_data() == ref_val
+                assert const.data == ref_val
             else:
-                res = np.equal(const.get_data(), np.array(ref_val))
+                res = np.equal(const.data, np.array(ref_val))
                 assert all(res)
-                assert const.get_data().shape == np.array(ref_val).shape
+                assert const.data.shape == np.array(ref_val).shape
 
         nodes = get_next_nodes(node, 0)
         assert len(nodes) == 1
@@ -296,7 +296,7 @@ def test_inplace_reduce_fn_dynamic_shapes(input_shape, raise_error):
     op = fn(input_1, 0)
     # check_const
     ref_const = np.array([0, 1, 2, 3])
-    assert all(np.equal(get_prev_node(op, 1).get_data(), ref_const))
+    assert all(np.equal(get_prev_node(op, 1).data, ref_const))
 
 
 @pytest.mark.parametrize("reduction_axes", [None, np.array([], dtype=np.int64)])