link-prediction

Follow these steps on the development server (not your local machine).

Setup

Clone this repo with git clone --recurse-submodules <URI>.

1. Follow instructions to download llama models

• NOTE: The repo is already added as a submodule, so no need to clone separately.
• Instructions per https://github.com/facebookresearch/llama
  • https://ai.meta.com/resources/models-and-libraries/llama-downloads/

2. Install miniconda

mkdir -p ~/miniconda3

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.sh

bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3

rm -rf ~/miniconda3/miniconda.sh

3. Create the conda virtual environment

conda env create -f environment.yml

conda activate link-prediction

• This will install all dependencies required by this project via pip (including the llama repo as an editable package).

Run

torchrun --nproc_per_node 2 main.py --path_to_cfg ./experiment_config.json --ckpt_dir /home/parklab/models/llama2/llama-2-13b-chat/ --tokenizer_path /home/parklab/models/llama2/tokenizer.model

Miscellaneous

llama submodule

The Llama repo was added as a submodule via the following. For now, we will use a git submodule to version control the llama editable pip package. git submodule add https://github.com/facebookresearch/llama

git config -f .gitmodules submodule.llama.ignore dirty # ignore dirty commits in the submodule

git config -f .gitmodules submodule.llama.branch main

If you forgot to clone with --recurse-submodules, you'll need to run the following command:

git submodule update --init --recursive

Updating Llama submodule

https://stackoverflow.com/questions/1777854/how-can-i-specify-a-branch-tag-when-adding-a-git-submodule/18799234

git submodule foreach -q --recursive \
  'git switch \
  $(git config -f $toplevel/.gitmodules submodule.$name.branch || echo main)'

git submodule update --recursive --remote

Model directory

The model files should be downloaded to a shared directory so the single copy can be used by multiple users. You will point the llama APIs to the model in this directory.

requirements.txt

The pytorch packages were chosen based on https://pytorch.org/get-started/locally/ and selecting:

• Stable
• Linux
• Pip
• Python
• Cuda 12.1