Chat with Meta's LLaMA models at home made easy

This repository is a chat example with LLaMA (arXiv) models running on a typical home PC. You will just need a NVIDIA videocard and some RAM to chat with model.

Conda Environment Setup Example for Windows 10+

Download and install Anaconda Python https://www.anaconda.com and run Anaconda Prompt

conda create -n llama python=3.10
conda activate llama
conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia

Setup

In a conda env with pytorch / cuda available, run

pip install -r requirements.txt

Then in this repository

pip install -e .

Download tokenizer and models

magnet:?xt=urn:btih:ZXXDAUWYLRUXXBHUYEMS6Q5CE5WA3LVA&dn=LLaMA

or

magnet:xt=urn:btih:b8287ebfa04f879b048d4d4404108cf3e8014352&dn=LLaMA&tr=udp%3a%2f%2ftracker.opentrackr.org%3a1337%2fannounce

Prepare model

First, you need to unshard model checkpoints to a single file. Let's do this for 30B model.

python merge_weights.py --input_dir D:\Downloads\LLaMA --model_size 30B

In this example, D:\Downloads\LLaMA is a root folder of downloaded torrent with weights.

This will create merged.pth file in the root folder of this repo.

Place this file and corresponding (torrentroot)/30B/params.json of model into [/model] folder.

Place (torrentroot)/tokenizer.model file to the [/tokenizer] folder of this repo. Now you are ready to go.

python example-chat.py ./model ./tokenizer/tokenizer.model