huggingface/optimum-intel v1.27.0

v1.27.0: Transformers 4.56/4.57, Zamba2, BitNet, Granite-4, Exaone-4, LFM2

on GitHub

️🏗 ​New architectures support

• Add OpenVINO support for Zamba2 by @rkazants in #1354
• Add OpenVINO support for BitNet by @rkazants in #1518
• Add OpenVINO support for LFM2 by @popovaan in #1515
• Add OpenVINO support for EXAONE 4.0 by @zhaohb in #1491
• Add OpenVINO support for Granite-4.0 family by @rkazants in #1514

🧹 Deprecations

• Removed nf4_fp8 quantization modes by @ljaljushkin in #1493
• Add depreciation warnings for INC and IPEX by @echarlaix in #1568

🔧 Enhancements & Fixes

• [OpenVINO] Transformers 4.56/4.57 support by @IlyasMoutawwakil in #1541
• [OpenVINO] Fix OpenVINO model inference not being affected by static quantization by @nikita-savelyevv in #1461
• [IPEX] Fix IPEX models for transformers v4.55 by @kaixuanliu in #1485
• [OpenVINO]Add default config for Qwen3-30B-A3B by @ljaljushkin in #1506
• [OpenVINO] Fix preprocess_inputs method for Gemma3 by @rkazants in #1507
• [IPEX] Fix IPEX models can_compile method by @jiqing-feng in #1511
• [OpenVINO] Fix TasksManager._TRANSFORMERS_TASKS_TO_MODEL_LOADERS by @echarlaix in #1501
• [OpenVINO] Add default int4 config for inceptionai/jais-13b by @nikita-savelyevv in #1519
• [OpenVINO] Add cache_position input inside prepare_inputs method for Mamba by @nikita-savelyevv in #1517
• [OpenVINO] Refactor from_pretrained quantization by @nikita-savelyevv in #1520
• [OpenVINO] fix bug for attention_mask when model is not patched by @kaixuanliu in #1526
• [OpenVINO] Add custom int4 config for SmolVLM2-256M-Video by @nikita-savelyevv in #1532
• [OpenVINO] Fix whisper inference for models exported without pkv by @echarlaix in #1534
• [OpenVINO] Optimize IR for Mamba models by @rkazants in #1538
• [OpenVINO] NNCF 2.19 update by @nikita-savelyevv in #1522
• [OpenVINO] Update optimum-intel to OV 2025.4 release by @rkazants in #1544
• [OpenVINO] Update InferRequestWrapper to collect samples depending on stateful models state by @nikita-savelyevv in #1505
• [OpenVINO] Add gsm8k as a dataset option for CausalLM quantization by @nikita-savelyevv in #1547
• [OpenVINO] Take into account that pillow may be not installed by @nikita-savelyevv in #1546
• [OpenVINO] Refactor CLI quantization by @nikita-savelyevv in #1525
• [OpenVINO] Streamline opevino-genai base pipelines testing by @IlyasMoutawwakil in #1545
• [OpenVINO] Fix VLM mixed quantization by @nikita-savelyevv in #1553
• [OpenVINO] Remove using nncf.torch.patch_torch_operators by @AlexanderDokuchaev in #1555
• [OpenVINO] Deprecate providing trust_remote_code to quantization configs by @nikita-savelyevv in #1558
• [OpenVINO] Add model-specific quantization ignored scopes by @nikita-savelyevv in #1556
• [OpenVINO] Save models immediately after quantization via CLI by @nikita-savelyevv in #1559

New Contributors

• @ml0mbardi made their first contribution in #1508
• @almilosz made their first contribution in #1533
• @zhaohb made their first contribution in #1491

What's Changed

Full Changelog: v1.26.1...v1.27.0

Compatible with transformers>=v4.45,<v5