2025.4.0.0

What's Changed

• Bump product version 2025.4 by @akladiev in #2620
• Fix CLEANUP_CACHE by @Wovchena in #2617
• [wwb] Add tests for mac/win to ci by @sbalandi in #2603
• xfail embed by @Wovchena in #2618
• Bump py-build-cmake from 0.4.3 to 0.5.0 by @dependabot[bot] in #2624
• Bump actions/upload-pages-artifact from 3.0.1 to 4.0.0 by @dependabot[bot] in #2625
• Bump optimum-intel[nncf] from 1.25.1 to 1.25.2 by @dependabot[bot] in #2613
• Bump optimum-intel from 1.25.1 to 1.25.2 in /tests/python_tests by @dependabot[bot] in #2614
• [GGUF] Fix Q4_1 accuracy by @wine99 in #2563
• [llm bench] Add support of arcee model by @sbalandi in #2636
• Warn about older transformers by @Wovchena in #2634
• Limited max GPU KV-cache considering max allocatable GPU memory size by @popovaan in #2633
• Bump actions/dependency-review-action from 4.7.1 to 4.7.2 by @dependabot[bot] in #2639
• [llm_bench] Override max_length to preserve max_new_tokens by @Wovchena in #2641
• Cache images by @Wovchena in #2629
• Test logging by @Wovchena in #2621
• [CMAKE] Fix samples installation by @mryzhov in #2649
• [JS] Add prettier and align eslint by @almilosz in #2631
• [WWB] Remove use_flash_attention_2 argument for phi4mm by @nikita-savelyevv in #2653
• Implement text embedding pipeline shape fix by @as-suvorov in #2449
• [CMAKE] Solve pybind targets conflict by @mryzhov in #2655
• [GHA] Disable Cacheopt tests on mac by @mryzhov in #2663
• Fix Coverity by @Wovchena in #2601
• Bump peft from 0.17.0 to 0.17.1 in /samples by @dependabot[bot] in #2658
• Bump peft from 0.17.0 to 0.17.1 in /tests/python_tests by @dependabot[bot] in #2660
• downgrade xgrammar version in master by @pavel-esir in #2668
• Extend chat template test models by @yatarkan in #2648
• [NPU]Enable chunk prefill for VLM. by @intelgaoxiong in #2657
• [JS] Add build NodeJS bindings into Manylinux 2_28 by @Retribution98 in #2537
• WWB empty_adapters mode by @likholat in #2671
• Bump actions/dependency-review-action from 4.7.2 to 4.7.3 by @dependabot[bot] in #2674
• Bump aquasecurity/trivy-action from 0.32.0 to 0.33.0 by @dependabot[bot] in #2677
• Bump langchain-core from 0.3.74 to 0.3.75 in /tests/python_tests by @dependabot[bot] in #2673
• Bump actions/download-artifact from 4.3.0 to 5.0.0 by @dependabot[bot] in #2676
• [OV JS] Add perfMetrics grammar getters & update docstrings by @almilosz in #2681
• Bump langchain-community from 0.3.27 to 0.3.29 in /tests/python_tests by @dependabot[bot] in #2680
• Bump actions/checkout from 4.2.2 to 5.0.0 by @dependabot[bot] in #2685
• Fix StructuredOutputConfig pybind11-subgen signatures generation by @pavel-esir in #2669
• [llm_bench] Add start memory info by @sbalandi in #2686
• Updating KVCrush hyperparameters by @gopikrishnajha in #2678
• [Docs] Convert whisper as stateless in the quantization example by @nikita-savelyevv in #2690
• print genai version by @wgzintel in #2684
• Fix initializer for the sparse attention mode by @vshampor in #2689
• Add docs entry about building GenAI with free threaded Python by @p-wysocki in #2679
• Reduce structured output controller mutex locking scope by @mzegla in #2687
• [speculative decoding] Move from ManualTimer to pure metrics by @sbalandi in #2695
• [CI] [GHA] Use custom actions/download-artifact action with the fixed retries logic by @akashchi in #2692
• Enable VLM generation on NPU without image input by @AlexanderKalistratov in #2694
• [llm bench] Add possibility to setup cache eviction config for LLM by @sbalandi in #2693
• Remove not supported rerank models from docs by @as-suvorov in #2702
• Tune automatic memory allocation by @popovaan in #2697
• Bump actions/setup-python from 5.6.0 to 6.0.0 by @dependabot[bot] in #2705
• Bump pytest from 8.4.1 to 8.4.2 in /tests/python_tests by @dependabot[bot] in #2710
• [WWB] friendly error message for wrong model type by @isanghao in #2672
• [CI] [GHA] Use smaller runners for image generation samples by @akashchi in #2682
• Add image generation pipeline reuse into README by @JohnLeFeng in #2701
• Align benchmark_vlm.py and cpp by @Wovchena in #2711
• [CI] Fix NodeJS tests for manylinux by @Retribution98 in #2715
• fix checking tokenizers version by @pavel-esir in #2667
• Optimize qwen2vl encoder by @WeldonWangwang in #2630
• Check available memory before allocating KV-cache. by @popovaan in #2683
• Fixed clearing of kv-cache for GPU by @popovaan in #2717
• [GHA] w/a to build ov samples by @mryzhov in #2734
• [OV JS] Initial support for SchedulerConfig by @almilosz in #2696
• Test LLM samples with GGUF models by @Retribution98 in #2464
• [llm_bench] LLMPipeline fix negative time by @sbalandi in #2742
• Bump pydantic from 2.11.7 to 2.11.9 in /samples by @dependabot[bot] in #2732
• [llm bench] Add mem info on initial/compilation phase to json/csv by @sbalandi in #2741
• Increase timeouts by @Wovchena in #2743
• Allow additional_params for tokenizer decode in TextStreamer by @dkalinowski in #2729
• Fix attention mask pass for whisper (static) by @eshiryae in #2665
• Support from-onnx parameter by @sstrehlk in #2441
• Use model path property for caching by @praasz in #2720
• Increase GGUF timeouts by @Wovchena in #2756
• [VLM] Fixed measuring of embeddings preparation. by @popovaan in #2752
• Bump timm from 1.0.19 to 1.0.20 by @dependabot[bot] in #2754
• [llm_bench] Fix OpenVINO config not being passed for speech-to-text and Whisper models by @aobolensk in #2763
• OPT & Clean code of openvino_vision_embeddings_merger_model inputs processing by @zhaixuejun1993 in #2726
• Add .github/pull_request_template.md by @Wovchena in #2765
• Update transformers to 4.53.3 by @as-suvorov in #2760
• Set xfail for gguf reader tests on windows. by @popovaan in #2766
• [llm_bench] Add reranking pipeline by @sbalandi in #2728
• remove extra cpu->gpu tensor copies in genai VLM pipeline by @liangali in #2507
• torchcodec for linux to use the latest datasets by @Wovchena in #2424
• [VLM] Add nanoLLaVA by @popovaan in #2733
• Simplify update_config_from_kwargs by @Wovchena in #2769
• Remove unused get_model_kv_cache_precision() by @Wovchena in #2750
• Update Tokenizers Submodule by @apaniukov in #2767
• Bump minja with call blocks support, remove chat template fallback for MiniCPM3-4B by @yatarkan in #2718
• Bump aquasecurity/trivy-action from 0.33.0 to 0.33.1 by @dependabot[bot] in #2706
• Bump langchain-core from 0.3.75 to 0.3.76 in /tests/python_tests by @dependabot[bot] in #2721
• hook supports transformers 4.55.0 by @wgzintel in #2776
• [Embeddings] Add last token pooling by @as-suvorov in #2757
• Remove hardcoded fields and pass special variables to apply_chat_template() by @yatarkan in #2768
• [CI] [GHA] Migrate mac workflow to mac-14 arm64 by @akashchi in #2533
• [CI] [GHA] Use older version of the Smart CI action by @akashchi in #2797
• Safe VLM JSON config parsing with read_json_param() by @yatarkan in #2785
• [NPUW] Disable chunking and F16IC for gemma3 as they are not supported currently by @AlexanderKalistratov in #2800
• Skip qwen3 embeddings test for mac by @as-suvorov in #2802
• [wwb] Add text embeddings pipeline by @sbalandi in #2787
• [llm_bench] Refactor class mapping and add qwen3 to text_embeds/rerank supported list by @sbalandi in #2782
• Support One Element Vector in Chat Scenario by @apaniukov in #2775
• Bump pydantic from 2.11.9 to 2.12.0 by @dependabot[bot] in #2806
• [WWB] Fix to support miniCPM-o. by @popovaan in #2808
• [wwb] Add text reranking pipeline by @sbalandi in #2786
• Introduce JsonContainer for chat history by @yatarkan in #2799
• [llm_bench] Add possibility to run tool with models in gguf format by @sbalandi in #2771
• Bump tar-fs from 3.0.9 to 3.1.1 in /samples/js in the npm_and_yarn group across 1 directory by @dependabot[bot] in #2774
• [llm_bench] Add minicpmo to allowded model types by @sbalandi in #2810
• Text2Image pipeline export/import by @as-suvorov in #2716
• [JS] Bump version openvino-genai-node up to 2025.4.0 by @Retribution98 in #2803
• [WWB] Fix to support nanoLLaVA. by @popovaan in #2801
• [TextRerankPipeline] Support Qwen3 reranker model by @as-suvorov in #2809
• feat: Add GGUF file support for GenAI and HF text pipelines by @sswierze in #2798
• Bump langchain-core from 0.3.76 to 0.3.78 in /tests/python_tests by @dependabot[bot] in #2794
• [VLM] Add LLaVa-NeXT-Video. by @popovaan in #2793
• Fix cache eviction header - warning as error. by @rasapala in #2784
• Add minicpm4 in supported model list by @openvino-dev-samples in #2827
• Remove redundant WWB nanoLLaVA test runs from GA workflows by @as-suvorov in #2822
• Fix output wrong model name for CSV and json by @wgzintel in #2821
• Adding non-CB pipeline for Speculative Decoding by @AsyaPronina in #2544
• Bump pydantic from 2.12.0 to 2.12.1 in /samples by @dependabot[bot] in #2830
• Add GPT OSS to Supported Models List by @apaniukov in #2833
• Bump langchain-core from 0.3.78 to 0.3.79 in /tests/python_tests by @dependabot[bot] in #2823
• Bump diffusers from 0.34.0 to 0.35.2 in /samples by @dependabot[bot] in #2842
• add_request() to support token_type_ids with prompt by @zhaohb in #2738
• Bump pydantic from 2.12.1 to 2.12.2 by @dependabot[bot] in #2837
• Split cache eviction tests by @as-suvorov in #2844
• [OV JS] Fix TextEmbeddingPipeline config param by @almilosz in #2832
• Fix missing result_dir parameter in evaluator constructors by @sswierze in #2825
• Bump diffusers from 0.34.0 to 0.35.2 in /tests/python_tests by @dependabot[bot] in #2836
• Coverity issues fix by @sgonorov in #2834
• Bump pillow from 11.3.0 to 12.0.0 in /samples by @dependabot[bot] in #2848
• Bump actions/dependency-review-action from 4.7.3 to 4.8.1 by @dependabot[bot] in #2824
• Sparse attention documentation by @vshampor in #2698
• Increase cacheopt timeout by @Wovchena in #2852
• Bump actions/cache from 4.2.4 to 4.3.0 by @dependabot[bot] in #2770
• Bump langchain-community from 0.3.29 to 0.3.31 in /tests/python_tests by @dependabot[bot] in #2805
• Bump timm from 1.0.19 to 1.0.20 in /tests/python_tests by @dependabot[bot] in #2753
• [StructuredOutput] Update XGrammar by @apaniukov in #2817
• Expose get_original_chat_template method in Tokenizer by @mzegla in #2722
• [DOCS] Fix ENABLE_PYTHON cmake option by @as-suvorov in #2849
• Encourage requirements txt usage by @Wovchena in #2851
• Fix MiniCPM-o-2_6 by @Wovchena in #2847
• Enable model caching for Whisper pipeline on GPU and NPU by @luke-lin-vmc in #2759
• Increase wwb timeout by @Wovchena in #2854
• Show how to run from PR by @Wovchena in #2261
• Introduce chat history class by @yatarkan in #2816
• [CI] [GHA] Remove redundant args from the coverity command by @akashchi in #2761
• Reduce logging for windows by @Wovchena in #2853
• Bump actions/setup-node from 4.4.0 to 6.0.0 by @dependabot[bot] in #2856
• [llm_bench] Model type resolution fix by @sgonorov in #2877
• Fix whisper hook generate decorator arguments passing by @as-suvorov in #2873
• [llm_bench] Fix text use_case in case of complex_model_types by @sbalandi in #2874
• [JS API] Add text_generation/benchmark_genai.js sample by @almilosz in #2826
• Fixture-based VLM models reusing by @sgonorov in #2719
• Update Tokenizers Submodule by @apaniukov in #2872
• Support ChatHistory in .generate() for LLMs by @yatarkan in #2871
• Print the scheduler_config info by @wgzintel in #2467
• [llm_bench] Fix gpt-oss LLM bench support by @sgonorov in #2889
• C API: implemented VlmPipeline by @zhaohb in #2735
• [JS] Implement StructuredOutputConfig by @Retribution98 in #2876
• Fix unclassified issues by @sgonorov in #2881
• [llm_bench] Fix help message by @isanghao in #2887
• Limit torch version by @sgonorov in #2858
• Relax DecodedResults.perfMetrics test by @Wovchena in #2900
• [wwb] Fix reranker tests by @sbalandi in #2890
• Bump pydantic from 2.12.2 to 2.12.3 in /samples by @dependabot[bot] in #2868
• [JS] Disable the randomly failed tests by @Retribution98 in #2910
• Enable WhisperStatefulImpl for NPU, fix Whisper pipelines for transformers 4.53.3 & 4.55 by @eshiryae in #2126
• Extend using of full chat history mode for stateful pipeline for VLM and LLM with encoded inputs by @AlexanderKalistratov in #2835
• Upgrade optimum-intel and transformers by @Wovchena in #2611
• [wwb] Similarity for reranking models become based on relevance scores by @sbalandi in #2883
• Bump langchain-core from 0.3.79 to 1.0.0 in /tests/python_tests by @dependabot[bot] in #2867
• Bump langchain-community from 0.3.31 to 0.4 in /tests/python_tests by @dependabot[bot] in #2866
• Add parsing by @pavel-esir in #2772
• Bump timm from 1.0.20 to 1.0.21 by @dependabot[bot] in #2915
• WWB Text Generation with LoRA by @likholat in #2723
• [RAG] Add qwen3 models to docs by @as-suvorov in #2920
• Restore 4 missing VLM test cases by @sgonorov in #2914
• Bump actions/upload-artifact from 4.6.2 to 5.0.0 by @dependabot[bot] in #2917
• Bump timm from 1.0.20 to 1.0.21 in /tests/python_tests by @dependabot[bot] in #2918
• update gpu block size based on xattn by @rnwang04 in #2764
• [JS] Create ChatHistory bindings by @Retribution98 in #2922
• Update C++ and Python chat samples with using ChatHistory class by @yatarkan in #2931
• Update llm bench by @rnwang04 in #2924
• configure_sparse_attention for wwb model loader by @rnwang04 in #2927
• Align whisper log with other GenAI pipelines by @eshiryae in #2939
• add gil_scoped_acquire to fix segfault for streaming parser by @pavel-esir in #2938
• [llm_bench] Limit the MAX Input Text In log by @peterchen-intel in #2911
• Revert torch freeze by @sgonorov in #2935
• Fix spelling errors & Grammar issues by @as-suvorov in #2882
• [GHA] Add missing jobs to workflow overal status by @as-suvorov in #2893
• Drop Python 3.9 by @Wovchena in #2948
• Add gemma-3-270m to supported models by @yatarkan in #2949
• Clarify llm_bench help by @Wovchena in #2944
• Fix for another batch of coverity issues by @sgonorov in #2897
• [RAG] Add text rerank models task by @as-suvorov in #2941
• [Rag] Add new embedding parameters by @as-suvorov in #2932
• [llm_bench] Add support smolvlm for optimum only by @sbalandi in #2957
• Add threaded callback for image generation pipeline by @JohnLeFeng in #2864
• Fix KVCrushAnchorPointMode::ALTERNATE conflicting with Windows headers by @yatarkan in #2958
• Fixed LLM bench readme to use CPU torch by @sgonorov in #2936
• Modify sample to demonstrate parsing by @pavel-esir in #2940
• Readme improvements by @MaximProshin in #2961
• [VLM] Use position_ids from inputs embedder in CB. by @popovaan in #2703
• Update README for threaded callback of image generation pipeline by @JohnLeFeng in #2962
• Support QWen VL video inputs (#2514) by @peterchen-intel in #2985
• [Port] Update condition for enum value conflicting with windows headers by @yatarkan in #2997
• Release fix for callback handling by @sgonorov in #3056

New Contributors

• @wine99 made their first contribution in #2563
• @JohnLeFeng made their first contribution in #2701
• @aobolensk made their first contribution in #2763
• @zhaixuejun1993 made their first contribution in #2726
• @liangali made their first contribution in #2507
• @rasapala made their first contribution in #2784
• @openvino-dev-samples made their first contribution in #2827
• @rnwang04 made their first contribution in #2764
• @MaximProshin made their first contribution in #2961

Full Changelog: 2025.3.0.0...2025.4.0.0