OV Core fix for text2image (openvinotoolkit#990)

Ticket: CVS-155261
iefode · Oct 17, 2024 · bc270f4 · bc270f4
1 parent 0f71d80
commit bc270f4
Show file tree

Hide file tree

Showing 8 changed files with 38 additions and 18 deletions.
diff --git a/src/cpp/include/openvino/genai/text2image/unet2d_condition_model.hpp b/src/cpp/include/openvino/genai/text2image/unet2d_condition_model.hpp
@@ -7,13 +7,13 @@
 #include <string>
 #include <memory>
 
-#include "openvino/genai/visibility.hpp"
-
 #include "openvino/core/any.hpp"
 #include "openvino/core/model.hpp"
 #include "openvino/runtime/tensor.hpp"
 #include "openvino/runtime/infer_request.hpp"
 #include "openvino/runtime/properties.hpp"
+
+#include "openvino/genai/visibility.hpp"
 #include "openvino/genai/lora_adapter.hpp"
 
 namespace ov {

diff --git a/src/cpp/src/lora_adapter.cpp b/src/cpp/src/lora_adapter.cpp
@@ -27,10 +27,11 @@
 #include "openvino/pass/pattern/op/wrap_type.hpp"
 #include "openvino/pass/graph_rewrite.hpp"
 #include "openvino/pass/manager.hpp"
-#include "openvino/runtime/core.hpp"
 
 #include "openvino/genai/lora_adapter.hpp"
 
+#include "utils.hpp"
+
 extern "C" {
     #include "safetensors.h"
 }
@@ -561,6 +562,7 @@ class InferRequestSignatureCache {
     }
 
     void insert (const Signature& signature, std::shared_ptr<ov::Model> model) {
+        ov::Core core = ov::genai::utils::singleton_core();
         requests[signature] = core.compile_model(model, device).create_infer_request();
     }
 
@@ -589,7 +591,6 @@ class InferRequestSignatureCache {
 
 private:
 
-    ov::Core core;
     std::unordered_map<Signature, ov::InferRequest> requests;
     std::string device;
 };

diff --git a/src/cpp/src/text2image/models/autoencoder_kl.cpp b/src/cpp/src/text2image/models/autoencoder_kl.cpp
@@ -13,8 +13,11 @@
 #include "openvino/op/multiply.hpp"
 #include "openvino/op/constant.hpp"
 
+#include "utils.hpp"
+
 #include "json_utils.hpp"
 #include "lora_helper.hpp"
+
 namespace ov {
 namespace genai {
 
@@ -34,7 +37,8 @@ AutoencoderKL::Config::Config(const std::string& config_path) {
 
 AutoencoderKL::AutoencoderKL(const std::string& root_dir)
     : m_config(root_dir + "/config.json") {
-    m_model = ov::Core().read_model(root_dir + "/openvino_model.xml");
+    ov::Core core = utils::singleton_core();
+    m_model = core.read_model(root_dir + "/openvino_model.xml");
     // apply VaeImageProcessor postprocessing steps by merging them into the VAE decoder model
     merge_vae_image_processor();
 }
@@ -69,7 +73,8 @@ AutoencoderKL& AutoencoderKL::reshape(int batch_size, int height, int width) {
 
 AutoencoderKL& AutoencoderKL::compile(const std::string& device, const ov::AnyMap& properties) {
     OPENVINO_ASSERT(m_model, "Model has been already compiled. Cannot re-compile already compiled model");
-    ov::CompiledModel compiled_model = ov::Core().compile_model(m_model, device, properties);
+    ov::Core core = utils::singleton_core();
+    ov::CompiledModel compiled_model = core.compile_model(m_model, device, properties);
     m_request = compiled_model.create_infer_request();
     // release the original model
     m_model.reset();

diff --git a/src/cpp/src/text2image/models/clip_text_model.cpp b/src/cpp/src/text2image/models/clip_text_model.cpp
@@ -5,10 +5,9 @@
 
 #include <fstream>
 
-#include "openvino/runtime/core.hpp"
-
 #include "json_utils.hpp"
 #include "lora_helper.hpp"
+#include "utils.hpp"
 
 namespace ov {
 namespace genai {
@@ -28,7 +27,8 @@ CLIPTextModel::Config::Config(const std::string& config_path) {
 CLIPTextModel::CLIPTextModel(const std::string root_dir) :
     m_clip_tokenizer(root_dir + "/../tokenizer"),
     m_config(root_dir + "/config.json") {
-    m_model = ov::Core().read_model(root_dir + "/openvino_model.xml");
+    ov::Core core = utils::singleton_core();
+    m_model = core.read_model(root_dir + "/openvino_model.xml");
 }
 
 CLIPTextModel::CLIPTextModel(const std::string& root_dir,
@@ -64,7 +64,8 @@ CLIPTextModel& CLIPTextModel::reshape(int batch_size) {
 
 CLIPTextModel& CLIPTextModel::compile(const std::string& device, const ov::AnyMap& properties) {
     OPENVINO_ASSERT(m_model, "Model has been already compiled. Cannot re-compile already compiled model");
-    ov::CompiledModel compiled_model = ov::Core().compile_model(m_model, device, properties);
+    ov::Core core = utils::singleton_core();
+    ov::CompiledModel compiled_model = core.compile_model(m_model, device, properties);
     m_request = compiled_model.create_infer_request();
     // release the original model
     m_model.reset();

diff --git a/src/cpp/src/text2image/models/clip_text_model_with_projection.cpp b/src/cpp/src/text2image/models/clip_text_model_with_projection.cpp
@@ -5,9 +5,8 @@
 
 #include <fstream>
 
-#include "openvino/runtime/core.hpp"
-
 #include "json_utils.hpp"
+#include "utils.hpp"
 
 namespace ov {
 namespace genai {
@@ -27,7 +26,8 @@ CLIPTextModelWithProjection::Config::Config(const std::string& config_path) {
 CLIPTextModelWithProjection::CLIPTextModelWithProjection(const std::string root_dir) :
     m_clip_tokenizer(root_dir + "/../tokenizer_2"),
     m_config(root_dir + "/config.json") {
-    m_model = ov::Core().read_model(root_dir + "/openvino_model.xml");
+    ov::Core core = utils::singleton_core();
+    m_model = core.read_model(root_dir + "/openvino_model.xml");
 }
 
 CLIPTextModelWithProjection::CLIPTextModelWithProjection(const std::string& root_dir,
@@ -57,7 +57,8 @@ CLIPTextModelWithProjection& CLIPTextModelWithProjection::reshape(int batch_size
 
 CLIPTextModelWithProjection& CLIPTextModelWithProjection::compile(const std::string& device, const ov::AnyMap& properties) {
     OPENVINO_ASSERT(m_model, "Model has been already compiled. Cannot re-compile already compiled model");
-    ov::CompiledModel compiled_model = ov::Core().compile_model(m_model, device, properties);
+    ov::Core core = utils::singleton_core();
+    ov::CompiledModel compiled_model = core.compile_model(m_model, device, properties);
     m_request = compiled_model.create_infer_request();
     // release the original model
     m_model.reset();

diff --git a/src/cpp/src/text2image/models/unet2d_condition_model.cpp b/src/cpp/src/text2image/models/unet2d_condition_model.cpp
@@ -5,10 +5,9 @@
 
 #include <fstream>
 
-#include "openvino/runtime/core.hpp"
-
 #include "json_utils.hpp"
 #include "lora_helper.hpp"
+#include "utils.hpp"
 
 namespace ov {
 namespace genai {
@@ -28,7 +27,8 @@ UNet2DConditionModel::Config::Config(const std::string& config_path) {
 
 UNet2DConditionModel::UNet2DConditionModel(const std::string root_dir) :
     m_config(root_dir + "/config.json") {
-    m_model = ov::Core().read_model(root_dir + "/openvino_model.xml");
+    ov::Core core = utils::singleton_core();
+    m_model = core.read_model(root_dir + "/openvino_model.xml");
     // compute VAE scale factor
     m_vae_scale_factor = std::pow(2, m_config.block_out_channels.size() - 1);
 }
@@ -86,7 +86,8 @@ UNet2DConditionModel& UNet2DConditionModel::reshape(int batch_size, int height,
 
 UNet2DConditionModel& UNet2DConditionModel::compile(const std::string& device, const ov::AnyMap& properties) {
     OPENVINO_ASSERT(m_model, "Model has been already compiled. Cannot re-compile already compiled model");
-    ov::CompiledModel compiled_model = ov::Core().compile_model(m_model, device, properties);
+    ov::Core core = utils::singleton_core();
+    ov::CompiledModel compiled_model = core.compile_model(m_model, device, properties);
     m_request = compiled_model.create_infer_request();
     // release the original model
     m_model.reset();

diff --git a/src/cpp/src/utils.cpp b/src/cpp/src/utils.cpp
@@ -259,6 +259,12 @@ void slice_matmul_statefull_model(std::shared_ptr<ov::Model> model) {
         matmul->input(0).replace_source_output(slice);
     }
 }
+
+ov::Core singleton_core() {
+    static ov::Core core;
+    return core;
+}
+
 }  // namespace utils
 }  // namespace genai
 }  // namespace ov
diff --git a/src/cpp/src/utils.hpp b/src/cpp/src/utils.hpp
@@ -4,6 +4,8 @@
 #pragma once
 
 #include "openvino/genai/llm_pipeline.hpp"
+#include "openvino/runtime/core.hpp"
+
 #include "visual_language/processor_config.hpp"
 
 namespace ov {
@@ -57,6 +59,9 @@ std::pair<ov::AnyMap, ov::AnyMap> split_core_complile_config(const ov::AnyMap& p
 ov::genai::TokenizedInputs subtract_chat_tokenized_inputs(const ov::genai::TokenizedInputs& minuend, const ov::genai::TokenizedInputs& subtrahend);
 
 void slice_matmul_statefull_model(std::shared_ptr<ov::Model> model);
+
+ov::Core singleton_core();
+
 }  // namespace utils
 }  // namespace genai
 }  // namespace ov