aihandlerwindows 1.8.18

AI Handler: An engine which wraps certain huggingface models

AI Handler

This is a simple framework for running AI models. It makes use of the huggingface API which gives you a queue, threading, a simple API, and the ability to run Stable Diffusion and LLMs seamlessly from your local hardware.

This is not intended to be used as a standalone application.

It can easily be extended and used to power interfaces or it can be run from the command line.

AI Handler is a work in progress. It powers two projects at the moment, but may not be ready for general use.

Installation

This is a work in progress.

Pre-requisites

System requirements

• Windows 10+
• Python 3.10.8
• pip 23.0.1
• CUDA toolkit 11.7
• CUDNN 8.6.0.163
• Cuda capable GPU
• 16gb+ ram

Install

pip install torch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 --index-url https://download.pytorch.org/whl/cu117
pip install https://github.com/w4ffl35/diffusers/archive/refs/tags/v0.14.0.ckpt_fix.tar.gz
pip install https://github.com/w4ffl35/transformers/archive/refs/tags/tensor_fix-v1.0.2.tar.gz
pip install https://github.com/acpopescu/bitsandbytes/releases/download/v0.37.2-win.0/bitsandbytes-0.37.2-py3-none-any.whl
pip install aihandlerwindows

Optional

These are optional instructions for installing TensorRT and Deepspeed for Windows

Install Tensor RT:

1. Download TensorRT-8.4.3.1.Windows10.x86_64.cuda-11.6.cudnn8.4
2. Git clone TensorRT 8.4.3.1
3. Follow their instructions to build TensorRT-8.4.3.1 python wheel
4. Install TensorRT pip install tensorrt-*.whl

Install Deepspeed:

1. Git clone Deepspeed 0.8.1
2. Follow their instructions to build Deepspeed python wheel
3. Install Deepspeed `pip install deepspeed-*.whl

---

Environment variables

• AIRUNNER_ENVIRONMENT - dev or prod. Defaults to dev. This controls the LOG_LEVEL
• LOG_LEVEL - FATAL for production, DEBUG for development. Override this to force a log level

Huggingface variables

Offline mode

These environment variables keep you offline until you need to download a model. This prevents unwanted online access and speeds up usage of huggingface libraries.

• DISABLE_TELEMETRY Keep this set to 1 at all times. Huggingface collects minimal telemetry when downloading a model from their repository but this will keep it disabled. See more info in this github thread
• HF_HUB_OFFLINE When loading a diffusers model, huggingface libraries will attempt to download an updated cache before running the model. This prevents that check from happening (long with a boolean passed to load_pretrained see the runner.py file for examples)
• TRANSFORMERS_OFFLINE Similar to HF_HUB_OFFLINE but for transformers models