️🏗 New architectures support
- Add OpenVINO support for gpt-oss by @echarlaix in #1428
- Add OpenVINO support for MiniCPM-o by @rkazants in #1454
🚀 New Features
- Add feature-extraction and text-classification support for Qwen3 export by @openvino-dev-samples in #1415
- VLM Vision Encoder full quantization by @nikita-savelyevv in #1394
- Add support transformers v4.54 v4.55 by @echarlaix in #1406
- New pipelines by @IlyasMoutawwakil in #1462
- IPEX transformers upgrade to 4.55 by @kaixuanliu in #1467
- Adopt new NNCF mxfp4 quantization logic by @nikita-savelyevv in #1465
🔧 Enhancements & Fixes
- Fix for diffusers 0.35 (and fix and speedup documentation build with uv) by @IlyasMoutawwakil in #1426
- Fix regexp error with searching onnx model by @sbalandi in #1442
- Fix high memory consumption during vision encoder NNCF quantization by @nikita-savelyevv in #1440
- Fix disk by distributing tests by @IlyasMoutawwakil in #1438
- Fix CI rate limiting by @IlyasMoutawwakil in #1449
- Add custom task inferring logic for mistralai/Mistral-7B-Instruct-v0.3 by @nikita-savelyevv in #1413
- Fix editable mode and cleanup/adapt for optimum v2 by @IlyasMoutawwakil in #1457
- Faster cli startup / helper by @IlyasMoutawwakil in #1460
- Fix OpenVINO VLM in-place static quantization by @nikita-savelyevv in #1464
- Remove datasets as required dependency by @echarlaix in #1466
- Fix OpenVINO data-free pipeline quantization by @nikita-savelyevv in #1477
- Skip concat_qkv creation for TP mode by @kaixuanliu in #1481
- Introduce OPENVINO_TEST_DEVICE by @IlyasMoutawwakil in #1479
- Add workaround logic for OpenVINO default int4 quantization of gpt-oss models by @nikita-savelyevv in #1490
- Apply chat_template for MiniCPM-o-2_6 by @Wovchena in #1484
🧹 Deprecations
- Deprecate Quantizer support of nn.Module by @echarlaix in #1421
New Contributors
- @TharescdqTuA made their first contribution in #1436
What's Changed
* Fix for diffusers 0.35 (and fix and speedup documentation build with uv) by @IlyasMoutawwakil in https://github.com//pull/1426
* [OpenVINO]to support qwen3 embedding/rerank by @openvino-dev-samples in https://github.com//pull/1415
* Remove reference to docker based doc building in Makefile by @IlyasMoutawwakil in https://github.com//pull/1430
* VLM Vision Encoder full quantization by @nikita-savelyevv in https://github.com//pull/1394
* Fix main documentation build/push by @IlyasMoutawwakil in https://github.com//pull/1432
* [Docs] Export whisper as stateless during quantization by @nikita-savelyevv in https://github.com//pull/1435
* chore: Fix Misspellings by @TharescdqTuA in https://github.com//pull/1436
* Skip marian tests on OpenVINO 2025.3 by @nikita-savelyevv in https://github.com//pull/1437
* [OV] Remove local imports from quantization.py by @nikita-savelyevv in https://github.com//pull/1433
* Update references after OpenVINO 2025.3 release by @nikita-savelyevv in https://github.com//pull/1441
* Fix regexp error with searching onnx model by @sbalandi in https://github.com//pull/1442
* [OV] Fix high memory consumption during vision encoder quantization by @nikita-savelyevv in https://github.com//pull/1440
* Fix disk by distributing tests by @IlyasMoutawwakil in https://github.com//pull/1438
* chore: Fix Misspellings by @TharescdqTuA in https://github.com//pull/1443
* Benchmarks updates by @ezelanza in https://github.com//pull/1439
* Deprecate Quantizer support of nn.Module by @echarlaix in https://github.com//pull/1421
* Add support transformers v4.54 v4.55 by @echarlaix in https://github.com//pull/1406
* Fix CI rate limiting by @IlyasMoutawwakil in https://github.com//pull/1449
* Add OpenVINO support for gpt-oss by @echarlaix in https://github.com//pull/1428
* Remove legacy code needed for transformers < v4.45 by @echarlaix in https://github.com//pull/1451
* add missing is_transformers_version by @echarlaix in https://github.com//pull/1453
* Add custom task inferring logic for mistralai/Mistral-7B-Instruct-v0.3 by @nikita-savelyevv in https://github.com//pull/1413
* Fix editable mode and cleanup/adapt for optimum v2 by @IlyasMoutawwakil in https://github.com//pull/1457
* [OpenVINO] Support openbmb/MiniCPM-o-2_6 for image-text-to-text task by @rkazants in https://github.com//pull/1454
* Faster cli startup / helper by @IlyasMoutawwakil in https://github.com//pull/1460
* set MAX_TRANSFORMERS_VERSION as inclusive upper bound by @echarlaix in https://github.com//pull/1463
* [OV] Fix VLM in-place static quantization by @nikita-savelyevv in https://github.com//pull/1464
* New pipelines by @IlyasMoutawwakil in https://github.com//pull/1462
* Set optimum-onnx version in setup by @echarlaix in https://github.com//pull/1472
* IPEX transformers upgrade to 4.55 by @kaixuanliu in https://github.com//pull/1467
* CI secret with public token fallback by @IlyasMoutawwakil in https://github.com//pull/1473
* Notebook_updated with benchmark by @ezelanza in https://github.com//pull/1475
* Add note about supported transformers version to export docs by @helena-intel in https://github.com//pull/1476
* [OpenVINO] Fix data-free pipeline quantization by @nikita-savelyevv in https://github.com//pull/1477
* [OpenVINO] Adopt new mxfp4 quantization logic by @nikita-savelyevv in https://github.com//pull/1465
* Remove datasets as required dependency by @echarlaix in https://github.com//pull/1466
* [OpenVINO] Add model inference check to weight-only and pipeline quantization testing by @nikita-savelyevv in https://github.com//pull/1470
* Skip concat_qkv creation for TP mode by @kaixuanliu in https://github.com//pull/1481
* Update OpenVINO workflows to run on Python 3.10 by @nikita-savelyevv in https://github.com//pull/1483
* Introduce OPENVINO_TEST_DEVICE by @IlyasMoutawwakil in https://github.com//pull/1479
* Fix PR upload documentation by @echarlaix in https://github.com//pull/1487
* [OpenVINO][Docs] Update docs to mention how to apply default 4-bit config via Python API by @nikita-savelyevv in https://github.com//pull/1486
* [OpenVINO] Add transformers==4.52 constraint to vision_language_quantization notebook by @nikita-savelyevv in https://github.com//pull/1489
* [OpenVINO] Add workaround logic for default int4 quantization of openai/gpt-oss-20b model by @nikita-savelyevv in https://github.com//pull/1490
* `_GPTOSSQuantizationConfig` follow-up by @nikita-savelyevv in https://github.com//pull/1492
* Apply chat_template for MiniCPM-o-2_6 by @Wovchena in https://github.com//pull/1484
* Update inc workflow python version by @echarlaix in https://github.com//pull/1494
Compatible with transformers>=4.45,<4.56
Full Changelog: v1.25.2...v1.26.0