Nerdy Rodent
GPT-J-6B – Just like GPT-3 but you can actually download the weights and run it at home. No API sign-up required, unlike some other models we could mention, right? Any just like GPT-3, this has 3 super easy example ways you can run this -v ia the EleutherAI website, Google Colab TPU or single GPU on a local system.
Github: https://github.com/kingoflolz/mesh-transformer-jax
Requirements for this guide to running locally:
* Ubuntu 20.04
* 32 GB RAM and an Nvidia GPU with at least 24 GB VRAM
* Anaconda – https://www.anaconda.com/products/individual
== Python virtual environment (Anaconda) ==
conda create –name mesh-transformer-jax python=3.8
conda activate mesh-transformer-jax
== Download & install ==
git clone https://github.com/kingoflolz/mesh-transformer-jax.git
cd mesh-transformer-jax
sudo apt install zstd
wget https://the-eye.eu/public/AI/GPT-J-6B/step_383500_slim.tar.zstd
tar -I zstd -xf step_383500_slim.tar.zstd
pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
pip install jaxlib==0.1.67+cuda111 -f https://storage.googleapis.com/jax-releases/jax_releases.html
pip install -r requirements.txt # (Edited to use tensorflow-gpu 2.5.0 for no reason)
pip install jax==0.2.12