Merge branch 'master' into verify_chatglm3-6b

.github/workflows/causal_lm_cpp.yml

@@ -16,10 +16,10 @@ concurrency:
   cancel-in-progress: true
 
 env:
-  l_ov_link: https://storage.openvinotoolkit.org/repositories/openvino/packages/nightly/2025.0.0-17709-688f0428cfc/l_openvino_toolkit_ubuntu20_2025.0.0.dev20241224_x86_64.tgz
-  l_u22_ov_link: https://storage.openvinotoolkit.org/repositories/openvino/packages/nightly/2025.0.0-17709-688f0428cfc/l_openvino_toolkit_ubuntu22_2025.0.0.dev20241224_x86_64.tgz
-  m_ov_link: https://storage.openvinotoolkit.org/repositories/openvino/packages/nightly/2025.0.0-17709-688f0428cfc/m_openvino_toolkit_macos_12_6_2025.0.0.dev20241224_x86_64.tgz
-  w_ov_link: https://storage.openvinotoolkit.org/repositories/openvino/packages/nightly/2025.0.0-17709-688f0428cfc/w_openvino_toolkit_windows_2025.0.0.dev20241224_x86_64.zip
+  l_ov_link: https://storage.openvinotoolkit.org/repositories/openvino/packages/nightly/2025.0.0-17726-9ab2c1a18e7/l_openvino_toolkit_ubuntu20_2025.0.0.dev20241230_x86_64.tgz
+  l_u22_ov_link: https://storage.openvinotoolkit.org/repositories/openvino/packages/nightly/2025.0.0-17726-9ab2c1a18e7/l_openvino_toolkit_ubuntu22_2025.0.0.dev20241230_x86_64.tgz
+  m_ov_link: https://storage.openvinotoolkit.org/repositories/openvino/packages/nightly/2025.0.0-17726-9ab2c1a18e7/m_openvino_toolkit_macos_12_6_2025.0.0.dev20241230_x86_64.tgz
+  w_ov_link: https://storage.openvinotoolkit.org/repositories/openvino/packages/nightly/2025.0.0-17726-9ab2c1a18e7/w_openvino_toolkit_windows_2025.0.0.dev20241230_x86_64.zip
 jobs:
   cpp-multinomial-greedy_causal_lm-ubuntu:
     runs-on: ubuntu-20.04-8-cores

.github/workflows/job_vlm_sample_llava.yml

@@ -11,7 +11,7 @@ on:
         type: string
 
 env:
-  l_u22_ov_link: https://storage.openvinotoolkit.org/repositories/openvino/packages/nightly/2025.0.0-17709-688f0428cfc/l_openvino_toolkit_ubuntu22_2025.0.0.dev20241224_x86_64.tgz
+  l_u22_ov_link: https://storage.openvinotoolkit.org/repositories/openvino/packages/nightly/2025.0.0-17726-9ab2c1a18e7/l_openvino_toolkit_ubuntu22_2025.0.0.dev20241230_x86_64.tgz
 
 jobs:
   visual_language_chat_sample-ubuntu-llava:

.github/workflows/lcm_dreamshaper_cpp.yml

@@ -18,8 +18,8 @@ concurrency:
 
 env:
   PYTHON_VERSION: '3.9'
-  LINUX_OV_ARCHIVE_URL: https://storage.openvinotoolkit.org/repositories/openvino/packages/nightly/2025.0.0-17709-688f0428cfc/l_openvino_toolkit_ubuntu22_2025.0.0.dev20241224_x86_64.tgz
-  WINDOWS_OV_ARCHIVE_URL: https://storage.openvinotoolkit.org/repositories/openvino/packages/nightly/2025.0.0-17709-688f0428cfc/w_openvino_toolkit_windows_2025.0.0.dev20241224_x86_64.zip
+  LINUX_OV_ARCHIVE_URL: https://storage.openvinotoolkit.org/repositories/openvino/packages/nightly/2025.0.0-17726-9ab2c1a18e7/l_openvino_toolkit_ubuntu22_2025.0.0.dev20241230_x86_64.tgz
+  WINDOWS_OV_ARCHIVE_URL: https://storage.openvinotoolkit.org/repositories/openvino/packages/nightly/2025.0.0-17726-9ab2c1a18e7/w_openvino_toolkit_windows_2025.0.0.dev20241230_x86_64.zip
   OV_INSTALL_DIR: ${{ github.workspace }}/ov
 
 jobs:

.github/workflows/linux.yml

@@ -109,10 +109,10 @@ jobs:
           merge-multiple: true
 
       - name: CMake Build
-        run: |    
+        run: |
           source ${{ env.OV_INSTALL_DIR }}/setupvars.sh
           cmake -DCMAKE_BUILD_TYPE=${{ matrix.build-type }} -S ${{ env.SRC_DIR}} -B ${{ env.BUILD_DIR }}
-          cmake --build ${{ env.BUILD_DIR}} --config ${{ matrix.build-type }} --parallel $(nproc)
+          cmake --build ${{ env.BUILD_DIR}} --config ${{ matrix.build-type }} --parallel $(nproc) --verbose
           cmake --install ${{ env.BUILD_DIR }} --config ${{ matrix.build-type }} --prefix ${{ env.INSTALL_DIR }}
       
       - name: Pack Artifacts

.github/workflows/mac.yml

@@ -219,7 +219,7 @@ jobs:
         run: |
           source ${OV_INSTALL_DIR}/setupvars.sh
           cmake -DCMAKE_BUILD_TYPE=Release -S ./ -B ./build/
-          cmake --build ./build/ --config Release -j
+          cmake --build ./build/ --config Release --parallel --verbose
 
       - name: Test bindings
         run: |
@@ -284,7 +284,7 @@ jobs:
         run: |
           source ${OV_INSTALL_DIR}/setupvars.sh
           cmake -DCMAKE_BUILD_TYPE=Release -S ./ -B ./build/
-          cmake --build ./build/ --config Release --target py_openvino_genai -j
+          cmake --build ./build/ --config Release --target py_openvino_genai --parallel --verbose
 
       - name: Test bindings
         run: |
@@ -350,7 +350,7 @@ jobs:
         run: |
           source ${OV_INSTALL_DIR}/setupvars.sh
           cmake -DCMAKE_BUILD_TYPE=${{ matrix.build-type }} -S ./ -B ./build/
-          cmake --build ./build/ --config ${{ matrix.build-type }} --target package -j
+          cmake --build ./build/ --config ${{ matrix.build-type }} --target package --parallel --verbose
 
       - name: Build and Install dependencies
         run: |

.github/workflows/stable_diffusion_1_5_cpp.yml

@@ -122,6 +122,8 @@ jobs:
           source openvino_sd_cpp/bin/activate
           optimum-cli export openvino --model dreamlike-art/dreamlike-anime-1.0 --weight-format fp16 --task stable-diffusion models/dreamlike-art-dreamlike-anime-1.0/FP16
           wget -O ./models/soulcard.safetensors https://civitai.com/api/download/models/72591
+        env:
+          HF_HUB_ENABLE_HF_TRANSFER: 1
 
       - name: Run text2image app
         run: |
@@ -198,6 +200,8 @@ jobs:
           . "./openvino_sd_cpp/Scripts/Activate.ps1"
           optimum-cli export openvino --model dreamlike-art/dreamlike-anime-1.0 --task stable-diffusion --weight-format fp16 models/dreamlike-art-dreamlike-anime-1.0/FP16
           Invoke-WebRequest -Uri 'https://civitai.com/api/download/models/72591' -OutFile 'models/soulcard.safetensors'
+        env:
+          HF_HUB_ENABLE_HF_TRANSFER: 1
 
       - name: Run text2image app
         run: |

.github/workflows/windows.yml

@@ -230,7 +230,7 @@ jobs:
         run: |
           . "${{ env.OV_INSTALL_DIR }}/setupvars.ps1"
           cmake -DCMAKE_BUILD_TYPE=Release -S ./ -B ./build/
-          cmake --build ./build/ --config Release -j
+          cmake --build ./build/ --config Release --parallel --verbose
 
       - name: Test bindings
         run: |
@@ -295,7 +295,7 @@ jobs:
         run: |
           . "${{ env.OV_INSTALL_DIR }}/setupvars.ps1"
           cmake -DCMAKE_BUILD_TYPE=Release -S ./ -B ./build/
-          cmake --build ./build/ --config Release --target py_openvino_genai -j
+          cmake --build ./build/ --config Release --target py_openvino_genai --parallel --verbose
 
       - name: Test bindings
         run: |
@@ -360,7 +360,7 @@ jobs:
         run: |
           . "${{ env.OV_INSTALL_DIR }}/setupvars.ps1"
           cmake -DCMAKE_BUILD_TYPE=Release -S ./ -B ./build/
-          cmake --build ./build/ --config Release --target py_openvino_genai -j
+          cmake --build ./build/ --config Release --target py_openvino_genai --parallel --verbose
 
       - name: Test bindings
         run: |

CMakeLists.txt

@@ -85,7 +85,6 @@ if(MSVC AND MSVC_VERSION GREATER_EQUAL 1930 AND MSVC_VERSION LESS 1941)
     add_compile_definitions(_DISABLE_CONSTEXPR_MUTEX_CONSTRUCTOR)
 endif()
 
-
 add_subdirectory(thirdparty)
 add_subdirectory(src)
 if(EXISTS "${OpenVINOGenAI_SOURCE_DIR}/samples")

samples/export-requirements.txt

@@ -10,3 +10,4 @@ diffusers==0.32.1 # For image generation pipelines
 timm==1.0.12  # For exporting InternVL2
 torchvision  # For visual language models
 transformers>=4.43 # For Whisper
+hf_transfer # for faster models download, should used with env var HF_HUB_ENABLE_HF_TRANSFER=1

src/cpp/CMakeLists.txt

@@ -59,11 +59,13 @@ ov_genai_build_jinja2cpp()
 file(GLOB_RECURSE SOURCE_FILES "${CMAKE_CURRENT_SOURCE_DIR}/src/*.cpp" "${CMAKE_CURRENT_SOURCE_DIR}/src/*.c")
 
 set(TARGET_NAME openvino_genai)
+
 add_library(${TARGET_NAME} SHARED ${SOURCE_FILES})
+add_library(openvino::genai ALIAS ${TARGET_NAME})
+
 if(TARGET openvino_tokenizers)
     add_dependencies(${TARGET_NAME} openvino_tokenizers)
 endif()
-add_library(openvino::genai ALIAS ${TARGET_NAME})
 
 target_include_directories(${TARGET_NAME}
     PUBLIC "$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>" "$<INSTALL_INTERFACE:runtime/include>"
@@ -81,6 +83,7 @@ set_target_properties(${TARGET_NAME} PROPERTIES
     LIBRARY_OUTPUT_DIRECTORY "$<1:${CMAKE_BINARY_DIR}/openvino_genai/>"
     RUNTIME_OUTPUT_DIRECTORY "$<1:${CMAKE_BINARY_DIR}/openvino_genai/>"
 )
+
 # Extract two last digits from OpenVINOGenAI_VERSION_MAJOR because SOVERSION can only contain up to 4 symbols.
 string(REGEX MATCH [=[[0-9][0-9]$]=] MAJOR_SUFFIX ${OpenVINOGenAI_VERSION_MAJOR})
 if(DEFINED PY_BUILD_CMAKE_PACKAGE_NAME AND LINUX)

src/cpp/src/continuous_batching_adapter.hpp

@@ -33,7 +33,7 @@ class ContinuousBatchingAdapter final : public LLMPipelineImplBase {
         const std::string& device,
         const ov::AnyMap& plugin_config
     ): LLMPipelineImplBase{tokenizer, GenerationConfig()}, m_impl{
-        models_path.string(),
+        models_path,
         tokenizer,
         scheduler_config,
         device,
@@ -64,7 +64,7 @@ class ContinuousBatchingAdapter final : public LLMPipelineImplBase {
         const std::string& device,
         const ov::AnyMap& plugin_config
     ): LLMPipelineImplBase{Tokenizer(models_path), GenerationConfig()}, m_impl{
-        models_path.string(),
+        models_path,
         m_tokenizer,
         scheduler_config,
         device,

src/cpp/src/continuous_batching_impl.cpp

@@ -23,17 +23,13 @@ ContinuousBatchingPipeline::ContinuousBatchingImpl::ContinuousBatchingImpl(
     m_generation_config = generation_config;
     m_is_validation_mode_enabled = is_validation_mode_enabled;
 
-    ov::Core core;
-
-    auto [core_properties, compile_properties] = utils::split_core_compile_config(properties);
-    core.set_property(core_properties);
-
-    DeviceConfig device_config(core, scheduler_config, device, compile_properties);
+    ov::Core core = utils::singleton_core();
+    DeviceConfig device_config(core, scheduler_config, device, properties);
 
     bool is_need_per_layer_cache_control = scheduler_config.use_cache_eviction;
     utils::apply_paged_attention_transformations(model, device_config, is_need_per_layer_cache_control);
 
-    initialize_pipeline(model, scheduler_config, compile_properties, device_config, core);
+    initialize_pipeline(model, scheduler_config, properties, device_config, core);
 }
 
 void ContinuousBatchingPipeline::ContinuousBatchingImpl::_pull_awaiting_requests() {
@@ -199,7 +195,7 @@ void ContinuousBatchingPipeline::ContinuousBatchingImpl::step() {
     step_count++;
 #endif
 
-    // process generation_config.echo parameetr
+    // process generation_config.echo parameter
     _fill_prompt_log_probs(m_requests, logits);
 
     SamplerOutput sampler_output;

src/cpp/src/continuous_batching_pipeline.cpp

@@ -48,8 +48,7 @@ ContinuousBatchingPipeline::ContinuousBatchingPipeline( const std::filesystem::p
     auto draft_model_desr = extract_draft_model_from_config(properties_without_draft_model);
     auto is_prompt_lookup_enabled = extract_prompt_lookup_from_config(properties_without_draft_model);
 
-    std::filesystem::path openvino_model_name = "openvino_model.xml";
-    auto model = utils::singleton_core().read_model((models_path / openvino_model_name).string());
+    auto model = utils::singleton_core().read_model(models_path / "openvino_model.xml", {}, properties);
     auto tokenizer = ov::genai::Tokenizer(models_path, tokenizer_properties);
     auto generation_config = utils::from_config_json_if_exists(models_path);
 
@@ -74,7 +73,7 @@ ContinuousBatchingPipeline::ContinuousBatchingPipeline(
     auto draft_model_desr = extract_draft_model_from_config(properties_without_draft_model);
     auto is_prompt_lookup_enabled = extract_prompt_lookup_from_config(properties_without_draft_model);
     std::filesystem::path openvino_model_name = "openvino_model.xml";
-    auto model = utils::singleton_core().read_model((models_path / openvino_model_name).string());
+    auto model = utils::singleton_core().read_model(models_path / openvino_model_name, {}, properties_without_draft_model);
     auto generation_config = utils::from_config_json_if_exists(models_path);
 
     if (is_prompt_lookup_enabled) {

src/cpp/src/image_generation/models/autoencoder_kl.cpp

@@ -91,17 +91,15 @@ AutoencoderKL::Config::Config(const std::filesystem::path& config_path) {
 
 AutoencoderKL::AutoencoderKL(const std::filesystem::path& vae_decoder_path)
     : m_config(vae_decoder_path / "config.json") {
-    ov::Core core = utils::singleton_core();
-    m_decoder_model = core.read_model((vae_decoder_path / "openvino_model.xml").string());
+    m_decoder_model = utils::singleton_core().read_model(vae_decoder_path / "openvino_model.xml");
     // apply VaeImageProcessor postprocessing steps by merging them into the VAE decoder model
     merge_vae_image_post_processing();
 }
 
 AutoencoderKL::AutoencoderKL(const std::filesystem::path& vae_encoder_path,
                              const std::filesystem::path& vae_decoder_path)
     : AutoencoderKL(vae_decoder_path) {
-    ov::Core core = utils::singleton_core();
-    m_encoder_model = core.read_model((vae_encoder_path / "openvino_model.xml").string());
+    m_encoder_model = utils::singleton_core().read_model(vae_encoder_path / "openvino_model.xml");
 }
 
 AutoencoderKL::AutoencoderKL(const std::filesystem::path& vae_decoder_path,
@@ -131,8 +129,7 @@ AutoencoderKL::AutoencoderKL(const std::string& vae_decoder_model,
                              const Tensor& vae_decoder_weights,
                              const Config& vae_decoder_config)
     : m_config(vae_decoder_config) {
-    ov::Core core = utils::singleton_core();
-    m_decoder_model = core.read_model(vae_decoder_model, vae_decoder_weights);
+    m_decoder_model = utils::singleton_core().read_model(vae_decoder_model, vae_decoder_weights);
     // apply VaeImageProcessor postprocessing steps by merging them into the VAE decoder model
     merge_vae_image_post_processing();
 }
@@ -143,8 +140,7 @@ AutoencoderKL::AutoencoderKL(const std::string& vae_encoder_model,
                              const Tensor& vae_decoder_weights,
                              const Config& vae_decoder_config)
     : AutoencoderKL(vae_decoder_model, vae_decoder_weights, vae_decoder_config) {
-    ov::Core core = utils::singleton_core();
-    m_encoder_model = core.read_model(vae_encoder_model, vae_encoder_weights);
+    m_encoder_model = utils::singleton_core().read_model(vae_encoder_model, vae_encoder_weights);
 }
 
 AutoencoderKL::AutoencoderKL(const std::string& vae_decoder_model,

src/cpp/src/image_generation/models/clip_text_model.cpp

@@ -37,8 +37,7 @@ CLIPTextModel::Config::Config(const std::filesystem::path& config_path) {
 CLIPTextModel::CLIPTextModel(const std::filesystem::path& root_dir) :
     m_clip_tokenizer(get_tokenizer_path_by_text_encoder(root_dir)),
     m_config(root_dir / "config.json") {
-    ov::Core core = utils::singleton_core();
-    m_model = core.read_model((root_dir / "openvino_model.xml").string());
+    m_model = utils::singleton_core().read_model(root_dir / "openvino_model.xml");
 }
 
 CLIPTextModel::CLIPTextModel(const std::filesystem::path& root_dir,
@@ -53,8 +52,7 @@ CLIPTextModel::CLIPTextModel(const std::string& model,
                              const Config& config,
                              const Tokenizer& clip_tokenizer) :
     m_clip_tokenizer(clip_tokenizer), m_config(config) {
-    ov::Core core = utils::singleton_core();
-    m_model = core.read_model(model, weights);
+    m_model = utils::singleton_core().read_model(model, weights);
 }
 
 CLIPTextModel::CLIPTextModel(const std::string& model,

src/cpp/src/image_generation/models/clip_text_model_with_projection.cpp

@@ -28,8 +28,7 @@ CLIPTextModelWithProjection::Config::Config(const std::filesystem::path& config_
 CLIPTextModelWithProjection::CLIPTextModelWithProjection(const std::filesystem::path& root_dir) :
     m_clip_tokenizer(get_tokenizer_path_by_text_encoder(root_dir)),
     m_config(root_dir / "config.json") {
-    ov::Core core = utils::singleton_core();
-    m_model = core.read_model((root_dir / "openvino_model.xml").string());
+    m_model = utils::singleton_core().read_model(root_dir / "openvino_model.xml");
 }
 
 CLIPTextModelWithProjection::CLIPTextModelWithProjection(const std::filesystem::path& root_dir,
@@ -44,8 +43,7 @@ CLIPTextModelWithProjection::CLIPTextModelWithProjection(const std::string& mode
                                                          const Config& config,
                                                          const Tokenizer& clip_tokenizer) :
     m_clip_tokenizer(clip_tokenizer), m_config(config) {
-    ov::Core core = utils::singleton_core();
-    m_model = core.read_model(model, weights);
+    m_model = utils::singleton_core().read_model(model, weights);
 }
 
 CLIPTextModelWithProjection::CLIPTextModelWithProjection(const std::string& model,

src/cpp/src/image_generation/models/flux_transformer_2d_model.cpp

@@ -26,7 +26,7 @@ FluxTransformer2DModel::Config::Config(const std::filesystem::path& config_path)
 
 FluxTransformer2DModel::FluxTransformer2DModel(const std::filesystem::path& root_dir)
     : m_config(root_dir / "config.json") {
-    m_model = utils::singleton_core().read_model((root_dir / "openvino_model.xml").string());
+    m_model = utils::singleton_core().read_model(root_dir / "openvino_model.xml");
     m_vae_scale_factor = ov::genai::get_vae_scale_factor(root_dir.parent_path() / "vae_decoder" / "config.json");
 }
 
@@ -42,8 +42,7 @@ FluxTransformer2DModel::FluxTransformer2DModel(const std::string& model,
                                                const Config& config,
                                                const size_t vae_scale_factor) :
     m_config(config), m_vae_scale_factor(vae_scale_factor) {
-    ov::Core core = utils::singleton_core();
-    m_model = core.read_model(model, weights);
+    m_model = utils::singleton_core().read_model(model, weights);
 }
 
 FluxTransformer2DModel::FluxTransformer2DModel(const std::string& model,

src/cpp/src/image_generation/models/sd3_transformer_2d_model.cpp

@@ -28,7 +28,7 @@ SD3Transformer2DModel::Config::Config(const std::filesystem::path& config_path)
 
 SD3Transformer2DModel::SD3Transformer2DModel(const std::filesystem::path& root_dir)
     : m_config(root_dir / "config.json") {
-    m_model = utils::singleton_core().read_model((root_dir / "openvino_model.xml").string());
+    m_model = utils::singleton_core().read_model(root_dir / "openvino_model.xml");
     m_vae_scale_factor = get_vae_scale_factor(root_dir.parent_path() / "vae_decoder" / "config.json");
 }
 
@@ -44,8 +44,7 @@ SD3Transformer2DModel::SD3Transformer2DModel(const std::string& model,
                                              const Config& config,
                                              const size_t vae_scale_factor) :
     m_config(config), m_vae_scale_factor(vae_scale_factor) {
-    ov::Core core = utils::singleton_core();
-    m_model = core.read_model(model, weights);
+    m_model = utils::singleton_core().read_model(model, weights);
 }
 
 SD3Transformer2DModel::SD3Transformer2DModel(const std::string& model,

src/cpp/src/image_generation/models/t5_encoder_model.cpp

@@ -16,8 +16,7 @@ std::filesystem::path get_tokenizer_path_by_text_encoder(const std::filesystem::
 
 T5EncoderModel::T5EncoderModel(const std::filesystem::path& root_dir) :
     m_tokenizer(get_tokenizer_path_by_text_encoder(root_dir)) {
-    ov::Core core = utils::singleton_core();
-    m_model = core.read_model((root_dir / "openvino_model.xml").string());
+    m_model = utils::singleton_core().read_model(root_dir / "openvino_model.xml");
 }
 
 T5EncoderModel::T5EncoderModel(const std::filesystem::path& root_dir,
@@ -31,8 +30,7 @@ T5EncoderModel::T5EncoderModel(const std::string& model,
                                const Tensor& weights,
                                const Tokenizer& tokenizer) :
     m_tokenizer(tokenizer) {
-    ov::Core core = utils::singleton_core();
-    m_model = core.read_model(model, weights);
+    m_model = utils::singleton_core().read_model(model, weights);
 }
 
 T5EncoderModel::T5EncoderModel(const std::string& model,
@@ -60,9 +58,7 @@ T5EncoderModel& T5EncoderModel::reshape(int batch_size, int max_sequence_length)
 
 T5EncoderModel& T5EncoderModel::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::Core core = utils::singleton_core();
-    ov::CompiledModel compiled_model;
-    compiled_model = core.compile_model(m_model, device, properties);
+    ov::CompiledModel compiled_model = utils::singleton_core().compile_model(m_model, device, properties);
     ov::genai::utils::print_compiled_model_properties(compiled_model, "T5 encoder model");
     m_request = compiled_model.create_infer_request();
     // release the original model