RPE operation and fusing transformation (openvinotoolkit#20734)

* RPE operation and fusing transformation * Correct includes * Apply suggestions from code review Co-authored-by: Pawel Raasz <[email protected]> * Comments adressed * Misprints * Update src/common/transformations/src/transformations/common_optimizations/fuse_rotary_positional_embeddings.cpp Co-authored-by: Pawel Raasz <[email protected]> * Ivan comments adressed * Update src/common/transformations/src/transformations/common_optimizations/fuse_rotary_positional_embeddings.cpp Co-authored-by: Pawel Raasz <[email protected]> * Fix includes and adds comments --------- Co-authored-by: Pavel Durandin <[email protected]> Co-authored-by: Pawel Raasz <[email protected]>
akuporos · Dec 8, 2023 · b607fb7 · b607fb7
1 parent 5c5b393
commit b607fb7
Show file tree

Hide file tree

Showing 6 changed files with 307 additions and 0 deletions.
diff --git a/src/common/transformations/include/ov_ops/rotary_positional_embeddings.hpp b/src/common/transformations/include/ov_ops/rotary_positional_embeddings.hpp
@@ -0,0 +1,38 @@
+// Copyright (C) 2018-2023 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include "openvino/op/op.hpp"
+#include "transformations_visibility.hpp"
+
+namespace ov {
+namespace op {
+namespace internal {
+
+///
+/// \brief Rotary Positional Embeddings operation
+/// Internal operation which may change in the future
+/// \ingroup ov_ops_cpp_api
+class TRANSFORMATIONS_API RPE : public ov::op::Op {
+public:
+    OPENVINO_OP("RPE", "ie_internal_opset", op::Op);
+
+    RPE() = default;
+    RPE(const Output<Node>& data, const Output<Node>& sin, const Output<Node>& cos, int64_t axis);
+
+    void set_axis(int64_t axis);
+    int64_t get_axis() const;
+
+    bool visit_attributes(AttributeVisitor& visitor) override;
+    void validate_and_infer_types() override;
+    std::shared_ptr<Node> clone_with_new_inputs(const OutputVector& new_args) const override;
+
+private:
+    int64_t m_axis{};
+};
+
+}  // namespace internal
+}  // namespace op
+}  // namespace ov
diff --git a/...ations/include/transformations/common_optimizations/fuse_rotary_positional_embeddings.hpp b/...ations/include/transformations/common_optimizations/fuse_rotary_positional_embeddings.hpp
@@ -0,0 +1,24 @@
+// Copyright (C) 2018-2023 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include "openvino/pass/graph_rewrite.hpp"
+#include "transformations_visibility.hpp"
+
+namespace ov {
+namespace pass {
+class TRANSFORMATIONS_API RPE_Fusion;
+}  // namespace pass
+}  // namespace ov
+
+/**
+ * @ingroup ie_transformation_common_api
+ * @brief Fuses special sub-graph into an internal Rotary Positional Embedding operation
+ */
+class ov::pass::RPE_Fusion : public ov::pass::MatcherPass {
+public:
+    OPENVINO_RTTI("RPE_Fusion", "0");
+    RPE_Fusion();
+};
diff --git a/src/common/transformations/src/ov_ops/rotary_positional_embeddings.cpp b/src/common/transformations/src/ov_ops/rotary_positional_embeddings.cpp
@@ -0,0 +1,46 @@
+// Copyright (C) 2018-2023 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "ov_ops/rotary_positional_embeddings.hpp"
+
+#include "itt.hpp"
+
+namespace ov {
+namespace op {
+namespace internal {
+
+RPE::RPE(const Output<Node>& data, const Output<Node>& sin, const Output<Node>& cos, int64_t axis)
+    : Op({data, sin, cos}),
+      m_axis{axis} {
+    constructor_validate_and_infer_types();
+}
+
+void RPE::set_axis(int64_t axis) {
+    m_axis = axis;
+}
+
+int64_t RPE::get_axis() const {
+    return m_axis;
+}
+
+void RPE::validate_and_infer_types() {
+    INTERNAL_OP_SCOPE(internal_RoPE_validate_and_infer_types);
+    set_output_type(0, get_input_element_type(0), get_input_partial_shape(0));
+}
+
+bool RPE::visit_attributes(ov::AttributeVisitor& visitor) {
+    INTERNAL_OP_SCOPE(internal_RoPE_visit_attributes);
+    visitor.on_attribute("axis", m_axis);
+    return true;
+}
+
+std::shared_ptr<ov::Node> RPE::clone_with_new_inputs(const ov::OutputVector& new_args) const {
+    INTERNAL_OP_SCOPE(internal_RoPE_clone_with_new_inputs);
+    check_new_args_count(this, new_args);
+    return std::make_shared<RPE>(new_args.at(0), new_args.at(1), new_args.at(2), m_axis);
+}
+
+}  // namespace internal
+}  // namespace op
+}  // namespace ov
diff --git a/...formations/src/transformations/common_optimizations/fuse_rotary_positional_embeddings.cpp b/...formations/src/transformations/common_optimizations/fuse_rotary_positional_embeddings.cpp
@@ -0,0 +1,98 @@
+// Copyright (C) 2018-2023 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "transformations/common_optimizations/fuse_rotary_positional_embeddings.hpp"
+
+#include "itt.hpp"
+#include "openvino/core/validation_util.hpp"
+#include "openvino/op/add.hpp"
+#include "openvino/op/concat.hpp"
+#include "openvino/op/multiply.hpp"
+#include "openvino/op/op.hpp"
+#include "openvino/op/variadic_split.hpp"
+#include "openvino/pass/pattern/op/wrap_type.hpp"
+#include "ov_ops/rotary_positional_embeddings.hpp"
+#include "transformations/utils/utils.hpp"
+#include "validation_util.hpp"
+
+using ov::op::v0::Concat;
+using ov::op::v1::Add;
+using ov::op::v1::Multiply;
+using ov::op::v1::VariadicSplit;
+
+ov::pass::RPE_Fusion::RPE_Fusion() {
+    MATCHER_SCOPE(RPE_Fusion);
+
+    auto sin = pattern::any_input();
+    auto cos = pattern::any_input();
+
+    // FIXME: should be a single node match
+    auto source_1 = pattern::any_input();
+    auto source = pattern::any_input();
+    // BEGIN: rotate_half
+
+    // Variadic Split into two equal parts
+    auto axis = pattern::any_input();
+    auto split_length = INT_CONSTANT_WITH_PREDICATE(value.size() == 2 && value[0] == value[1]);
+    auto vsplit = pattern::wrap_type<VariadicSplit>({source, axis, split_length});
+    vsplit->set_output_size(2);
+
+    // Negate
+    auto minus_1 = FLOAT_CONSTANT_WITH_PREDICATE(value.size() == 1 && value[0] == -1);
+    auto neg = pattern::wrap_type<Multiply>({vsplit->output(1), minus_1});
+
+    // Concat two splitted parts in the opposite order, first of them is negated
+    auto concat = pattern::wrap_type<Concat>({neg, vsplit->output(0)});  // make sure axis eq to vsplit eq -1
+
+    // END: rotate half
+
+    auto mul_sin = pattern::wrap_type<Multiply>({concat, sin});
+    auto mul_cos = pattern::wrap_type<Multiply>({source_1, cos});
+    auto add = pattern::wrap_type<Add>({mul_cos, mul_sin});
+
+    ov::matcher_pass_callback matcher_pass_callback = [=](pattern::Matcher& m) {
+        auto value_map = m.get_pattern_value_map();
+
+        auto actual_source = value_map.at(vsplit).get_node_shared_ptr()->input_value(0);
+        auto potential_source = value_map.at(mul_cos).get_node_shared_ptr()->input_value(0);
+        auto cos_output = value_map.at(mul_cos).get_node_shared_ptr()->input_value(1);
+
+        if (actual_source != potential_source && actual_source != cos_output)
+            return false;  // flawed match
+        if (actual_source == potential_source && actual_source == cos_output)
+            return false;  // flawed match
+        if (actual_source != potential_source && actual_source == cos_output)
+            cos_output = potential_source;
+
+        auto input = value_map.at(source);
+        auto concat_node = ov::as_type_ptr<Concat>(value_map.at(concat).get_node_shared_ptr());
+        if (!concat_node)
+            return false;
+        auto split_axis_node = ov::util::get_constant_from_source(value_map.at(axis));
+        if (!split_axis_node)
+            return false;
+        auto value = split_axis_node->cast_vector<int64_t>();
+        if (value.size() != 1)
+            return false;
+        auto concat_axis = concat_node->get_concatenation_axis();
+        auto split_axis = value[0];
+        if (concat_axis != split_axis) {
+            if (input.get_partial_shape().rank().is_static()) {
+                auto rank = input.get_partial_shape().rank().get_length();
+                concat_axis = ov::util::normalize(concat_axis, rank);
+                split_axis = ov::util::normalize(split_axis, rank);
+            }
+            if (concat_axis != split_axis)
+                return false;
+        }
+        auto rpe =
+            std::make_shared<ov::op::internal::RPE>(input, value_map.at(sin), cos_output, concat_node->get_axis());
+
+        for (const auto& label : {vsplit, neg, concat, mul_sin, mul_cos, add})
+            ov::copy_runtime_info(value_map.at(label).get_node_shared_ptr(), rpe);
+        return ov::replace_output_update_name(value_map.at(add), rpe->output(0));
+    };
+    auto m = std::make_shared<pattern::Matcher>(add, matcher_name);
+    register_matcher(m, matcher_pass_callback);
+}
diff --git a/src/common/transformations/tests/common_optimizations/fuse_rotary_positional_embeddings.cpp b/src/common/transformations/tests/common_optimizations/fuse_rotary_positional_embeddings.cpp
@@ -0,0 +1,97 @@
+// Copyright (C) 2018-2023 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "transformations/common_optimizations/fuse_rotary_positional_embeddings.hpp"
+
+#include <gtest/gtest.h>
+
+#include "common_test_utils/ov_test_utils.hpp"
+#include "openvino/op/parameter.hpp"
+#include "openvino/op/variadic_split.hpp"
+#include "ov_ops/rotary_positional_embeddings.hpp"
+#include "transformations/utils/utils.hpp"
+
+using namespace std;
+using namespace ov;
+using namespace ov::op;
+
+void name_node_and_output(const shared_ptr<Node>& op, const std::string& name) {
+    op->set_friendly_name(name);
+    op->output(0).set_names({name});
+}
+
+TEST_F(TransformationTestsF, FuseRPE) {
+    {
+        auto data = make_shared<v0::Parameter>(element::f32, PartialShape::dynamic());
+        name_node_and_output(data, "source");
+        auto sin = make_shared<v0::Parameter>(element::f32, PartialShape::dynamic());
+        name_node_and_output(sin, "sin");
+        auto cos = make_shared<v0::Parameter>(element::f32, PartialShape::dynamic());
+        name_node_and_output(cos, "cos");
+        auto axis = v0::Constant::create(element::i64, {}, {-1});
+        auto split_lengths = v0::Constant::create(element::i64, {2}, {10, 10});
+        auto split = make_shared<v1::VariadicSplit>(data, axis, split_lengths);
+
+        auto minus_one = v0::Constant::create(element::f32, {}, {-1});
+        auto negate = make_shared<v1::Multiply>(split->output(1), minus_one);
+
+        auto concat = make_shared<v0::Concat>(OutputVector{negate, split->output(0)}, -1);
+
+        auto mul_sin = make_shared<op::v1::Multiply>(concat, sin);
+        auto mul_cos = make_shared<op::v1::Multiply>(data, cos);
+        auto add = make_shared<op::v1::Add>(mul_cos, mul_sin);
+        name_node_and_output(add, "rpe");
+
+        model = std::make_shared<Model>(NodeVector{add}, ParameterVector{data, sin, cos});
+
+        manager.register_pass<ov::pass::RPE_Fusion>();
+    }
+    {
+        auto data = make_shared<v0::Parameter>(element::f32, PartialShape::dynamic());
+        name_node_and_output(data, "source");
+        auto sin = make_shared<v0::Parameter>(element::f32, PartialShape::dynamic());
+        name_node_and_output(sin, "sin");
+        auto cos = make_shared<v0::Parameter>(element::f32, PartialShape::dynamic());
+        name_node_and_output(cos, "cos");
+        auto rpe = make_shared<ov::op::internal::RPE>(data, sin, cos, -1);
+        name_node_and_output(rpe, "rpe");
+        model_ref = std::make_shared<Model>(NodeVector{rpe}, ParameterVector{data, sin, cos});
+    }
+    comparator.enable(FunctionsComparator::CmpValues::NAMES);
+}
+
+TEST_F(TransformationTestsF, FuseRPESorcesAreMultiOutputed) {
+    /* Transformation matcher searches for a single source as a beginning of the pattern:
+                 VariadicSplit ...
+     source ____/
+                \
+                 Multiply ...
+    This test is designed to check that in case we feed VariadicSplit and Multiply from different outputs of the same
+    node, the transformation won't happen since the source isn't the same
+     */
+    {
+        auto data_ = make_shared<v0::Parameter>(element::f32, PartialShape::dynamic());
+        auto sin = make_shared<v0::Parameter>(element::f32, PartialShape::dynamic());
+        auto cos = make_shared<v0::Parameter>(element::f32, PartialShape::dynamic());
+
+        auto data = make_shared<v1::Split>(data_, v0::Constant::create(element::i64, {}, {-1}), 2);
+
+        auto axis = v0::Constant::create(element::i64, {}, {-1});
+        auto split_lengths = v0::Constant::create(element::i64, {2}, {10, 10});
+        auto split = make_shared<v1::VariadicSplit>(data->output(0), axis, split_lengths);
+
+        auto minus_one = v0::Constant::create(element::f32, {}, {-1});
+        auto negate = make_shared<v1::Multiply>(split->output(1), minus_one);
+
+        auto concat = make_shared<v0::Concat>(OutputVector{negate, split->output(0)}, -1);
+
+        auto mul_sin = make_shared<op::v1::Multiply>(concat, sin);
+        auto mul_cos = make_shared<op::v1::Multiply>(data->output(1), cos);
+        auto add = make_shared<op::v1::Add>(mul_cos, mul_sin);
+
+        model = std::make_shared<Model>(NodeVector{add}, ParameterVector{data_, sin, cos});
+
+        manager.register_pass<ov::pass::RPE_Fusion>();
+    }
+}
diff --git a/src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp b/src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp
@@ -30,6 +30,7 @@
 #include "transformations/common_optimizations/common_optimizations.hpp"
 #include "transformations/common_optimizations/wrap_interpolate_into_transposes.hpp"
 #include "transformations/common_optimizations/matmul_const_transposes_extraction.hpp"
+#include "transformations/common_optimizations/fuse_rotary_positional_embeddings.hpp"
 #include "transformations/control_flow/unroll_tensor_iterator.hpp"
 #include "transformations/fp16_compression/mark_decompression_convert_constant_folding.hpp"
 #include "transformations/op_conversions/convert_batch_to_space.hpp"
@@ -326,6 +327,7 @@ void Transformations::PreLpt(const std::vector<ov::element::Type>& defaultPrecis
 
     CPU_REGISTER_PASS_COMMON(manager, ov::pass::AUGRUCellFusion);
     CPU_REGISTER_PASS_COMMON(manager, ov::pass::CommonOptimizations);
+    CPU_REGISTER_PASS_COMMON(manager, ov::pass::RPE_Fusion);
     CPU_REGISTER_PASS_COMMON(manager, ov::pass::WrapInterpolateIntoTransposes);
     CPU_REGISTER_PASS_COMMON(manager, ov::pass::TransposeSinking);
     CPU_REGISTER_PASS_COMMON(manager, ov::pass::ConvertSequenceToTensorIterator);
@@ -473,6 +475,8 @@ void Transformations::PreLpt(const std::vector<ov::element::Type>& defaultPrecis
     CPU_DISABLE_PASS_COMMON(manager, ov::pass::ConvertTopK11ToTopK3);
     CPU_DISABLE_PASS_COMMON(manager, ov::pass::HSwishDecomposition);
     CPU_DISABLE_PASS_COMMON(manager, ov::pass::MatMulConstTransposesExtraction);
+    // CVS-126827: should be disabled until CPU supports this internal op
+    CPU_DISABLE_PASS_COMMON(manager, ov::pass::RPE_Fusion);
     CPU_DISABLE_PASS_X64(manager, ov::pass::HSigmoidDecomposition);
 
     CPU_DISABLE_PASS_X64(manager, ov::pass::ReduceL1Decomposition);