Create an LLM-powered Chatbot using OpenVINO Generate API#

This Jupyter notebook can be launched after a local installation only.

Github

In the rapidly evolving world of artificial intelligence (AI), chatbots have emerged as powerful tools for businesses to enhance customer interactions and streamline operations. Large Language Models (LLMs) are artificial intelligence systems that can understand and generate human language. They use deep learning algorithms and massive amounts of data to learn the nuances of language and produce coherent and relevant responses. While a decent intent-based chatbot can answer basic, one-touch inquiries like order management, FAQs, and policy questions, LLM chatbots can tackle more complex, multi-touch questions. LLM enables chatbots to provide support in a conversational manner, similar to how humans do, through contextual memory. Leveraging the capabilities of Language Models, chatbots are becoming increasingly intelligent, capable of understanding and responding to human language with remarkable accuracy.

Previously, we already discussed how to build an instruction-following pipeline using OpenVINO, please check out this tutorial for reference. In this tutorial, we consider how to use the power of OpenVINO for running Large Language Models for chat. We will use a pre-trained model from the Hugging Face Transformers library. The Hugging Face Optimum Intel library converts the models to OpenVINO™ IR format. To simplify the user experience, we will use OpenVINO Generate API for generation pipeline.

The tutorial consists of the following steps:

Installation Instructions#

This is a self-contained example that relies solely on its own code.

We recommend running the notebook in a virtual environment. You only need a Jupyter server to start. For details, please refer to Installation Guide.

Table of contents:

Prerequisites#

Install required dependencies

import os

os.environ["GIT_CLONE_PROTECTION_ACTIVE"] = "false"

%pip install -Uq pip
%pip uninstall -q -y optimum optimum-intel
%pip install -q -U "openvino>=2024.3.0" openvino-tokenizers[transformers] openvino-genai
%pip install -q --extra-index-url https://download.pytorch.org/whl/cpu\
"git+https://github.com/huggingface/optimum-intel.git"\
"git+https://github.com/openvinotoolkit/nncf.git"\
"torch>=2.1"\
"datasets" \
"accelerate" \
"gradio>=4.19" \
"onnx<=1.16.1; sys_platform=='win32'" "einops" "transformers_stream_generator" "tiktoken" "bitsandbytes"

Note: you may need to restart the kernel to use updated packages.
Note: you may need to restart the kernel to use updated packages.
Note: you may need to restart the kernel to use updated packages.
Note: you may need to restart the kernel to use updated packages.

import os
from pathlib import Path
import requests
import shutil

# fetch model configuration

config_shared_path = Path("../../utils/llm_config.py")
config_dst_path = Path("llm_config.py")

if not config_dst_path.exists():
    if config_shared_path.exists():
        try:
            os.symlink(config_shared_path, config_dst_path)
        except Exception:
            shutil.copy(config_shared_path, config_dst_path)
    else:
        r = requests.get(url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/utils/llm_config.py")
        with open("llm_config.py", "w", encoding="utf-8") as f:
            f.write(r.text)
elif not os.path.islink(config_dst_path):
    print("LLM config will be updated")
    if config_shared_path.exists():
        shutil.copy(config_shared_path, config_dst_path)
    else:
        r = requests.get(url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/utils/llm_config.py")
        with open("llm_config.py", "w", encoding="utf-8") as f:
            f.write(r.text)

if not Path("genai_gradio_helper.py").exists():
    r = requests.get(url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/notebooks/llm-chatbot/genai_gradio_helper.py")
    open("gradio_helper.py", "w").write(r.text)

if not Path("notebook_utils.py").exists():
    r = requests.get(url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/utils/notebook_utils.py")
    open("notebook_utils.py", "w").write(r.text)

Select model for inference#

The tutorial supports different models, you can select one from the provided options to compare the quality of open source LLM solutions. Model conversion and optimization is time- and memory-consuming process. For your convenience, we provide a collection of optimized models on HuggingFace hub. You can skip the model conversion step by selecting one of the available on HuggingFace hub model. If you want to reproduce optimization process locally, please unset Use preconverted models checkbox.

Note: conversion of some models can require additional actions from user side and at least 64GB RAM for conversion.

Weight compression is a technique for enhancing the efficiency of models, especially those with large memory requirements. This method reduces the model’s memory footprint, a crucial factor for Large Language Models (LLMs). We provide several options for model weight compression:

  • FP16 reducing model binary size on disk using save_model with enabled compression weights to FP16 precision. This approach is available in OpenVINO from scratch and is the default behavior.

  • INT8 is an 8-bit weight-only quantization provided by NNCF: This method compresses weights to an 8-bit integer data type, which balances model size reduction and accuracy, making it a versatile option for a broad range of applications.

  • INT4 is an 4-bit weight-only quantization provided by NNCF. involves quantizing weights to an unsigned 4-bit integer symmetrically around a fixed zero point of eight (i.e., the midpoint between zero and 15). in case of symmetric quantization or asymmetrically with a non-fixed zero point, in case of asymmetric quantization respectively. Compared to INT8 compression, INT4 compression improves performance even more, but introduces a minor drop in prediction quality. INT4 it ideal for situations where speed is prioritized over an acceptable trade-off against accuracy.

  • INT4 AWQ is an 4-bit activation-aware weight quantization. Activation-aware Weight Quantization (AWQ) is an algorithm that tunes model weights for more accurate INT4 compression. It slightly improves generation quality of compressed LLMs, but requires significant additional time for tuning weights on a calibration dataset. We will use wikitext-2-raw-v1/train subset of the Wikitext dataset for calibration.

Click here to see available models options

  • tiny-llama-1b-chat - This is the chat model finetuned on top of TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T. The TinyLlama project aims to pretrain a 1.1B Llama model on 3 trillion tokens with the adoption of the same architecture and tokenizer as Llama 2. This means TinyLlama can be plugged and played in many open-source projects built upon Llama. Besides, TinyLlama is compact with only 1.1B parameters. This compactness allows it to cater to a multitude of applications demanding a restricted computation and memory footprint. More details about model can be found in model card

  • mini-cpm-2b-dpo - MiniCPM is an End-Size LLM developed by ModelBest Inc. and TsinghuaNLP, with only 2.4B parameters excluding embeddings. After Direct Preference Optimization (DPO) fine-tuning, MiniCPM outperforms many popular 7b, 13b and 70b models. More details can be found in model_card.

  • gemma-2b-it - Gemma is a family of lightweight, state-of-the-art open models from Google, built from the same research and technology used to create the Gemini models. They are text-to-text, decoder-only large language models, available in English, with open weights, pre-trained variants, and instruction-tuned variants. Gemma models are well-suited for a variety of text generation tasks, including question answering, summarization, and reasoning. This model is instruction-tuned version of 2B parameters model. More details about model can be found in model card. >Note: run model with demo, you will need to accept license agreement. >You must be a registered user in Hugging Face Hub. Please visit HuggingFace model card, carefully read terms of usage and click accept button. You will need to use an access token for the code below to run. For more information on access tokens, refer to this section of the documentation. >You can login on Hugging Face Hub in notebook environment, using following code:

# login to huggingfacehub to get access to pretrained model


from huggingface_hub import notebook_login, whoami

try:
    whoami()
    print('Authorization token already provided')
except OSError:
    notebook_login()

  • phi3-mini-instruct - The Phi-3-Mini is a 3.8B parameters, lightweight, state-of-the-art open model trained with the Phi-3 datasets that includes both synthetic data and the filtered publicly available websites data with a focus on high-quality and reasoning dense properties. More details about model can be found in model card, Microsoft blog and technical report.

  • red-pajama-3b-chat - A 2.8B parameter pre-trained language model based on GPT-NEOX architecture. It was developed by Together Computer and leaders from the open-source AI community. The model is fine-tuned on OASST1 and Dolly2 datasets to enhance chatting ability. More details about model can be found in HuggingFace model card.

  • gemma-7b-it - Gemma is a family of lightweight, state-of-the-art open models from Google, built from the same research and technology used to create the Gemini models. They are text-to-text, decoder-only large language models, available in English, with open weights, pre-trained variants, and instruction-tuned variants. Gemma models are well-suited for a variety of text generation tasks, including question answering, summarization, and reasoning. This model is instruction-tuned version of 7B parameters model. More details about model can be found in model card. >Note: run model with demo, you will need to accept license agreement. >You must be a registered user in Hugging Face Hub. Please visit HuggingFace model card, carefully read terms of usage and click accept button. You will need to use an access token for the code below to run. For more information on access tokens, refer to this section of the documentation. >You can login on Hugging Face Hub in notebook environment, using following code:

# login to huggingfacehub to get access to pretrained model

from huggingface_hub import notebook_login, whoami

try:
    whoami()
    print('Authorization token already provided')
except OSError:
    notebook_login()

  • llama-2-7b-chat - LLama 2 is the second generation of LLama models developed by Meta. Llama 2 is a collection of pre-trained and fine-tuned generative text models ranging in scale from 7 billion to 70 billion parameters. llama-2-7b-chat is 7 billions parameters version of LLama 2 finetuned and optimized for dialogue use case. More details about model can be found in the paper, repository and HuggingFace model card. >Note: run model with demo, you will need to accept license agreement. >You must be a registered user in Hugging Face Hub. Please visit HuggingFace model card, carefully read terms of usage and click accept button. You will need to use an access token for the code below to run. For more information on access tokens, refer to this section of the documentation. >You can login on Hugging Face Hub in notebook environment, using following code:

# login to huggingfacehub to get access to pretrained model

from huggingface_hub import notebook_login, whoami

try:
    whoami()
    print('Authorization token already provided')
except OSError:
    notebook_login()

  • llama-3-8b-instruct - Llama 3 is an auto-regressive language model that uses an optimized transformer architecture. The tuned versions use supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) to align with human preferences for helpfulness and safety. The Llama 3 instruction tuned models are optimized for dialogue use cases and outperform many of the available open source chat models on common industry benchmarks. More details about model can be found in Meta blog post, model website and model card. >Note: run model with demo, you will need to accept license agreement. >You must be a registered user in Hugging Face Hub. Please visit HuggingFace model card, carefully read terms of usage and click accept button. You will need to use an access token for the code below to run. For more information on access tokens, refer to this section of the documentation. >You can login on Hugging Face Hub in notebook environment, using following code:

# login to huggingfacehub to get access to pretrained model

from huggingface_hub import notebook_login, whoami

try:
    whoami()
    print('Authorization token already provided')
except OSError:
    notebook_login()

  • llama-3.1-8b-instruct - The Llama 3.1 instruction tuned text only models (8B, 70B, 405B) are optimized for multilingual dialogue use cases and outperform many of the available open source and closed chat models on common industry benchmarks. More details about model can be found in Meta blog post, model website and model card. >Note: run model with demo, you will need to accept license agreement. >You must be a registered user in Hugging Face Hub. Please visit HuggingFace model card, carefully read terms of usage and click accept button. You will need to use an access token for the code below to run. For more information on access tokens, refer to this section of the documentation. >You can login on Hugging Face Hub in notebook environment, using following code:

# login to huggingfacehub to get access to pretrained model

from huggingface_hub import notebook_login, whoami

try:
    whoami()
    print('Authorization token already provided')
except OSError:
    notebook_login()

  • qwen2-1.5b-instruct/qwen2-7b-instruct - Qwen2 is the new series of Qwen large language models.Compared with the state-of-the-art open source language models, including the previous released Qwen1.5, Qwen2 has generally surpassed most open source models and demonstrated competitiveness against proprietary models across a series of benchmarks targeting for language understanding, language generation, multilingual capability, coding, mathematics, reasoning, etc. For more details, please refer to model_card, blog, GitHub, and Documentation.

  • qwen1.5-0.5b-chat/qwen1.5-1.8b-chat/qwen1.5-7b-chat - Qwen1.5 is the beta version of Qwen2, a transformer-based decoder-only language model pretrained on a large amount of data. Qwen1.5 is a language model series including decoder language models of different model sizes. It is based on the Transformer architecture with SwiGLU activation, attention QKV bias, group query attention, mixture of sliding window attention and full attention. You can find more details about model in the model repository.

  • qwen-7b-chat - Qwen-7B is the 7B-parameter version of the large language model series, Qwen (abbr. Tongyi Qianwen), proposed by Alibaba Cloud. Qwen-7B is a Transformer-based large language model, which is pretrained on a large volume of data, including web texts, books, codes, etc. For more details about Qwen, please refer to the GitHub code repository.

  • chatglm3-6b - ChatGLM3-6B is the latest open-source model in the ChatGLM series. While retaining many excellent features such as smooth dialogue and low deployment threshold from the previous two generations, ChatGLM3-6B employs a more diverse training dataset, more sufficient training steps, and a more reasonable training strategy. ChatGLM3-6B adopts a newly designed Prompt format, in addition to the normal multi-turn dialogue. You can find more details about model in the model card

  • mistral-7b - The Mistral-7B-v0.1 Large Language Model (LLM) is a pretrained generative text model with 7 billion parameters. You can find more details about model in the model card, paper and release blog post.

  • zephyr-7b-beta - Zephyr is a series of language models that are trained to act as helpful assistants. Zephyr-7B-beta is the second model in the series, and is a fine-tuned version of mistralai/Mistral-7B-v0.1 that was trained on on a mix of publicly available, synthetic datasets using Direct Preference Optimization (DPO). You can find more details about model in technical report and HuggingFace model card.

  • neural-chat-7b-v3-1 - Mistral-7b model fine-tuned using Intel Gaudi. The model fine-tuned on the open source dataset Open-Orca/SlimOrca and aligned with Direct Preference Optimization (DPO) algorithm. More details can be found in model card and blog post.

  • notus-7b-v1 - Notus is a collection of fine-tuned models using Direct Preference Optimization (DPO). and related RLHF techniques. This model is the first version, fine-tuned with DPO over zephyr-7b-sft. Following a data-first approach, the only difference between Notus-7B-v1 and Zephyr-7B-beta is the preference dataset used for dDPO. Proposed approach for dataset creation helps to effectively fine-tune Notus-7b that surpasses Zephyr-7B-beta and Claude 2 on AlpacaEval. More details about model can be found in model card.

  • youri-7b-chat - Youri-7b-chat is a Llama2 based model. Rinna Co., Ltd. conducted further pre-training for the Llama2 model with a mixture of English and Japanese datasets to improve Japanese task capability. The model is publicly released on Hugging Face hub. You can find detailed information at the rinna/youri-7b-chat project page.

  • baichuan2-7b-chat - Baichuan 2 is the new generation of large-scale open-source language models launched by Baichuan Intelligence inc. It is trained on a high-quality corpus with 2.6 trillion tokens and has achieved the best performance in authoritative Chinese and English benchmarks of the same size.

  • internlm2-chat-1.8b - InternLM2 is the second generation InternLM series. Compared to the previous generation model, it shows significant improvements in various capabilities, including reasoning, mathematics, and coding. More details about model can be found in model repository.

  • glm-4-9b-chat - GLM-4-9B is the open-source version of the latest generation of pre-trained models in the GLM-4 series launched by Zhipu AI. In the evaluation of data sets in semantics, mathematics, reasoning, code, and knowledge, GLM-4-9B and its human preference-aligned version GLM-4-9B-Chat have shown superior performance beyond Llama-3-8B. In addition to multi-round conversations, GLM-4-9B-Chat also has advanced features such as web browsing, code execution, custom tool calls (Function Call), and long text reasoning (supporting up to 128K context). More details about model can be found in model card, technical report and repository

from llm_config import get_llm_selection_widget

form, lang, model_id_widget, compression_variant, use_preconverted = get_llm_selection_widget()

form

Box(children=(Box(children=(Label(value='Language:'), Dropdown(options=('English', 'Chinese', 'Japanese'), val…

model_configuration = model_id_widget.value
model_id = model_id_widget.label
print(f"Selected model {model_id} with {compression_variant.value} compression")

Selected model qwen2-0.5b-instruct with INT4 compression

Convert model using Optimum-CLI tool#

Optimum Intel is the interface between the Transformers and Diffusers libraries and OpenVINO to accelerate end-to-end pipelines on Intel architectures. It provides ease-to-use cli interface for exporting models to OpenVINO Intermediate Representation (IR) format.

Click here to read more about Optimum CLI usage

The command bellow demonstrates basic command for model export with optimum-cli

optimum-cli export openvino --model <model_id_or_path> --task <task> <out_dir>

where --model argument is model id from HuggingFace Hub or local directory with model (saved using .save_pretrained method), --task is one of supported task that exported model should solve. For LLMs it is recommended to use text-generation-with-past. If model initialization requires to use remote code, --trust-remote-code flag additionally should be passed.

Weights Compression using Optimum-CLI#

You can also apply fp16, 8-bit or 4-bit weight compression on the Linear, Convolutional and Embedding layers when exporting your model with the CLI.

Click here to read more about weights compression with Optimum CLI

Setting --weight-format to respectively fp16, int8 or int4. This type of optimization allows to reduce the memory footprint and inference latency. By default the quantization scheme for int8/int4 will be asymmetric, to make it symmetric you can add --sym.

For INT4 quantization you can also specify the following arguments : - The --group-size parameter will define the group size to use for quantization, -1 it will results in per-column quantization. - The --ratio parameter controls the ratio between 4-bit and 8-bit quantization. If set to 0.9, it means that 90% of the layers will be quantized to int4 while 10% will be quantized to int8.

Smaller group_size and ratio values usually improve accuracy at the sacrifice of the model size and inference latency. You can enable AWQ to be additionally applied during model export with INT4 precision using --awq flag and providing dataset name with --datasetparameter (e.g. --dataset wikitext2)

Note: Applying AWQ requires significant memory and time.

Note: It is possible that there will be no matching patterns in the model to apply AWQ, in such case it will be skipped.

from llm_config import convert_and_compress_model

model_dir = convert_and_compress_model(model_id, model_configuration, compression_variant.value, use_preconverted.value)

⌛ qwen2-0.5b-instruct conversion to INT4 started. It may takes some time.

Export command:

optimum-cli export openvino --model Qwen/Qwen2-0.5B-Instruct --task text-generation-with-past --weight-format int4 --group-size 128 --ratio 0.8 qwen2/INT4_compressed_weights

2024-08-28 02:56:11.289127: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable TF_ENABLE_ONEDNN_OPTS=0.
2024-08-28 02:56:11.322171: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
2024-08-28 02:56:11.841318: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
Framework not specified. Using pt to export the model.
Using framework PyTorch: 2.2.2+cpu
Overriding 1 configuration item(s)
    - use_cache -> True
We detected that you are passing past_key_values as a tuple and this is deprecated and will be removed in v4.43. Please use an appropriate Cache class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)
Mixed-Precision assignment ━━━━━━━━━━━━━━━━━━━━ 100% 168/168 • 0:00:03 • 0:00:00
INFO:nncf:Statistics of the bitwidth distribution:
┍━━━━━━━━━━━━━━━━┯━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┯━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┑
│   Num bits (N) │ % all parameters (layers)   │ % ratio-defining parameters (layers)   │
┝━━━━━━━━━━━━━━━━┿━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┿━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┥
│              8 │ 43% (81 / 169)              │ 21% (80 / 168)                         │
├────────────────┼─────────────────────────────┼────────────────────────���───────────────┤
│              4 │ 57% (88 / 169)              │ 79% (88 / 168)                         │
┕━━━━━━━━━━━━━━━━┷━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┷━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┙
Applying Weight Compression ━━━━━━━━━━━━━━━━━━━━━━━━━━━ 100% • 0:00:10 • 0:00:00

/opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-761/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/optimum/exporters/openvino/model_patcher.py:489: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if sequence_length != 1:
/opt/home/k8sworker/ci-ai/cibuilds/ov-notebook/OVNotebookOps-761/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/transformers/models/qwen2/modeling_qwen2.py:110: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if seq_len > self.max_seq_len_cached:
Set tokenizer padding side to left for text-generation-with-past task.
Replacing (?!S) pattern to (?:$|[^S]) in RegexSplit operation
✅ INT4 qwen2-0.5b-instruct model converted and can be found in qwen2/INT4_compressed_weights

Let’s compare model size for different compression types

from llm_config import compare_model_size

compare_model_size(model_dir)

Size of model with INT4 compressed weights is 358.80 MB

Select device for inference#

from notebook_utils import device_widget

device = device_widget(default="CPU", exclude=["NPU"])

device

Dropdown(description='Device:', options=('CPU', 'AUTO'), value='CPU')

The cell below demonstrates how to instantiate model based on selected variant of model weights and inference device

Instantiate pipeline with OpenVINO Generate API#

OpenVINO Generate API can be used to create pipelines to run an inference with OpenVINO Runtime.

Firstly we need to create a pipeline with LLMPipeline. LLMPipeline is the main object used for text generation using LLM in OpenVINO GenAI API. You can construct it straight away from the folder with the converted model. We will provide directory with model and device for LLMPipeline. Then we run generate method and get the output in text format. Additionally, we can configure parameters for decoding. We can get the default config with get_generation_config(), setup parameters, and apply the updated version with set_generation_config(config) or put config directly to generate(). It’s also possible to specify the needed options just as inputs in the generate() method, as shown below, e.g. we can add max_new_tokens to stop generation if a specified number of tokens is generated and the end of generation is not reached. We will discuss some of the available generation parameters more deeply later.

from openvino_genai import LLMPipeline

print(f"Loading model from {model_dir}\n")


pipe = LLMPipeline(str(model_dir), device.value)

generation_config = pipe.get_generation_config()

input_prompt = "The Sun is yellow bacause"
print(f"Input text: {input_prompt}")
print(pipe.generate(input_prompt, max_new_tokens=10))

Loading model from qwen2/INT4_compressed_weights

Input text: The Sun is yellow bacause
 it is made of hydrogen and oxygen atoms. The

Run Chatbot#

Now, when model created, we can setup Chatbot interface using Gradio.

Click here to see how pipeline works

The diagram below illustrates how the chatbot pipeline works

llm_diagram

llm_diagram#

As you can see, user input question passed via tokenizer to apply chat-specific formatting (chat template) and turn the provided string into the numeric format. OpenVINO Tokenizers are used for these purposes inside LLMPipeline. You can find more detailed info about tokenization theory and OpenVINO Tokenizers in this tutorial. Then tokenized input passed to LLM for making prediction of next token probability. The way the next token will be selected over predicted probabilities is driven by the selected decoding methodology. You can find more information about the most popular decoding methods in this blog. The sampler’s goal is to select the next token id is driven by generation configuration. Next, we apply stop generation condition to check the generation is finished or not (e.g. if we reached the maximum new generated tokens or the next token id equals to end of the generation). If the end of the generation is not reached, then new generated token id is used as the next iteration input, and the generation cycle repeats until the condition is not met. When stop generation criteria are met, then OpenVINO Detokenizer decodes generated token ids to text answer.

The difference between chatbot and instruction-following pipelines is that the model should have “memory” to find correct answers on the chain of connected questions. OpenVINO GenAI uses KVCache representation for maintain a history of conversation. By default, LLMPipeline resets KVCache after each generate call. To keep conversational history, we should move LLMPipeline to chat mode using start_chat() method.

More info about OpenVINO LLM inference can be found in LLM Inference Guide

Advanced generation options#

Click here to see detailed description of advanced options

There are several parameters that can control text generation quality, * Temperature is a parameter used to control the level of creativity in AI-generated text. By adjusting the temperature, you can influence the AI model’s probability distribution, making the text more focused or diverse.
Consider the following example: The AI model has to complete the sentence “The cat is ____.” with the following token probabilities:
playing: 0.5
sleeping: 0.25
eating: 0.15
driving: 0.05
flying: 0.05

  • Low temperature (e.g., 0.2): The AI model becomes more focused and deterministic, choosing tokens with the highest probability, such as “playing.”

    • Medium temperature (e.g., 1.0): The AI model maintains a balance between creativity and focus, selecting tokens based on their probabilities without significant bias, such as “playing,” “sleeping,” or “eating.”

    • High temperature (e.g., 2.0): The AI model becomes more adventurous, increasing the chances of selecting less likely tokens, such as “driving” and “flying.”

  • Top-p, also known as nucleus sampling, is a parameter used to control the range of tokens considered by the AI model based on their cumulative probability. By adjusting the top-p value, you can influence the AI model’s token selection, making it more focused or diverse. Using the same example with the cat, consider the following top_p settings:

    • Low top_p (e.g., 0.5): The AI model considers only tokens with the highest cumulative probability, such as “playing.”

    • Medium top_p (e.g., 0.8): The AI model considers tokens with a higher cumulative probability, such as “playing,” “sleeping,” and “eating.”

    • High top_p (e.g., 1.0): The AI model considers all tokens, including those with lower probabilities, such as “driving” and “flying.”

  • Top-k is an another popular sampling strategy. In comparison with Top-P, which chooses from the smallest possible set of words whose cumulative probability exceeds the probability P, in Top-K sampling K most likely next words are filtered and the probability mass is redistributed among only those K next words. In our example with cat, if k=3, then only “playing”, “sleeping” and “eating” will be taken into account as possible next word.

  • Repetition Penalty This parameter can help penalize tokens based on how frequently they occur in the text, including the input prompt. A token that has already appeared five times is penalized more heavily than a token that has appeared only one time. A value of 1 means that there is no penalty and values larger than 1 discourage repeated tokens.

from genai_gradio_helper import get_gradio_helper

demo = get_gradio_helper(pipe, model_configuration, model_id, lang.value)

try:
    demo.launch(debug=False)
except Exception:
    demo.launch(debug=False, share=True)
# if you are launching remotely, specify server_name and server_port
# demo.launch(server_name='your server name', server_port='server port in int')
# Read more in the docs: https://gradio.app/docs/

Running on local URL:  http://127.0.0.1:7860

To create a public link, set share=True in launch().