[Transformation] LoraSubgraph fusion (#27068)

### Details:
- *Introduced `LoraSubgraph` operation, which is used for LoRA subgraphs
fusion for further optimizations (for the details, please refer to the
description in the header)*
 - *Introduced `LoraSubgraphFusion` pass*
 - *The changes are covered by transformation tests*

### Tickets:
 - *CVS-153035*
 - *CVS-155112*

Author: v-Golubev

src/common/transformations/include/ov_ops/lora_subgraph.hpp

@@ -0,0 +1,38 @@
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include "openvino/op/op.hpp"
+#include "openvino/op/util/sub_graph_base.hpp"
+#include "transformations_visibility.hpp"
+
+namespace ov {
+namespace op {
+namespace internal {
+/**
+ * @interface LoraSubgraph
+ * @brief LoraSubgraph operation, which is used for LoRA subgraphs fusion.
+ * It always has only 1 output, and the following inputs, whose order is fixed:
+ * 1. main_flow_input: input from original model.
+ * 2. LoRA_input: input to which the Low-Rank adaptation is applied.
+ *    The adapted input is combined with `main_flow_input`.
+ * 3. LoRA_matrices: 3 Low-Rank adaptation matrices applied to `LoRA_input`.
+ * The fused subgraph can be optimized in runtime based on LoRA semantic.
+ * For instance, `main_flow_input` can be fast-forwarded to output in case of empty `LoRA_matrices`.
+ */
+class TRANSFORMATIONS_API LoraSubgraph : public ov::op::util::SubGraphOp {
+public:
+    OPENVINO_OP("LoraSubgraph", "ie_internal_opset", ov::op::util::SubGraphOp);
+
+    LoraSubgraph() = default;
+    LoraSubgraph(const OutputVector& args, const std::shared_ptr<ov::Model>& body);
+
+    void validate_and_infer_types() override;
+    std::shared_ptr<Node> clone_with_new_inputs(const ov::OutputVector& new_args) const override;
+};
+
+}  // namespace internal
+}  // namespace op
+}  // namespace ov

src/common/transformations/include/transformations/common_optimizations/lora_subgraph_fusion.hpp

@@ -0,0 +1,25 @@
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include <memory>
+#include <vector>
+
+#include "openvino/pass/matcher_pass.hpp"
+#include "transformations_visibility.hpp"
+
+namespace ov {
+namespace pass {
+
+class TRANSFORMATIONS_API LoraSubgraphFusion;
+
+}  // namespace pass
+}  // namespace ov
+
+class ov::pass::LoraSubgraphFusion : public ov::pass::MatcherPass {
+public:
+    OPENVINO_RTTI("LoraSubgraphFusion", "0");
+    LoraSubgraphFusion();
+};

src/common/transformations/src/ov_ops/lora_subgraph.cpp

@@ -0,0 +1,41 @@
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "ov_ops/lora_subgraph.hpp"
+
+#include "itt.hpp"
+
+namespace ov {
+namespace op {
+namespace internal {
+
+LoraSubgraph::LoraSubgraph(const OutputVector& args, const std::shared_ptr<ov::Model>& body) : SubGraphOp(args) {
+    SubGraphOp::set_function(body);
+    for (size_t i = 0; i < body->get_parameters().size(); ++i)
+        m_input_descriptions[0].push_back(std::make_shared<InvariantInputDescription>(i, i));
+    for (size_t i = 0; i < body->get_output_size(); ++i)
+        m_output_descriptions[0].push_back(std::make_shared<BodyOutputDescription>(i, i));
+    constructor_validate_and_infer_types();
+}
+
+std::shared_ptr<Node> LoraSubgraph::clone_with_new_inputs(const OutputVector& new_args) const {
+    INTERNAL_OP_SCOPE(internal_LoraSubgraph_clone_with_new_inputs);
+    check_new_args_count(this, new_args);
+    return std::make_shared<LoraSubgraph>(new_args, get_function()->clone());
+}
+
+void LoraSubgraph::validate_and_infer_types() {
+    INTERNAL_OP_SCOPE(internal_LoraSubgraph_validate_and_infer_types);
+    OPENVINO_ASSERT(get_input_size() == 5, "LoraSubgraph must have 5 inputs whereas it has ", get_input_size());
+    OPENVINO_ASSERT(get_output_size() == 1, "LoraSubgraph must have 1 output whereas it has ", get_output_size());
+    const auto& body = get_function();
+    OPENVINO_ASSERT(body, "LoraSubgraph must have initialized body");
+    validate_and_infer_type_body(body, m_input_descriptions[0]);
+    for (size_t i = 0; i < get_output_size(); ++i)
+        set_output_type(i, body->get_output_element_type(i), body->get_output_partial_shape(i));
+}
+
+}  // namespace internal
+}  // namespace op
+}  // namespace ov

src/common/transformations/src/transformations/common_optimizations/lora_subgraph_fusion.cpp

@@ -0,0 +1,108 @@
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "transformations/common_optimizations/lora_subgraph_fusion.hpp"
+
+#include <memory>
+#include <vector>
+
+#include "itt.hpp"
+#include "openvino/op/add.hpp"
+#include "openvino/op/convolution.hpp"
+#include "openvino/op/matmul.hpp"
+#include "openvino/op/multiply.hpp"
+#include "openvino/op/parameter.hpp"
+#include "openvino/op/transpose.hpp"
+#include "openvino/op/util/read_value_base.hpp"
+#include "openvino/pass/pattern/op/optional.hpp"
+#include "openvino/pass/pattern/op/wrap_type.hpp"
+#include "ov_ops/lora_subgraph.hpp"
+#include "transformations/utils/utils.hpp"
+
+ov::pass::LoraSubgraphFusion::LoraSubgraphFusion() {
+    MATCHER_SCOPE(LoraSubgraphFusion);
+    using namespace pass::pattern;
+    auto lora_input_m = any_input();
+    auto transpose_const1_m = wrap_type<ov::op::v0::Constant>(consumers_count(1));
+    auto transpose1_m = optional<ov::op::v1::Transpose>({lora_input_m, transpose_const1_m}, consumers_count(1));
+    auto read_value1_m = wrap_type<ov::op::util::ReadValueBase>();
+    auto matmul1_m = wrap_type<ov::op::v0::MatMul>({transpose1_m, read_value1_m}, consumers_count(1));
+    auto read_value2_m = wrap_type<ov::op::util::ReadValueBase>();
+    auto multiply_m = wrap_type<ov::op::v1::Multiply>({matmul1_m, read_value2_m}, consumers_count(1));
+    auto read_value3_m = wrap_type<ov::op::util::ReadValueBase>();
+    auto matmul2_m = wrap_type<ov::op::v0::MatMul>({multiply_m, read_value3_m}, consumers_count(1));
+    auto transpose_const2_m = wrap_type<ov::op::v0::Constant>(consumers_count(1));
+    auto transpose2_m = optional<ov::op::v1::Transpose>({matmul2_m, transpose_const2_m}, consumers_count(1));
+    auto main_flow_m = wrap_type<ov::op::v0::MatMul, ov::op::v1::Convolution>({lora_input_m, any_input()});
+    auto add_m = wrap_type<ov::op::v1::Add>({transpose2_m, main_flow_m});
+
+    ov::matcher_pass_callback callback = [OV_CAPTURE_CPY_AND_THIS](Matcher& m) {
+        const auto& pattern_map = m.get_pattern_value_map();
+        const auto& lora_input = pattern_map.at(lora_input_m);
+        const auto& matmul1 = pattern_map.at(matmul1_m);
+        const auto& read_value1 = pattern_map.at(read_value1_m);
+        const auto& multiply = pattern_map.at(multiply_m);
+        const auto& read_value2 = pattern_map.at(read_value2_m);
+        const auto& matmul2 = pattern_map.at(matmul2_m);
+        const auto& read_value3 = pattern_map.at(read_value3_m);
+        const auto& main_flow = pattern_map.at(main_flow_m);
+        const auto& add = pattern_map.at(add_m);
+
+        const auto add_node = add.get_node_shared_ptr();
+        if (transformation_callback(add_node)) {
+            return false;
+        }
+
+        auto find_connected_input = [](ov::Node* child, ov::Node* parent) {
+            for (size_t i = 0; i < child->get_input_size(); ++i) {
+                auto input = child->input(i);
+                if (input.get_source_output().get_node() == parent)
+                    return input;
+            }
+            OPENVINO_THROW("Ops are not connected");
+        };
+
+        // Note: internal_inputs/external_connections order corresponds to LoraSubgraph semantic
+        const std::vector<ov::Input<ov::Node>> internal_inputs{
+            // For commutative eltwise ops, input idx may be any, so it must be computed
+            find_connected_input(add.get_node(), main_flow.get_node()),
+            pattern_map.count(transpose1_m) ? pattern_map.at(transpose1_m).get_node()->input(0)
+                                            : matmul1.get_node()->input(0),
+            matmul1.get_node()->input(1),
+            find_connected_input(multiply.get_node(), read_value2.get_node()),
+            matmul2.get_node()->input(1),
+        };
+        const ov::OutputVector external_connections{
+            main_flow,
+            lora_input,
+            read_value1,
+            read_value2,
+            read_value3,
+        };
+
+        ov::ParameterVector subgraph_parameters;
+        subgraph_parameters.reserve(internal_inputs.size());
+        for (auto& in : internal_inputs) {
+            auto new_parameter = std::make_shared<ov::op::v0::Parameter>(in.get_element_type(), in.get_partial_shape());
+            subgraph_parameters.push_back(new_parameter);
+            in.replace_source_output(new_parameter);
+        }
+        // Note: lora consumers should be taken before lora_subgraph creation,
+        // because only original consumers should be replaced with lora's output
+        const auto& lora_consumers = add.get_target_inputs();
+        const auto lora_subgraph = std::make_shared<ov::Model>(ov::OutputVector{add}, subgraph_parameters);
+        const auto lora_node = std::make_shared<ov::op::internal::LoraSubgraph>(external_connections, lora_subgraph);
+        ov::copy_runtime_info(m.get_matched_nodes(), lora_node);
+        lora_node->set_friendly_name(add_node->get_friendly_name());
+
+        for (const auto& consumer : lora_consumers)
+            consumer.replace_source_output(lora_node->output(0));
+        if (!add.get_names().empty())
+            lora_node->output(0).set_names(add.get_names());
+        return true;
+    };
+
+    auto m = std::make_shared<Matcher>(add_m, matcher_name);
+    this->register_matcher(m, callback);
+}