OpenVINO™ GenAI is a library of the most popular Generative AI model pipelines, optimized execution methods, and samples that run on top of highly performant OpenVINO Runtime.

This library is friendly to PC and laptop execution, and optimized for resource consumption. It requires no external dependencies to run generative models as it already includes all the core functionality (e.g. tokenization via openvino-tokenizers).

Please follow the following blogs to setup your first hands-on experience with C++ and Python samples.

• How to Build OpenVINO™ GenAI APP in C++
• How to run Llama 3.2 locally with OpenVINO™

OpenVINO™ GenAI library provides very lightweight C++ and Python APIs to run following Generative Scenarios:

• Text generation using Large Language Models. For example, chat with local LLaMa model
• Image generation using Diffuser models, for example, generation using Stable Diffusion models
• Speech recognition using Whisper family models
• Text generation using Large Visual Models, for instance, Image analysis using LLaVa or miniCPM models family

Library efficiently supports LoRA adapters for Text and Image generation scenarios:

• Load multiple adapters per model
• Select active adapters for every generation
• Mix multiple adapters with coefficients via alpha blending

All scenarios are run on top of OpenVINO Runtime that supports inference on CPU, GPU and NPU. See here for platform support matrix.

OpenVINO™ GenAI library provides a transparent way to use state-of-the-art generation optimizations:

• Speculative decoding that employs two models of different sizes and uses the large model to periodically correct the results of the small model. See here for more detailed overview
• KVCache token eviction algorithm that reduces the size of the KVCache by pruning less impacting tokens.

Additionally, OpenVINO™ GenAI library implements a continuous batching approach to use OpenVINO within LLM serving. Continuous batching library could be used in LLM serving frameworks and supports the following features:

• Prefix caching that caches fragments of previous generation requests and corresponding KVCache entries internally and uses them in case of repeated query. See here for more detailed overview

Continuous batching functionality is used within OpenVINO Model Server (OVMS) to serve LLMs, see here for more details.

    # Installing OpenVINO GenAI via pip
    pip install openvino-genai

    # Install optimum-intel to be able to download, convert and optimize LLMs from Hugging Face
    # Optimum is not required to run models, only to convert and compress
    pip install optimum-intel@git+https://github.com/huggingface/optimum-intel.git

    # (Optional) Install (TBD) to be able to download models from Model Scope

#(Basic) download and convert to OpenVINO TinyLlama-Chat-v1.0 model
optimum-cli export openvino --model "TinyLlama/TinyLlama-1.1B-Chat-v1.0" --weight-format fp16 --trust-remote-code "TinyLlama-1.1B-Chat-v1.0"

#(Recommended) download, convert to OpenVINO and compress to int4 TinyLlama-Chat-v1.0 model
optimum-cli export openvino --model "TinyLlama/TinyLlama-1.1B-Chat-v1.0" --weight-format int4 --trust-remote-code "TinyLlama-1.1B-Chat-v1.0"

import openvino_genai as ov_genai
#Will run model on CPU, GPU or NPU are possible options
pipe = ov_genai.LLMPipeline("./TinyLlama-1.1B-Chat-v1.0/", "CPU")
print(pipe.generate("The Sun is yellow because", max_new_tokens=100))

#include "openvino/genai/llm_pipeline.hpp"
#include <iostream>

int main(int argc, char* argv[]) {
    std::string models_path = argv[1];
    ov::genai::LLMPipeline pipe(models_path, "CPU");
    std::cout << pipe.generate("The Sun is yellow because", ov::genai::max_new_tokens(100)) << '\n';
}

optimum-cli export openvino --model openbmb/MiniCPM-V-2_6 --trust-remote-code MiniCPM-V-2_6

import openvino_genai as ov_genai
#Will run model on CPU, GPU is a possible option
pipe = ov_genai.VLMPipeline("./MiniCPM-V-2_6/", "CPU")
rgb = read_image("cat.jpg")
print(pipe.generate(prompt, image=rgb, max_new_tokens=100))

#include "load_image.hpp"
#include <openvino/genai/visual_language/pipeline.hpp>
#include <iostream>

int main(int argc, char* argv[]) {
    std::string models_path = argv[1];
    ov::genai::VLMPipeline pipe(models_path, "CPU");
    ov::Tensor rgb = utils::load_image(argv[2]);
    std::cout << pipe.generate(
        prompt,
        ov::genai::image(rgb),
        ov::genai::max_new_tokens(100)
    ) << '\n';
}

#Download and convert to OpenVINO dreamlike-anime-1.0 model
optimum-cli export openvino --model dreamlike-art/dreamlike-anime-1.0 --task stable-diffusion --weight-format fp16 dreamlike_anime_1_0_ov/FP16

import argparse
from PIL import Image
import openvino_genai

def main():
    parser = argparse.ArgumentParser()
    parser.add_argument('model_dir')
    parser.add_argument('prompt')
    args = parser.parse_args()

    device = 'CPU'  # GPU, NPU can be used as well
    pipe = openvino_genai.Text2ImagePipeline(args.model_dir, device)
    image_tensor = pipe.generate(
        args.prompt,
        width=512,
        height=512,
        num_inference_steps=20
    )

    image = Image.fromarray(image_tensor.data[0])
    image.save("image.bmp")

#include "openvino/genai/image_generation/text2image_pipeline.hpp"
#include "imwrite.hpp"
int main(int argc, char* argv[]) {

   const std::string models_path = argv[1], prompt = argv[2];
   const std::string device = "CPU";  // GPU, NPU can be used as well

   ov::genai::Text2ImagePipeline pipe(models_path, device);
   ov::Tensor image = pipe.generate(prompt,
        ov::genai::width(512),
        ov::genai::height(512),
        ov::genai::num_inference_steps(20));

   imwrite("image.bmp", image, true);
}

#Download and convert to OpenVINO whisper-base model
optimum-cli export openvino --trust-remote-code --model openai/whisper-base whisper-base

import openvino_genai
import librosa


def read_wav(filepath):
    raw_speech, samplerate = librosa.load(filepath, sr=16000)
    return raw_speech.tolist()

device = "CPU" # GPU can be used as well
pipe = openvino_genai.WhisperPipeline("whisper-base", device)
raw_speech = read_wav("sample.wav")
print(pipe.generate(raw_speech))

#include <iostream>

#include "audio_utils.hpp"
#include "openvino/genai/whisper_pipeline.hpp"

int main(int argc, char* argv[]) {
    std::filesystem::path models_path = argv[1];
    std::string wav_file_path = argv[2];
    std::string device = "CPU"; // GPU can be used as well

    ov::genai::WhisperPipeline pipeline(models_path, device);

    ov::genai::RawSpeechInput raw_speech = utils::audio::read_wav(wav_file_path);

    std::cout << pipeline.generate(raw_speech, ov::genai::max_new_tokens(100)) << '\n';
}

• List of supported models (NOTE: models can work, but were not tried yet)
• OpenVINO LLM inference Guide
• Optimum-intel and OpenVINO

The OpenVINO™ GenAI repository is licensed under Apache License Version 2.0. By contributing to the project, you agree to the license and copyright terms therein and release your contribution under these terms.