v1.15.2: Patch release

• Fix compatibility for transformers>=4.38.0 by @echarlaix in #570

v1.15.1: Patch release

• Relax dependency on accelerate and datasets in OVQuantizer by @eaidova in #547

• Disable compilation before applying 4-bit weight compression by @AlexKoff88 in #569

• Update Transformers dependency requirements by @echarlaix in #571

v1.15.0: OpenVINO Tokenizers, quantization configuration

• Add OpenVINO Tokenizers by @apaniukov #513

• Introduce the OpenVINO quantization configuration by @AlexKoff88 #538

• Enable model OpenVINO export by @echarlaix in #557

from diffusers import StableDiffusionPipeline
from optimum.exporters.openvino import export_from_model

model_id = "runwayml/stable-diffusion-v1-5"
model = StableDiffusionPipeline.from_pretrained(model_id)

export_from_model(model, output="ov_model", task="stable-diffusion")

v1.14.0: IPEX models

IPEX models

from optimum.intel import IPEXModelForCausalLM
from transformers import AutoTokenizer, pipeline

model_id = "Intel/q8_starcoder"
model = IPEXModelForCausalLM.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
results = pipe("He's a dreadful magician and")

• Add IPEX models by @echarlaix in #516 / #534 / #536

• Add IPEX models by @ofirzaf in #542 / #543 / #544

Fixes

• Fix position_ids initialization for first inference of stateful models by @eaidova in #532
• Relax requirements to have registered normalized config for decoder models #537 by @eaidova in #537

v1.13.0: 4-bit quantization, stateful models, Whisper

OpenVINO

Weight only 4-bit quantization

• Add weight only 4-bit quantization support by @AlexKoff88 in #469

optimum-cli export openvino --model gpt2 --weight-format int4_sym_g128 ov_model

Stateful

• Add support for stateful models by @eaidova in #493

New architectures

Whisper

• Add support for export and inference for whisper models by @eaidova in #470

v1.12.4: Patch release

• Fix compatibility with transformers v4.37.0 by @echarlaix in #515
• Fix compatibility with transformers v4.37.0 by @echarlaix in #527

v1.12.3: Patch release

• Fix compatibility with diffusers v0.25.0 by @eaidova in #497
• Modify minimum required transformers version by @echarlaix in #498

v1.12.2: Patch release

• Fix compatibility with timm latest release by @echarlaix in #482
• Fix causallm weights compression via quantizer by @eaidova #484
• Fix pkv dtype by @jiqing-feng #481
• Fix compatibility causallm models export with optimum 1.15 by @eaidova #487
• Fix trainer compatibility with transformers>=4.36.0 by @echarlaix #490
• Fix openvino export by @eaidova #470
• Fix INC quantized model loading by @echarlaix #492

v1.12.1: Patch release

• Fix causal language models export by @eaidova in #477

v1.12.0: Weight only quantization, LCM, Pix2Struct , GPTBigCode

OpenVINO

Export CLI

• Add OpenVINO export CLI by @echarlaix in #437

optimum-cli export openvino --model gpt2 ov_model

New architectures

LCMs

• Enable Latent Consistency models OpenVINO export and inference by @echarlaix in #463

from optimum.intel import OVLatentConsistencyModelPipeline

pipe = OVLatentConsistencyModelPipeline.from_pretrained("SimianLuo/LCM_Dreamshaper_v7", export=True)
prompt = "sailing ship in storm by Leonardo da Vinci"
images = pipe(prompt=prompt, num_inference_steps=4, guidance_scale=8.0).images

Pix2Struct

• Add support for export and inference for pix2struct models by @eaidova in #450

GPTBigCode

• Add support for export and inference for GPTBigCode models by @echarlaix in #459

Changes and bugfixes

• Move VAE execution to fp32 precision on GPU by @eaidova in #432
• Enable OpenVINO export without ONNX export step by @eaidova in #397
• Enable 8-bit weight compression for OpenVINO model by @l-bat in #415
• Add image reshaping for statically reshaped OpenVINO SD models by @echarlaix in #428
• OpenVINO device updates by @helena-intel in #434
• Fix decoder model without cache by @echarlaix in #438
• Fix export by @echarlaix in #439
• Added 8 bit weights compression by default for decoders larger than 1B by @AlexKoff88 in #444
• Add fp16 and int8 conversion to OVModels and export CLI by @echarlaix in #443

model = OVModelForCausalLM.from_pretrained(model_id, load_in_8bit=True)

• Create default attention mask when needed but not provided by @eaidova in #457
• Do not automatically cache models when exporting a model in a temporary directory by @helena-intel in #462

Neural Compressor

• Integrate INC weight-only quantization by @mengniwang95 in #417
• Support num_key_value_heads by @jiqing-feng in #447
• Enable ORT model support to INC quantizer by @echarlaix in #436
• fix INC model loading by @echarlaix in #452
• Fix INC modeling by @echarlaix in #453
• Add starcode past-kv shape for TSModelForCausal class by @changwangss in #371
• Fix transformers v4.35.0 compatibility by @echarlaix in #471
• Fix compatibility for optimum next release by @echarlaix in #460

Full Changelog: https://github.com/huggingface/optimum-intel/commits/v1.12.0