Merge branch 'master' into ak/wwb_llamacpp

Author: AlexKoff88

.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

@@ -1,4 +1,4 @@
-name: macOS (12, Python 3.9)
+name: macOS (12, Python 3.10)
 on:
   workflow_dispatch:
   pull_request:
@@ -16,7 +16,7 @@ concurrency:
   cancel-in-progress: true
 
 env:
-  PYTHON_VERSION: '3.9'
+  PYTHON_VERSION: '3.10'
   OV_BRANCH: master
   OV_TARBALL: ''
 
@@ -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/deployment-requirements.txt

@@ -2,4 +2,4 @@
 --extra-index-url https://storage.openvinotoolkit.org/simple/wheels/nightly
 openvino_genai~=2025.0.0.0.dev
 librosa==0.10.2.post1  # For Whisper
-pillow==11.0.0  # Image processing for VLMs
+pillow==11.1.0  # Image processing for VLMs

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 +23,14 @@ 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);
+    utils::apply_gather_before_matmul_transformation(model);
 
-    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 +196,7 @@ 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;
@@ -448,7 +445,7 @@ void ContinuousBatchingPipeline::ContinuousBatchingImpl::_fill_prompt_log_probs(
     const float * logits_data = logits.data<float>();
     ov::Shape logits_shape = logits.get_shape();
     OPENVINO_ASSERT(logits_shape.size() == 3);
-    size_t batch_seq_len = logits_shape[1], vocab_size = logits_shape[2];
+    size_t vocab_size = logits_shape[2];
     for (size_t sequence_group_id = 0, currently_processed_tokens = 0; sequence_group_id < sequence_groups.size(); ++sequence_group_id) {
         SequenceGroup::Ptr sequence_group = sequence_groups[sequence_group_id];
         // requests not scheduled, in decoding phase or not echoing are not processed
@@ -458,26 +455,25 @@ void ContinuousBatchingPipeline::ContinuousBatchingImpl::_fill_prompt_log_probs(
 
         size_t num_running_sequences = sequence_group->num_running_seqs();
         OPENVINO_ASSERT(num_running_sequences == 1);
-        size_t actual_seq_len = sequence_group->get_num_scheduled_tokens();
-        size_t padded_amount_of_processed_tokens = std::max(actual_seq_len, batch_seq_len);
+        size_t output_seq_len = sequence_group->get_output_seq_len();
 
         const float * sequence_group_logits_data = logits_data + vocab_size * currently_processed_tokens;
 
         size_t num_prompt_tokens_processed = sequence_group->get_num_processed_tokens();
-        OPENVINO_ASSERT(num_prompt_tokens_processed + actual_seq_len <= sequence_group->get_prompt_len());
+        OPENVINO_ASSERT(num_prompt_tokens_processed + output_seq_len <= sequence_group->get_prompt_len());
 
         // if we processed the whole prompt we don't include last logprob as it will be processed by the sampler (it's already completion)
         // otherwise we include it as it will be used in the next part of the prompt
         int exclude_last_logprob = 1;
-        if (num_prompt_tokens_processed + actual_seq_len < sequence_group->get_prompt_len())
+        if (num_prompt_tokens_processed + output_seq_len < sequence_group->get_prompt_len())
             exclude_last_logprob = 0;
 
         // if we start processing the prompt we add "fake" log prob for the first position (begin of sequence)
         if (num_prompt_tokens_processed == 0)
             sequence_group->append_prompt_log_prob(1.0);
 
         for (int token_logits_offset = 0, token_id_offset = num_prompt_tokens_processed + 1;
-             token_logits_offset < actual_seq_len - exclude_last_logprob;
+             token_logits_offset < output_seq_len - exclude_last_logprob;
              token_logits_offset++, token_id_offset++) {
 
             const float* token_logits = (sequence_group_logits_data + token_logits_offset * vocab_size);
@@ -502,7 +498,7 @@ void ContinuousBatchingPipeline::ContinuousBatchingImpl::_fill_prompt_log_probs(
 
             sequence_group->append_prompt_log_prob(token_logit - max_value - log_sum);
         }
-        currently_processed_tokens += padded_amount_of_processed_tokens * num_running_sequences;
+        currently_processed_tokens += output_seq_len * num_running_sequences;
         // For max_new_tokens == 0, we don't reach sampling so need to notify handle separately
         if(sequence_group->get_sampling_parameters().max_new_tokens == 0) {
             sequence_group->notify_handle_echo_only();

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/generation_config.cpp

@@ -230,9 +230,9 @@ void GenerationConfig::validate() const {
         OPENVINO_ASSERT(temperature > 0, "When 'do_sample' is true, temperature must be a strictly positive float, but got ", temperature);
     } else {
         // parameters requiring multinomial
-        OPENVINO_ASSERT(top_k == std::numeric_limits<size_t>::max(), "When 'do_sample' is false, top_k must be max of size_t, but got ", top_k);
-        OPENVINO_ASSERT(top_p == 1.0f, "When 'do_sample' is false, top_p must be 1.0f, but got ", top_p);
-        OPENVINO_ASSERT(temperature == 1.0f, "When 'do_sample' is false, temperature must be a 1.0f, but got ", temperature);
+        // OPENVINO_ASSERT(top_k == std::numeric_limits<size_t>::max(), "When 'do_sample' is false, top_k must be max of size_t, but got ", top_k);
+        // OPENVINO_ASSERT(top_p == 1.0f, "When 'do_sample' is false, top_p must be 1.0f, but got ", top_p);
+        // OPENVINO_ASSERT(temperature == 1.0f, "When 'do_sample' is false, temperature must be a 1.0f, but got ", temperature);
     }
 
     if (is_beam_search()) {
@@ -252,10 +252,10 @@ void GenerationConfig::validate() const {
         }
     } else {
         // parameters requiring beam search
-        OPENVINO_ASSERT(num_beam_groups == 1, "'num_beam_groups' is supported by beam search only and should be 1 otherwise, but got ", num_beam_groups);
-        OPENVINO_ASSERT(no_repeat_ngram_size == std::numeric_limits<size_t>::max(), "'no_repeat_ngram_size' is supported only by beam search, otherwise should be set to max of size_t, but got ", no_repeat_ngram_size);
-        OPENVINO_ASSERT(diversity_penalty == 0.0f, "'diversity_penalty' is set to ", diversity_penalty, " (default is 0.0f), which is supported only by beam search sampling");
-        OPENVINO_ASSERT(length_penalty == 1.0f, "'length_penalty' is set to ", length_penalty, " (default is 1.0f), which is supported only by beam search sampling");
+        // OPENVINO_ASSERT(num_beam_groups == 1, "'num_beam_groups' is supported by beam search only and should be 1 otherwise, but got ", num_beam_groups);
+        // OPENVINO_ASSERT(no_repeat_ngram_size == std::numeric_limits<size_t>::max(), "'no_repeat_ngram_size' is supported only by beam search, otherwise should be set to max of size_t, but got ", no_repeat_ngram_size);
+        // OPENVINO_ASSERT(diversity_penalty == 0.0f, "'diversity_penalty' is set to ", diversity_penalty, " (default is 0.0f), which is supported only by beam search sampling");
+        // OPENVINO_ASSERT(length_penalty == 1.0f, "'length_penalty' is set to ", length_penalty, " (default is 1.0f), which is supported only by beam search sampling");
     }
 
     // assistant generation