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 by @eaidova in #1199
• add gemma3 support by @eaidova in #1198
• enable awq tests by @jiqing-feng in #1195
• fix tests running with nightly by @eaidova in #1205
• fix internvl2 patching for transformes>=4.48 by @eaidova in #1206
• Support full int8 quantization for diffusers by @l-bat in #1193
• Update default int4 config for llama-2-7b-chat-hf by @nikita-savelyevv in #1216
• Patch by @jiqing-feng in #1200
• add falcon3 simplified chat template by @eaidova in #1217
• add support SmolVLM and Idefics3 models by @eaidova in #1210
• fix checking available files if from_onnx=True by @eaidova in #1208
• remove XPULinearXXX class definition for ipex by @kaixuanliu in #1212
• phi3moe support by @eaidova in #1215
• Fix crash issue of IPEX XPU's rotary_embedding API by @kaixuanliu in #1218
• Add simplified chat template for Mistral-7B-Instruct-v0.3 by @sbalandi in #1221
• Add OpenVINO Optimization Support Matrix table by @nikita-savelyevv in #1219
• Fix tests by @IlyasMoutawwakil in #1222
• [OV, Docs] Dark mode compatibility of optimization support matrix by @nikita-savelyevv in #1225
• [OV] INT4 configs for Qwen2.5-1.5B-Instruct and Llama-3.2-1B-Instruct by @nikita-savelyevv in #1231
• fix flux and sana export with diffusers 0.33+ by @eaidova in #1236
• Fix more ci by @IlyasMoutawwakil in #1228
• move sentencepiece to test requirements for unblocking python3.13 by @eaidova in #1240
• Update pkv precision at save_pretrained call by @nikita-savelyevv in #1235
• Install diffusers requirement during OV Full and Slow tests by @nikita-savelyevv in #1243
• Add nncf version as openvino model runtime flag by @nikita-savelyevv in #1244
• add support sana-sprint by @eaidova in #1245
• Replace compvis safety checker model with a tiny version by @nikita-savelyevv in #1250
• restore cache_position input in whisper by @eaidova in #1254
• Introduce OVCalibrationDatasetBuilder (part 1/2) by @nikita-savelyevv in #1232
• Update transformers by @IlyasMoutawwakil in #1242
• fix typo in unpatching decoder models by @eaidova in #1259
• Enable GLM4 for openvino by @openvino-dev-samples in #1249
• Fix INC CI tests failed by @changwangss in #1262
• [OV] Update --sym argument description by @nikita-savelyevv in #1263
• PTQ support for OVModelForFeatureExtraction and OVSentenceTransformer by @nikita-savelyevv in #1257
• Fix ipex CI by @jiqing-feng in #1260
• add OVSamModel for inference by @eaidova in #1229
• Fix documentation by @echarlaix in #1266
• PTQ support for OVModelForMaskedLM by @nikita-savelyevv in #1268
• upgrade transformers for ipex by @kaixuanliu in #1267
• Clean ipex tests by @jiqing-feng in #1270
• support LTX-video by @eaidova in #1264
• fix gptj export for transformers>4.49 by @eaidova in #1271
• Support SpeechT5 text-to-speech pipeline by OpenVINO by @rkazants in #1230
• Enable Qwen3 and Qwen3-MOE for OpenVINO by @openvino-dev-samples in #1214
• Add Mistral modeling optimization support for ipex by @kaixuanliu in #1269
• Avoid use of deprecated openvino.runtime by @rkazants in #1274
• upgrade ipex to 2.7 by @jiqing-feng in #1277
• support Zero-shot-Image-Classification by @eaidova in #1273
• fix bug when bs > 1 and do not provide position_ids for Mistral model by @kaixuanliu in #1276
• Fix minimum diffusers version for LTXPipeline by @echarlaix in #1279
• removal onnx fallback by @eaidova in #1272
• Replace deprecated openvino.runtime by @echarlaix in #1280
• int4 configs for Qwen3-1.7B and Qwen3-4B by @MaximProshin in #1282
• int4 config for starcoder2-15b by @MaximProshin in #1285
• Fix spaces in chat_template simplification for Mistral-7B-Instruct-v0.3 by @Wovchena in #1281
• Enable compile for ipex patched model with paged attention by @jiqing-feng in #1253
• Fix typos by @omahs in #1284
• int4 config for Qwen/Qwen3-8B by @MaximProshin in #1287
• add support phi4 multimodal by @eaidova in #1201
• Fix custom dataset processing for text encoding tasks by @nikita-savelyevv in #1286
• llama4 by @eaidova in #1226
• Default config for microsoft/Phi-4-multimodal-instruct by @nikita-savelyevv in #1289
• Remove OVBaseModelForSeq2SeqLM by @echarlaix in #1278
• Add compression tests for phi4mm by @nikita-savelyevv in #1292
• Update tiny model for ipex lanchain tests by @echarlaix in #1296
• PTQ support for OVModelForZeroShotImageClassification by @nikita-savelyevv in #1283
• add token classification for qwen2 by @eaidova in #1299
• fix beam search test for latest optimum by @eaidova in #1290
• fix automatic task detection for phi4-multimodal during data-aware quant by @eaidova in #1293
• provide workaround for export openai whisper models by @eaidova in #1301
• Introduce named test reference values per submodel by @nikita-savelyevv in #1300
• set optimum version in setup by @echarlaix in #1298
• add gemma to skip check trace models by @echarlaix in #1306
• remove forcing input_ids by @eaidova in #1307
• Fixes openvino decoder models output by @echarlaix in #1308
• Make a WA to avoid XPU crash for API PagedAttention.reshape_and_cache_flash by @kaixuanliu in #1288

Compatible with transformers>=4.36,<=4.51

Full Changelog: v1.22.0...v1.23.0