gai-aio 0.7

Gai/Gen is the Universal Multi-Modal Wrapper Library for LLM. The library is designed to provide a simplified and unified interface for seamless switching between multi-modal open source language models on a local machine and OpenAI APIs.

Gai/Gen: Universal LLM Wrapper

This is a Universal Multi-Modal Wrapper Library for LLM inferencing.

This is a Universal Multi-Modal Wrapper Library for LLM inferencing. The library provides a simplified and unified interface for seamless switching between multi-modal open source language models on a local machine and OpenAI APIs, adhering closely to the OpenAI API specs.

This is intended for Developers who are targetting the use of multi-modal LLMs for both OpenAI API and local machine models.

As the main focus is on 7 billion parameters and below open source models, running on commodity hardware, the library is designed to be a singleton wrapper. Only one model is loaded and cached into memory at any one time.

The wrappers are organised under the gen folder according to 4 categories:

• ttt: Text-to-Text
• tts: Text-to-Speech
• stt: Speech-to-Text
• itt: Image-to-Text

---

Setting Up

pip install gai-aio

or

git clone https://www.github.com/kakkoii1337/gai-aio

Configuration

• Create the default application directory ~/gai and the default models directory ~/gai/models.

• Copy gai.json from this repository into ~/gai. This file contains the names and paths of the models and their respective loaders.

API Key

• All API keys should be stored in a .env file in the root directory of the project. For example,

  OPENAI_API_KEY=<--replace-with-your-api-key-->
  ANTHROPIC_API_KEY=<--replace-with-your-api-key-->
  

Requirements

• The instructions are tested mainly on:
  • Windows 11 with WSL
  • Ubuntu 20.04.2 LTS
  • NVIDIA RTX 2060 GPU with 8GB VRAM
  • CUDA Toolkit is required for the GPU accelerated models. Run nvidia-smi to check if CUDA is installed. If you need help, refer to this https://gist.github.com/kakkoii1337/8a8d4d0bc71fa9c099a683d1601f219e

Credits

This library is made possible by the generosity and hardwork of the following open source projects. You are highly encouraged to check out the original source and documentations.

TTT

• TheBloke for all the quantized models in the demo
• turboderp for ExLlama
• Meta Team for the LLaMa2 Model
• HuggingFace team for the Transformers library and open source models
• [Mistral AI Team for Mistral7B Model
• Georgi Gerganov for LLaMaCpp

ITT

• Liu HaoTian for the LLaVa Model and Library

TTS

• Coqui-AI for the xTTS Model

STT

• OpenAI for Open Sourcing Whisper v3

---

Quick Start

1. Setup OpenAI API Key.

Save your OpenAI API key in a .env file in the root directory of your project.

OPENAI_API_KEY=<--replace-with-your-api-key-->

2. Run GPT4.

Run Text-to-Text generation using OpenAI by loading gpt-4 wrapper.

from gai.gen import Gaigen
gen = Gaigen.GetInstance().load('gpt-4')

response = gen.create(messages=[{'role':'USER','content':'Tell me a one paragraph short story.'},{'role':'ASSISTANT','content':''}])
print(response)

3. Install Mistral7B.

Open ~/gai/gai.json and refer to model_path:

• under mistral-8k, refer to model-path for the repo name.
• Go to huggingface
• Download the repo Mistral-7B-Instruct-v0.1-GPTQ into the ~/gai/models folder.

4. Run Mistral7B.

Run Text-to-Text generation using Mistral7B by replacing gpt-4 with mistral-8k.

from gai.gen import Gaigen
gen = Gaigen.GetInstance().load('mistral-8k')

response = gen.create(messages=[{'role':'USER','content':'Tell me a one paragraph short story.'},{'role':'ASSISTANT','content':''}])
print(response)

Examples

• Text-to-Text Generation
• Speech-to-Text Generation
• Text-to-Speech Generation
• Image-to-Text Generation