HuggingChatAllInOne

HuggingChatAllInOne copied to clipboard

HuggingChatAllInOne

I like HuggingChat UI, and their text generation infra structure, so wanted to build All in one docker file to run HuggingChat Frontend and Backend in one docker and on runpod.io and it should be fairly easily to run this single docker offline given the models already downloaded.

The project built in a way that it can be easily kept up-to-date with latest changes on the original repos:

• https://github.com/huggingface/text-generation-inference/
• https://github.com/huggingface/chat-ui

How it works

mostly the docker file is just the combination of the two dockerfile from huggingface repos, but one difficult part was the fact .env variable parameters, some of them, were being integrated at build time, this made me choose the easiset path, dynamically generate .env variable and re-build the website at run time. and I added the logic to apply some patches to original source code in some cases where its needed

In addition, Im checking if this is running under http://runpod.io, so I'll be setting the parameter PUBLIC_ORIGIN automatically (thankfully runpod.io providing env variable called RUNPOD_POD_ID)

To Build

Building the docker image will take really long time...

./UpdateAllSources.sh
./BuildAllInOneDockerFile.sh

• this will build up to the latest commit I've tested, you can modify the two script availble: GetLatestBackEndSources.sh, GetLatestFrontEnd.sh to choose different commit

to Run

with Patches, no quantization

docker run --gpus all -p 8080:8080 -v $PWD/Data:/data -e APPLY_PATCHES=1 -e PUBLIC_ORIGIN="http://exmaple.com:8080" -e MODEL_ID='TheBloke/Wizard-Vicuna-7B-Uncensored-HF' ghcr.io/bodaay/huggingchatallinone:latest

without Patches, no quantization

docker run --gpus all -p 8080:8080 -v $PWD/Data:/data -e PUBLIC_ORIGIN="http://localhost:8080" -e MODEL_ID='TheBloke/Wizard-Vicuna-7B-Uncensored-HF' ghcr.io/bodaay/huggingchatallinone:latest

without Patches, gptq

docker run --gpus all -p 8080:8080 -v $PWD/Data:/data -e PUBLIC_ORIGIN="http://localhost:8080" -e QUANTIZE='gptq' -e MODEL_ID='Ichigo2899/WIZVIC-7b-TGI-GPTQ' ghcr.io/bodaay/huggingchatallinone:latest

You need to set APPLY_PATCHES in the following cases for PUBLIC_ORIGIN:

• you are not running this in https://runpod.io, or a secure https connection
• you need to access this site using ip address or non secure domain name (not http://localhost:8080), ex: http://10.0.0.1:8080, http://example.com

Runpod Template

The most important parameter is: MODEL_ID this will the inference server based on that model

I've created two templates, just to make it easy to go between GPTQ and HF format, they are basically same, just different Env parameters

BoDaay - HuggingChatAllInOne - HF

• https://runpod.io/gsc?template=k8qitdzihe&ref=8s08lrw8

BoDaay - HuggingChatAllInOne - GTPQ

• https://runpod.io/gsc?template=2kkkk99tnd&ref=8s08lrw8

you can replace any of the parameters here and they will follow the default value, or you can use any other parameter from original repos of huggingface/chat-ui,huggingface/text-generation-inference

• Note that PUBLIC_ORIGIN will be automatically overwritten if you are running this under runpod.io

GPTQ Support

Only Model made with latest gptq-for-llama will work, in case of GPTQ, you need pass QUANTIZE parameter and set it to gptq

Environment Variables