Inference with OpenVINO GenAI

OpenVINO™ GenAI is a library of pipelines and methods, extending the OpenVINO runtime to work with generative AI models more efficiently. This article provides reference code and guidance on its usage. Note that the base OpenVINO version will not work with these instructions, make sure to install OpenVINO with GenAI.

Here is sample code for several Generative AI use case scenarios. Note that these are very basic examples and may need adjustments for your specific needs, like changing the inference device.

For a more extensive instruction and additional options, see the step-by-step chat-bot guide below.

Text-to-Image Generation

OpenVINO GenAI introduces openvino_genai.Text2ImagePipeline for inference of text-to-image models such as: as Stable Diffusion 1.5, 2.1, XL, LCM, Flex, and more. See the following usage example for reference.

Python

text2image.py

import argparse

import openvino_genai
from PIL import Image


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

    device = 'CPU'  # GPU 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,
        num_images_per_prompt=1)

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

lora_text2image.py

import openvino as ov
import openvino_genai

def image_write(path: str, image_tensor: ov.Tensor):
    from PIL import Image
    image = Image.fromarray(image_tensor.data[0])
    image.save(path)


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument('models_path')
    parser.add_argument('prompt')
    args, adapters = parser.parse_known_args()

    prompt = args.prompt

    device = "CPU"  # GPU, NPU can be used as well
    adapter_config = openvino_genai.AdapterConfig()

    # Multiple LoRA adapters applied simultaneously are supported, parse them all and corresponding alphas from cmd parameters:
    for i in range(int(len(adapters) / 2)):
        adapter = openvino_genai.Adapter(adapters[2 * i])
        alpha = float(adapters[2 * i + 1])
        adapter_config.add(adapter, alpha)

    # LoRA adapters passed to the constructor will be activated by default in next generates
    pipe = openvino_genai.Text2ImagePipeline(args.models_path, device, adapters=adapter_config)

    print("Generating image with LoRA adapters applied, resulting image will be in lora.bmp")
    image = pipe.generate(prompt,
                          width=512,
                          height=896,
                          num_inference_steps=20,
                          rng_seed=42)

    image_write("lora.bmp", image)
    print("Generating image without LoRA adapters applied, resulting image will be in baseline.bmp")
    image = pipe.generate(prompt,
                          # passing adapters in generate overrides adapters set in the constructor; openvino_genai.AdapterConfig() means no adapters
                          adapters=openvino_genai.AdapterConfig(),
                          width=512,
                          height=896,
                          num_inference_steps=20,
                          rng_seed=42)
    image_write("baseline.bmp", image)

For more information, refer to the Python sample

C++

text2image.cpp

#include "openvino/genai/image_generation/text2image_pipeline.hpp"

#include "imwrite.hpp"

int32_t main(int32_t argc, char* argv[]) try {
    OPENVINO_ASSERT(argc == 3, "Usage: ", argv[0], " <MODEL_DIR> '<PROMPT>'");

    const std::string models_path = argv[1], prompt = argv[2];
    const std::string device = "CPU";  // GPU 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),
        ov::genai::num_images_per_prompt(1));

    // writes `num_images_per_prompt` images by pattern name
    imwrite("image_%d.bmp", image, true);

    return EXIT_SUCCESS;
} catch (const std::exception& error) {
    try {
        std::cerr << error.what() << '\n';
    } catch (const std::ios_base::failure&) {}
    return EXIT_FAILURE;
} catch (...) {
    try {
        std::cerr << "Non-exception object thrown\n";
    } catch (const std::ios_base::failure&) {}
    return EXIT_FAILURE;
}

lora_text2image.cpp

#include "openvino/genai/image_generation/text2image_pipeline.hpp"

#include "imwrite.hpp"

int32_t main(int32_t argc, char* argv[]) try {
    OPENVINO_ASSERT(argc >= 3 && (argc - 3) % 2 == 0, "Usage: ", argv[0], " <MODEL_DIR> '<PROMPT>' [<LORA_SAFETENSORS> <ALPHA> ...]]");

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

    ov::genai::AdapterConfig adapter_config;
    // Multiple LoRA adapters applied simultaneously are supported, parse them all and corresponding alphas from cmd parameters:
    for(size_t i = 0; i < (argc - 3)/2; ++i) {
        ov::genai::Adapter adapter(argv[3 + 2*i]);
        float alpha = std::atof(argv[3 + 2*i + 1]);
        adapter_config.add(adapter, alpha);
    }

    // LoRA adapters passed to the constructor will be activated by default in next generates
    ov::genai::Text2ImagePipeline pipe(models_path, device, ov::genai::adapters(adapter_config));

    std::cout << "Generating image with LoRA adapters applied, resulting image will be in lora.bmp\n";
    ov::Tensor image = pipe.generate(prompt,
        ov::genai::width(512),
        ov::genai::height(896),
        ov::genai::num_inference_steps(20),
        ov::genai::rng_seed(42));
    imwrite("lora.bmp", image, true);

    std::cout << "Generating image without LoRA adapters applied, resulting image will be in baseline.bmp\n";
    image = pipe.generate(prompt,
        ov::genai::adapters(),  // passing adapters in generate overrides adapters set in the constructor; adapters() means no adapters
        ov::genai::width(512),
        ov::genai::height(896),
        ov::genai::num_inference_steps(20),
        ov::genai::rng_seed(42));
    imwrite("baseline.bmp", image, true);

    return EXIT_SUCCESS;
} catch (const std::exception& error) {
    try {
        std::cerr << error.what() << '\n';
    } catch (const std::ios_base::failure&) {}
    return EXIT_FAILURE;
} catch (...) {
    try {
        std::cerr << "Non-exception object thrown\n";
    } catch (const std::ios_base::failure&) {}
    return EXIT_FAILURE;
}

For more information, refer to the C++ sample

Speech Recognition

The application performs inference on speech recognition Whisper Models. The samples include the WhisperPipeline class and use audio files in WAV format at a sampling rate of 16 kHz as input.

Python

import openvino_genai
import librosa


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


def infer(model_dir: str, wav_file_path: str):
    device = "CPU"  # GPU or NPU can be used as well.
    pipe = openvino_genai.WhisperPipeline(model_dir, device)

    # The pipeline expects normalized audio with a sampling rate of 16kHz.
    raw_speech = read_wav(wav_file_path)
    result = pipe.generate(
        raw_speech,
        max_new_tokens=100,
        language="<|en|>",
        task="transcribe",
        return_timestamps=True,
    )

    print(result)

    for chunk in result.chunks:
        print(f"timestamps: [{chunk.start_ts}, {chunk.end_ts}] text: {chunk.text}")

For more information, refer to the Python sample.

C++

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

int main(int argc, char* argv[]) try {
    if (3 > argc) {
        throw std::runtime_error(std::string{"Usage: "} + argv[0] + " <MODEL_DIR> \"<WAV_FILE_PATH>\"");
    }

    std::filesystem::path models_path = argv[1];
    std::string wav_file_path = argv[2];
    std::string device = "CPU";  // GPU or NPU can be used as well.

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

    ov::genai::WhisperGenerationConfig config(models_path / "generation_config.json");
    config.max_new_tokens = 100;
    config.language = "<|en|>";
    config.task = "transcribe";
    config.return_timestamps = true;

    // The pipeline expects normalized audio with a sampling rate of 16kHz.
    ov::genai::RawSpeechInput raw_speech = utils::audio::read_wav(wav_file_path);
    auto result = pipeline.generate(raw_speech, config);

    std::cout << result << "\n";

    for (auto& chunk : *result.chunks) {
        std::cout << "timestamps: [" << chunk.start_ts << ", " << chunk.end_ts << "] text: " << chunk.text << "\n";
    }

} catch (const std::exception& error) {
    try {
        std::cerr << error.what() << '\n';
    } catch (const std::ios_base::failure&) {
    }
    return EXIT_FAILURE;
} catch (...) {
    try {
        std::cerr << "Non-exception object thrown\n";
    } catch (const std::ios_base::failure&) {
    }
    return EXIT_FAILURE;
}

For more information, refer to the C++ sample.

Using GenAI in Chat Scenario

For chat scenarios where inputs and outputs represent a conversation, maintaining KVCache across inputs may prove beneficial. The start_chat and finish_chat chat-specific methods are used to mark a conversation session, as shown in the samples below:

Python

import openvino_genai


def streamer(subword):
    print(subword, end='', flush=True)
    return False


def infer(model_dir: str):
    device = 'CPU'  # GPU can be used as well.
    pipe = openvino_genai.LLMPipeline(model_dir, device)

    config = openvino_genai.GenerationConfig()
    config.max_new_tokens = 100

    pipe.start_chat()
    while True:
        try:
            prompt = input('question:\n')
        except EOFError:
            break
        pipe.generate(prompt, config, streamer)
        print('\n----------')
    pipe.finish_chat()

For more information, refer to the Python sample.

C++

#include "openvino/genai/llm_pipeline.hpp"

int main(int argc, char* argv[]) try {
    if (2 != argc) {
        throw std::runtime_error(std::string{"Usage: "} + argv[0] + " <MODEL_DIR>");
    }
    std::string prompt;
    std::string models_path = argv[1];

    std::string device = "CPU";  // GPU, NPU can be used as well
    ov::genai::LLMPipeline pipe(models_path, device);

    ov::genai::GenerationConfig config;
    config.max_new_tokens = 100;
    std::function<bool(std::string)> streamer = [](std::string word) {
        std::cout << word << std::flush;
        return false;
    };

    pipe.start_chat();
    std::cout << "question:\n";
    while (std::getline(std::cin, prompt)) {
        pipe.generate(prompt, config, streamer);
        std::cout << "\n----------\n"
            "question:\n";
    }
    pipe.finish_chat();
} catch (const std::exception& error) {
    try {
        std::cerr << error.what() << '\n';
    } catch (const std::ios_base::failure&) {}
    return EXIT_FAILURE;
} catch (...) {
    try {
        std::cerr << "Non-exception object thrown\n";
    } catch (const std::ios_base::failure&) {}
    return EXIT_FAILURE;
}

For more information, refer to the C++ sample

Using GenAI with Vision Language Models

OpenVINO GenAI introduces the openvino_genai.VLMPipeline pipeline for inference of multimodal text-generation Vision Language Models (VLMs). With a text prompt and an image as input, VLMPipeline can generate text using models such as LLava or MiniCPM-V. See the chat scenario presented in the samples below:

Python

import numpy as np
import openvino_genai
from PIL import Image
from openvino import Tensor
from pathlib import Path


def streamer(subword: str) -> bool:
    print(subword, end='', flush=True)


def read_image(path: str) -> Tensor:
    pic = Image.open(path).convert("RGB")
    image_data = np.array(pic.getdata()).reshape(1, pic.size[1], pic.size[0], 3).astype(np.uint8)
    return Tensor(image_data)


def read_images(path: str) -> list[Tensor]:
    entry = Path(path)
    if entry.is_dir():
        return [read_image(str(file)) for file in sorted(entry.iterdir())]
    return [read_image(path)]


def infer(model_dir: str, image_dir: str):
    rgbs = read_images(image_dir)
    device = 'CPU'  # GPU can be used as well.
    enable_compile_cache = dict()
    if "GPU" == device:
        enable_compile_cache["CACHE_DIR"] = "vlm_cache"
    pipe = openvino_genai.VLMPipeline(model_dir, device, **enable_compile_cache)

    config = openvino_genai.GenerationConfig()
    config.max_new_tokens = 100

    pipe.start_chat()
    prompt = input('question:\n')
    pipe.generate(prompt, images=rgbs, generation_config=config, streamer=streamer)

    while True:
        try:
            prompt = input("\n----------\n"
                "question:\n")
        except EOFError:
            break
        pipe.generate(prompt, generation_config=config, streamer=streamer)
    pipe.finish_chat()

For more information, refer to the Python sample.

C++

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

bool print_subword(std::string&& subword) {
    return !(std::cout << subword << std::flush);
}

int main(int argc, char* argv[]) try {
    if (3 != argc) {
        throw std::runtime_error(std::string{"Usage "} + argv[0] + " <MODEL_DIR> <IMAGE_FILE OR DIR_WITH_IMAGES>");
    }

    std::vector<ov::Tensor> rgbs = utils::load_images(argv[2]);

    std::string device = "CPU";  // GPU can be used as well.
    ov::AnyMap enable_compile_cache;
    if ("GPU" == device) {
        enable_compile_cache.insert({ov::cache_dir("vlm_cache")});
    }
    ov::genai::VLMPipeline pipe(argv[1], device, enable_compile_cache);

    ov::genai::GenerationConfig generation_config;
    generation_config.max_new_tokens = 100;

    std::string prompt;

    pipe.start_chat();
    std::cout << "question:\n";

    std::getline(std::cin, prompt);
    pipe.generate(prompt,
                  ov::genai::images(rgbs),
                  ov::genai::generation_config(generation_config),
                  ov::genai::streamer(print_subword));
    std::cout << "\n----------\n"
        "question:\n";
    while (std::getline(std::cin, prompt)) {
        pipe.generate(prompt,
                      ov::genai::generation_config(generation_config),
                      ov::genai::streamer(print_subword));
        std::cout << "\n----------\n"
            "question:\n";
    }
    pipe.finish_chat();
} catch (const std::exception& error) {
    try {
        std::cerr << error.what() << '\n';
    } catch (const std::ios_base::failure&) {}
    return EXIT_FAILURE;
} catch (...) {
    try {
        std::cerr << "Non-exception object thrown\n";
    } catch (const std::ios_base::failure&) {}
    return EXIT_FAILURE;
}

For more information, refer to the C++ sample

Chat-bot use case - step by step

This example will show you how to create a chat-bot functionality, using the ov_genai.LLMPipeline and a chat-tuned TinyLlama model. Apart from the basic implementation, it provides additional optimization methods.

Although CPU is used as inference device in the samples below, you may choose GPU instead. Note that tasks such as token selection, tokenization, and detokenization are always handled by CPU only. Tokenizers, represented as a separate model, are also run on CPU.

Running the model

You start with exporting an LLM model via Hugging Face Optimum-Intel. Note that the precision of int4 is used, instead of the original fp16, for better performance. The weight compression is done by NNCF at the model export stage. The exported model contains all the information necessary for execution, including the tokenizer/detokenizer and the generation config, ensuring that its results match those generated by Hugging Face.

The LLMPipeline is the main object to setup the model for text generation. You can provide the converted model to this object, specify the device for inference, and provide additional parameters.

Python

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
pipe = ov_genai.LLMPipeline(model_path, "CPU")
print(pipe.generate("The Sun is yellow because", max_new_tokens=100))

C++

optimum-cli export openvino --model "TinyLlama/TinyLlama-1.1B-Chat-v1.0" --weight-format int4 --trust-remote-code "TinyLlama-1.1B-Chat-v1.0"

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

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

Streaming the Output

For more interactive UIs during generation, you can stream output tokens. In this example, a lambda function outputs words to the console immediately upon generation:

Python

import openvino_genai as ov_genai
pipe = ov_genai.LLMPipeline(model_path, "CPU")

streamer = lambda x: print(x, end='', flush=True)
pipe.generate("The Sun is yellow because", streamer=streamer, max_new_tokens=100)

C++

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

int main(int argc, char* argv[]) {
   std::string model_path = argv[1];
   ov::genai::LLMPipeline pipe(model_path, "CPU");

   auto streamer = [](std::string word) {
      std::cout << word << std::flush;
      // Return flag indicating whether generation should be stopped.
      // false means continue generation.
      return false;
   };
   pipe.generate("The Sun is yellow because", ov::genai::streamer(streamer), ov::genai::max_new_tokens(100));
}

You can also create your custom streamer for more sophisticated processing:

Python

import openvino_genai as ov_genai

class CustomStreamer(ov_genai.StreamerBase):
   def __init__(self, tokenizer):
      ov_genai.StreamerBase.__init__(self)
      self.tokenizer = tokenizer
   def put(self, token_id) -> bool:
      # Decode tokens and process them.
      # Streamer returns a flag indicating whether generation should be stopped.
      # In Python, `return` can be omitted. In that case, the function will return None
      # which will be converted to False, meaning that generation should continue.
      # return stop_flag
   def end(self):
      # Decode tokens and process them.

pipe = ov_genai.LLMPipeline(model_path, "CPU")
pipe.generate("The Sun is yellow because", streamer=CustomStreamer(), max_new_tokens=100)

C++

#include <streamer_base.hpp>

class CustomStreamer: publict StreamerBase {
public:
   bool put(int64_t token) {
      bool stop_flag = false;
      /*
      custom decoding/tokens processing code
      tokens_cache.push_back(token);
      std::string text = m_tokenizer.decode(tokens_cache);
      ...
      */
      return stop_flag;  // Flag indicating whether generation should be stopped. If True, generation stops.
   };

   void end() {
      /* custom finalization */
   };
};

int main(int argc, char* argv[]) {
   auto custom_streamer = std::make_shared<CustomStreamer>();

   std::string model_path = argv[1];
   ov::genai::LLMPipeline pipe(model_path, "CPU");
   pipe.generate("The Sun is yellow because", ov::genai::streamer(custom_streamer), ov::genai::max_new_tokens(100));
}

Optimizing Generation with Grouped Beam Search

For better text generation quality and more efficient batch processing, specify generation_config to leverage grouped beam search decoding.

Python

import openvino_genai as ov_genai
pipe = ov_genai.LLMPipeline(model_path, "CPU")
config = pipe.get_generation_config()
config.max_new_tokens = 256
config.num_beam_groups = 3
config.num_beams = 15
config.diversity_penalty = 1.0
pipe.generate("The Sun is yellow because", config)

C++

int main(int argc, char* argv[]) {
   std::string model_path = argv[1];
   ov::genai::LLMPipeline pipe(model_path, "CPU");
   ov::genai::GenerationConfig config = pipe.get_generation_config();
   config.max_new_tokens = 256;
   config.num_beam_groups = 3;
   config.num_beams = 15;
   config.diversity_penalty = 1.0f;

   cout << pipe.generate("The Sun is yellow because", config);
}

Efficient Text Generation via Speculative Decoding

Speculative decoding (or assisted-generation) enables faster token generation when an additional smaller draft model is used alongside the main model. This reduces the number of infer requests to the main model, increasing performance.

The draft model predicts the next K tokens one by one in an autoregressive manner. The main model validates these predictions and corrects them if necessary - in case of a discrepancy, the main model prediction is used. Then, the draft model acquires this token and runs prediction of the next K tokens, thus repeating the cycle.

Python

import openvino_genai
import queue
import threading

def streamer(subword):
        print(subword, end='', flush=True)
        return False

def infer(model_dir: str, draft_model_dir: str, prompt: str):
    main_device = 'CPU'  # GPU can be used as well.
    draft_device = 'CPU'

    scheduler_config = openvino_genai.SchedulerConfig()
    scheduler_config.cache_size = 2

    draft_model = openvino_genai.draft_model(draft_model_dir, draft_device)

    pipe = openvino_genai.LLMPipeline(model_dir, main_device, scheduler_config=scheduler_config, draft_model=draft_model)

    config = openvino_genai.GenerationConfig()
    config.max_new_tokens = 100
    config.num_assistant_tokens = 5

    pipe.generate("The Sun is yellow because", config, streamer)

For more information, refer to the Python sample.

C++

#include <openvino/openvino.hpp>

#include "openvino/genai/llm_pipeline.hpp"

int main(int argc, char* argv[]) try {
    if (4 != argc) {
        throw std::runtime_error(std::string{"Usage: "} + argv[0] + " <MODEL_DIR> <DRAFT_MODEL_DIR> '<PROMPT>'");
    }

    ov::genai::GenerationConfig config;
    config.max_new_tokens = 100;
    config.num_assistant_tokens = 5;

    std::string main_model_path = argv[1];
    std::string draft_model_path = argv[2];
    std::string prompt = argv[3];

    std::string main_device = "CPU", draft_device = "CPU";

    ov::genai::SchedulerConfig scheduler_config;
    scheduler_config.cache_size = 5;

    ov::genai::LLMPipeline pipe(
        main_model_path,
        main_device,
        ov::genai::draft_model(draft_model_path, draft_device),
        ov::genai::scheduler_config(scheduler_config));

    auto streamer = [](std::string subword) {
        std::cout << subword << std::flush;
        return false;
    };

    pipe.generate("The Sun is yellow because", config, streamer);
} catch (const std::exception& error) {
    try {
        std::cerr << error.what() << '\n';
    } catch (const std::ios_base::failure&) {}
    return EXIT_FAILURE;
} catch (...) {
    try {
        std::cerr << "Non-exception object thrown\n";
    } catch (const std::ios_base::failure&) {}
    return EXIT_FAILURE;
}

For more information, refer to the C++ sample

Comparing with Hugging Face Results

You can compare the results of the above example with those generated by Hugging Face models by running the following code:

Python

from transformers import AutoTokenizer, AutoModelForCausalLM
import openvino_genai as ov_genai

tokenizer = AutoTokenizer.from_pretrained("TinyLlama/TinyLlama-1.1B-Chat-v1.0")
model = AutoModelForCausalLM.from_pretrained("TinyLlama/TinyLlama-1.1B-Chat-v1.0")

max_new_tokens = 32
prompt = 'table is made of'

encoded_prompt = tokenizer.encode(prompt, return_tensors='pt', add_special_tokens=False)
hf_encoded_output = model.generate(encoded_prompt, max_new_tokens=max_new_tokens, do_sample=False)
hf_output = tokenizer.decode(hf_encoded_output[0, encoded_prompt.shape[1]:])
print(f'hf_output: {hf_output}')

pipe = ov_genai.LLMPipeline('TinyLlama-1.1B-Chat-v1.0')
ov_output = pipe.generate(prompt, max_new_tokens=max_new_tokens)
print(f'ov_output: {ov_output}')

assert hf_output == ov_output

GenAI API

The use case described here regards the following OpenVINO GenAI API classes:

• generation_config - defines a configuration class for text generation, enabling customization of the generation process such as the maximum length of the generated text, whether to ignore end-of-sentence tokens, and the specifics of the decoding strategy (greedy, beam search, or multinomial sampling).

• llm_pipeline - provides classes and utilities for processing inputs, text generation, and managing outputs with configurable options.

• streamer_base - an abstract base class for creating streamers.

• tokenizer - the tokenizer class for text encoding and decoding.

Learn more from the GenAI API reference.

Additional Resources

• OpenVINO GenAI Repo

• OpenVINO GenAI Samples

• A Jupyter notebook demonstrating Visual-language assistant with MiniCPM-V2 and OpenVINO

• OpenVINO Tokenizers

• Neural Network Compression Framework