v1.25.2: Patch release

• Fix tokenizer conversion #1414 by @nikita-savelyevv
• Fix and test stateless encoder decoders #1423 by @IlyasMoutawwakil
• Use eager mask all the time #1424 by @IlyasMoutawwakil

Full Changelog: v1.25.1...v1.25.2

Compatible with transformers>=4.36,<=4.53

v1.25.1: Patch release

• Fix gemma3 for older transformers versions and llava next with mistral decoder #1408 by @IlyasMoutawwakil
• Handle deprecation of forced_decoder_ids in transformers generation_config #1402 by @aleksandr-mokrov and @echarlaix

Full Changelog: v1.25.0...v1.25.1

Compatible with transformers>=4.36,<=4.53

v1.25.0: Text-to-Text generation models quantization

🚀 New Features & Enhancements

• Add quantization for text2text-generation models by @nikita-savelyevv in #1359
• Add OpenVINO support for Mamba and Falcon-mamba by @rkazants in #1360
• Add quantization for SegmentAnything model by @nikita-savelyevv in #1384
• Add support for cb4_f8e4m3 quantization mode by @nikita-savelyevv in #1378
• Add quantization statistics path argument by @nikita-savelyevv in #1392
• Add Transformers 4.53 support by @IlyasMoutawwakil in #1377

New Contributors

• @mitruska made their first contribution in #1375
• @ezelanza made their first contribution in #1385

What's Changed

• Add OpenVINO weight compression tests for llama4 by @nikita-savelyevv in #1369
• Fix IPEX model loading for sentence-transformers v5 by @echarlaix in #1370
• Update OpenVINO documentation with newly supported tasks by @rkazants in #1371
• [Docs] Optimization table on click feedback logic by @nikita-savelyevv in #1372
• Fix attr name typo in model_configs for llava-next compatibility with transformers 4.51.3 by @mitruska in #1375
• [OV] Add quantization for text2text-generation models by @nikita-savelyevv in #1359
• free up disk for slow/full ci by @IlyasMoutawwakil in #1376
• Original model types by @IlyasMoutawwakil in #1329
• Add openvino VLM quantization notebook by @echarlaix in #1382
• Remove notebook redundant quantization configs by @echarlaix in #1383
• [OV] Prepare quantization dataset collection logic to transition to datasets v4.0 by @nikita-savelyevv in #1381
• [OpenVINO] Add support for Mamba and Falcon-mamba by @rkazants in #1360
• Improve VLM quantization notebook structure by @ezelanza in #1385
• [OV] Add quantization for SegmentAnything model by @nikita-savelyevv in #1384
• [OV] Update the reference number of int8 nodes for SANA model by @nikita-savelyevv in #1386
• Add notebook quantization config paragraph by @echarlaix in #1390
• [TTS] Fix second generation for Speech T5 TSS by @rkazants in #1389
• fix auto_model_class for OVModelForVisualCausalLM by @echarlaix in #1391
• Add support for cb4_f8e4m3 quantization mode. by @nikita-savelyevv in #1378
• Add quantization statistics path argument by @nikita-savelyevv in #1392
• Transformers 4.53 support by @IlyasMoutawwakil in #1377

Compatible with transformers>=4.36,<=4.53

Full Changelog: v1.24.0...v1.25.0

v1.24.0: OVPipelineQuantizationConfig

🚀 New Features & Enhancements

Optimum 1.26 compatibility by @IlyasMoutawwakil in #1352

OpenVINO

• Introduce default full quantization configs for clip models by @nikita-savelyevv in #1302
• Introduce OVPipelineQuantizationConfig by @nikita-savelyevv in #1310
• Add int8 PTQ configs for some fill-mask models by @nikita-savelyevv in #1331
• Add transformers v4.52 compatibility by @eaidova in #1319
• Add compression config for Qwen/Qwen2.5-Coder-3B-Instruct by @MaximProshin in #1355
• [OV] Add support for data-free AWQ by @nikita-savelyevv in #1349
• Convert dataclasses to dicts in quantization config before saving by @nikita-savelyevv in #1362
• Remove reshaping for stateful decoders by @echarlaix in #1333

IPEX

• Add transformers v4.52 compatibility by @jiqing-feng in #1317

🔧 Key Fixes & Optimizations

• Raise if converted subcomponent not found by @echarlaix in #1303
• Keep Hybrid Quantization only for diffusion pipelines by @nikita-savelyevv in #1313
• Fix whisper with auto language detection by @eaidova in #1314
• Fix vision embeddings export for maira by @eaidova in #1320
• Fix VLM calibration dataset collection by @nikita-savelyevv in #1321
• Resize large images during VLM calibration data collection by @nikita-savelyevv in #1322
• Resolve logger warnings by @emmanuel-ferdman in #1324
• Fix progress bar during calibration dataset collection by @nikita-savelyevv in #1323
• Fix ESM models export and add it to supported by @eaidova in #1328
• Allow skip trace check for sentence stransformers by @eaidova in #1332
• Fix int value recompile by @jiqing-feng in #1335
• Fix TP tensor dimension dismatch for IPEX models by @kaixuanliu in #1340
• Updated Qwen3-8b compression config by @MaximProshin in #1341

New Contributors

• @kilavvy made their first contribution in #1345
• @maximevtush made their first contribution in #1347
• @leopardracer made their first contribution in #1351

What's Changed

• Dev version by @echarlaix in #1309
• Update number of int8 nodes for Segment Anything model by @nikita-savelyevv in #1311
• [OV][Docs] Keep Hybrid Quantization only for diffusion pipelines by @nikita-savelyevv in #1313
• raise if converted subcomponent not found by @echarlaix in #1303
• [OV] Introduce default full quantization configs for clip models by @nikita-savelyevv in #1302
• fix whisper with auto language detection by @eaidova in #1314
• fix vision embeddings export for maira by @eaidova in #1320
• [OV] Fix VLM calibration dataset collection by @nikita-savelyevv in #1321
• [OV] Resize large images during VLM calibration data collection by @nikita-savelyevv in #1322
• Resolve logger warnings by @emmanuel-ferdman in #1324
• [OV] Fix progress bar during calibration dataset collection by @nikita-savelyevv in #1323
• Limit INC version to fix CI. by @changwangss in #1325
• [OV] Update AWQ test to pass on NNCF develop by @nikita-savelyevv in #1326
• Fix ESM models export and add it to supported by @eaidova in #1328
• Introduce OVPipelineQuantizationConfig by @nikita-savelyevv in #1310
• [OV] Add int8 PTQ configs for some fill-mask models. by @nikita-savelyevv in #1331
• allow skip trace check for sentence stransformers by @eaidova in #1332
• fix int value recompile by @jiqing-feng in #1335
• Add style bot by @echarlaix in #1337
• Fix setup.py to support INC latest version 3.4.1 by @changwangss in #1339
• fix bug when using tp, tensor dimension dismatch by @kaixuanliu in #1340
• fix optimum version by @echarlaix in #1344
• Updated Qwen3-8b compression config by @MaximProshin in #1341
• Fix Typo in Error Message for Sequence Length Validation by @kilavvy in #1345
• Fix Typographical Errors in Documentation String by @maximevtush in #1347
• upgrade windows runner image by @echarlaix in #1350
• Upgrade transformers version to 4.52 for ipex patching by @jiqing-feng in #1317
• Minor Typo Fixes in Comments for Quantized Generation Demo Notebook by @leopardracer in #1351
• fix openvino for compatibility with transformers 4.52 by @eaidova in #1319
• Optimum 2.26 compatibility by @IlyasMoutawwakil in #1352
• [OV] Update reference number of fp8 fake convert nodes by @nikita-savelyevv in #1348
• Compression config for Qwen/Qwen2.5-Coder-3B-Instruct by @MaximProshin in #1355
• Docs: Fix typos in quantized generation demo notebook by @kilavvy in #1356
• update style bot permission and token by @echarlaix in #1357
• [OV] Add support for data-free AWQ by @nikita-savelyevv in #1349
• Add documentation workflow by @echarlaix in #1361
• Fix style by @echarlaix in #1363
• fix by @echarlaix in #1364
• Fix documentation workflow by @echarlaix in #1365
• Convert dataclasses to dicts in quantization config before saving by @nikita-savelyevv in #1362
• Remove reshaping for stateful decoders by @echarlaix in #1333

Compatible with transformers>=4.36,<=4.52

Full Changelog: v1.23.0...v1.24.0

v1.23.1: Patch release

Full Changelog: v1.23.0...v1.23.1

v1.23.0: DeepSeek, Llama 4, LTX-Video

🚀 New Features & Enhancements

OpenVINO

• Add MAIRA-2 support by @eaidova in #1145
• Add support for nf4_f8e4m3 quantization mode by @nikita-savelyevv in #1148
• Add DeepSeek support by @eaidova in #1155
• Add Qwen2.5-VL support by @eaidova in #1163
• Add LLaVA-Next-Video support by @eaidova in #1183
• Add GOT-OCR2 support by @eaidova in #1202
• Add Gemma 3 support by @eaidova in #1198
• Add SmolVLM and Idefics3 support by @eaidova in #1210
• Add Phi-3-MoE support by @eaidova in #1215
• Add OVSamModel for inference by @eaidova in #1229
• Add Phi-4-multimodal support by @eaidova in #1201
• Add Llama 4 support by @eaidova in #1226
• Add zero-shot-Image-classification support by @eaidova in #1273
• Add PTQ support for OVModelForZeroShotImageClassification by @nikita-savelyevv in #1283
• Add diffuers full int8 quantization Support by @l-bat in #1193
• Add SANA-Sprint support by @eaidova in #1245
• Add PTQ support for OVModelForMaskedLM by @nikita-savelyevv in #1268
• Add LTX-Video support by @eaidova in #1264
• Add Qwen3 and Qwen3-MOE support by @openvino-dev-samples in #1214
• Add SpeechT5 text-to-speech support by OpenVINO by @rkazants in #1230
• Add GLM4 support by @openvino-dev-samples in #1249
• PTQ support for OVModelForFeatureExtraction and OVSentenceTransformer by @nikita-savelyevv in #1257
• Introduce OVCalibrationDatasetBuilder by @nikita-savelyevv in #1232

IPEX

• Add Qwen2 support by @jiqing-feng in #1107
• Enable quantization model support by @jiqing-feng in #1074
• Add support for flash decoding on xpu by @kaixuanliu in #1118
• Add Phi support by @jiqing-feng in #1175
• Enable compilation for patched model with paged attention by @jiqing-feng in #1253
• Add Mistral modeling optimization support for ipex by @kaixuanliu in #1269

Transformers compatibility

• Add compatibility with transformers v4.49 by @echarlaix in #1172
• Add compatibility with transformers v4.50 and v4.51 by @IlyasMoutawwakil in #1242

🔧 Key Fixes & Optimizations

• Fix misplaced configs saving by @eaidova in #1159
• Check if nncf is installed before running quantization from optimum-cli by @nikita-savelyevv in #1154
• Fix automatic-speech-recognition-with-past quantization from CLI by @nikita-savelyevv in #1180
• Propagate OV QuantizationConfig kwargs to nncf calls by @nikita-savelyevv in #1179
• Fix model field names for OVBaseModelForSeq2SeqLM by @nikita-savelyevv in #1184
• Align loading dtype logic for diffusers with other models by @eaidova in #1187
• Fix generation for statically reshaped diffusion pipeline by @eaidova in #1199
• Add ov_submodels property to OVBaseModel by @nikita-savelyevv in #1177
• Fix flux and sana export with diffusers 0.33+ by @eaidova in #1236
• Update pkv precision at save_pretrained call by @nikita-savelyevv in #1235
• Remove ONNX fallback when converting to OpenVINO by @eaidova in #1272
• Fix custom dataset processing for text encoding tasks by @nikita-savelyevv in #1286
• Fix openvino decoder models output by @echarlaix in #1308

What's Changed

• fix export phi3 with --trust-remote-code by @eaidova in #1147
• Skip test_aware_training_quantization test by @nikita-savelyevv in #1149
• Check if nncf is installed before running quantization from optimum-cli by @nikita-savelyevv in #1154
• enable qwen2 model by @jiqing-feng in #1107
• maira2 support by @eaidova in #1145
• Add slow tests for lower transformers version by @echarlaix in #1144
• fix misplaced configs saving by @eaidova in #1159
• Add default int4 config for DeepSeek-R1-Distill-Llama-8B by @nikita-savelyevv in #1158
• Remove unnecessary SD reload from saved dir by @l-bat in #1162
• resolve complicated chat templates during tokenizer saving by @eaidova in #1151
• Trigger tests for maira2 for compatible transformers version by @echarlaix in #1161
• use Tensor.numpy() instead np.array(Tensor) by @eaidova in #1153
• [OV] Add support for nf4_f8e4m3 quantization mode by @nikita-savelyevv in #1148
• support updated chat template for llava-next by @eaidova in #1166
• avoid extra reshaping to max_model_lenght for unet by @eaidova in #1164
• Enable quant model support by @jiqing-feng in #1074
• [OV] Add default int4 configurations for DeepSeek-R1-Distill-Qwen models by @nikita-savelyevv in #1168
• Deprecate OVTrainer by @nikita-savelyevv in #1167
• Support deeepseek models export by @eaidova in #1155
• add support for flash decoding on xpu by @kaixuanliu in #1118
• deprecate TSModelForCausalLM by @echarlaix in #1173
• transformers 4.49 by @echarlaix in #1172
• Update ipex Ci to torch 2.6 by @jiqing-feng in #1176
• add support qwen2.5vl by @eaidova in #1163
• enable phi by @jiqing-feng in #1175
• Add ov_submodels property to OVBaseModel by @nikita-savelyevv in #1177
• [OV] Fix automatic-speech-recognition-with-past quantization from CLI by @nikita-savelyevv in #1180
• Propagate OV*QuantizationConfig kwargs to nncf calls by @nikita-savelyevv in #1179
• [OV] Add int4 config for Llama-3.1-8b model id aliases by @nikita-savelyevv in #1182
• Fix model field names for OVBaseModelForSeq2SeqLM by @nikita-savelyevv in #1184
• [OV] Enable back phi3_v 4bit compression test by @nikita-savelyevv in #1185
• align loading dtype logic for diffusers with other models by @eaidova in #1187
• attempt to resolve 4.49 compatibility issues and fix input processing… by @eaidova in #1190
• fix logits_to_keep by @jiqing-feng in #1188
• warm up do not work for compiled model by @jiqing-feng in #1189
• Add default int4 configs for Phi-4-mini-instruct and Qwen2.5-7B-Instruct by @nikita-savelyevv in #1194
• add support llava-next-video by @eaidova in #1183
• upgrade transformers to 4.49 for patching models by @jiqing-feng in #1196
• add support got-ocr2 by @eaidova in #1202
• fix generation for statically reshaped diffusion pipeline...

v1.22.0: Qwen2-VL, Granite, Sana, Sentence Transformers

OpenVINO

• Add quantization of Whisper pipeline by @nikita-savelyevv in #1040
• Add Qwen2-VL support by @eaidova in #1042
• Add AWQ models support by @mvafin in #1049
• Update default OV configuration by @KodiaqQ in #1057
• Introduce --quant-mode cli argument enabling full quantization via optimum-cli by @nikita-savelyevv in #1061
• Merge decoder and decoder with past to stateful for seq2seq models by @eaidova in #1078
• Add transformers 4.47 support by @IlyasMoutawwakil in #1088
• Add GLM-Edge models support by @eaidova in #1089
• Add Granite and GraniteMoe models support by @eaidova in #1099
• Add fp8 implementation by @KodiaqQ in #1100
• Add Flux Fill inpainting pipeline support by @eaidova in #1095
• Add Sana support by @eaidova in #1106
• Add v4.48 transformers support by @IlyasMoutawwakil in #1136

IPEX

• Add support to sentence transformers models by @echarlaix in #1034

from optimum.intel import IPEXSentenceTransformer

model = IPEXSentenceTransformer.from_pretrained(model_id)

• Add support to text-to-text task by @jiqing-feng in #1054

from optimum.intel import IPEXModelForSeq2SeqLM

model = IPEXModelForSeq2SeqLM.from_pretrained(model_id)

• Enable Flash Attention by @jiqing-feng in #1065

Compatible with transformers>=4.36,<=4.48

Full Changelog: v1.21.0...v1.22.0

v1.21.0: SD3, Flux, MiniCPM, NanoLlava, VLM Quantization, XPU, PagedAttention

What's Changed

OpenVINO

Diffusers

• SD3 and Flux pipelines support by @eaidova in #916

VLMs Modeling

• MiniCPMv support by @eaidova in #972
• NanoLlava support by @eaidova in #969
• Phi3v support by @eaidova in #977

NNCF

• Quantization support for CausalVisualLMs by @nikita-savelyevv in #951
• NF4 data type support for OV weight compression by @l-bat in #988
• NNCF 2.14 new features support by @nikita-savelyevv in #997

IPEX

• Unified XPU/CPU modeling with custom PagedAttention cache for LLMs by @sywangyi in #1009

INC

• Layer-wise quantization support by @changwangss in #1040

New Contributors

• @emmanuel-ferdman made their first contribution in #974
• @mvafin made their first contribution in #1033

Compatible with transformers>=4.36,<=4.46

Full Changelog: v1.20.0...v1.21.0

v1.20.1: Patch release

• Fix lora unscaling in diffusion pipelines by @eaidova in #937
• Fix compatibility with diffusers < 0.25.0 by @eaidova in #952
• Allow to use SDPA in clip models by @eaidova in #941
• Updated OVPipelinePart to have separate ov_config by @e-ddykim in #957
• Symbol use in optimum: fix misprint by @jane-intel in #948
• Fix temporary directory saving by @eaidova in #959
• Disable warning about tokenizers version for ov tokenizers >= 2024.5 by @eaidova in #962
• Restore original model_index.json after save_pretrained call by @eaidova in #961
• Add v4.46 transformers support by @echarlaix in #960

v1.20.0: multi-modal and OpenCLIP models support, transformers v4.45

OpenVINO

Multi-modal models support

Adding OVModelForVisionCausalLM by @eaidova in #883

OpenCLIP models support

Adding OpenCLIP models support by @sbalandi in #857

from optimum.intel import OVModelCLIPVisual, OVModelCLIPText

visual_model = OVModelCLIPVisual.from_pretrained(model_name_or_path)
text_model  = OVModelCLIPText.from_pretrained(model_name_or_path)
image = processor(image).unsqueeze(0)
text = tokenizer(["a diagram", "a dog", "a cat"])
image_features = visual_model(image).image_features
text_features = text_model(text).text_features

Diffusion pipeline

Adding OVDiffusionPipeline to simplify diffusers model loading by @IlyasMoutawwakil in #889

  model_id = "stabilityai/stable-diffusion-xl-base-1.0"
- pipeline = OVStableDiffusionXLPipeline.from_pretrained(model_id)
+ pipeline = OVDiffusionPipeline.from_pretrained(model_id)
  image = pipeline("sailing ship in storm by Leonardo da Vinci").images[0]

NNCF GPTQ support

GPTQ support by @nikita-savelyevv in #912

Transformers v4.45

Transformers v4.45 support by @echarlaix in #902

Subfolder

Remove the restriction for the model's config to be in the model's subfolder by @tomaarsen in #933

New Contributors

• @jane-intel made their first contribution in #696
• @andreyanufr made their first contribution in #903
• @MaximProshin made their first contribution in #905
• @tomaarsen made their first contribution in #931

Compatible with transformers>=4.36,<=4.45

Full Changelog: v1.19.0...v1.20.0