Skip to content

Latest commit

 

History

History
204 lines (139 loc) · 7.35 KB

README_old.md

File metadata and controls

204 lines (139 loc) · 7.35 KB
Raw

Perez Astra Fellowship Universal & Transferable VLM Jailbreaks

Runpod IO Setup

apt-get update && apt-get install vim -y && apt-get install nano -y && apt-get install tmux
cd /workspace && wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh && chmod +x Miniconda3-latest-Linux-x86_64.sh && ./Miniconda3-latest-Linux-x86_64.sh

Create SSH key and register it with GitHub:

ssh-keygen -t ed25519 -C "[email protected]"
mkdir /workspace/.ssh && cp /root/.ssh/* /workspace/.ssh/ && ls /workspace/.ssh/
eval "$(ssh-agent -s)"
ssh-add /workspace/.ssh/id_ed25519
cat ~/.ssh/id_ed25519.pub

Then go to GitHub and add the SSH key to your account. Once you're setup, clone the repo:

git clone [email protected]:RylanSchaeffer/PerezAstraFellowship-Universal-VLM-Jailbreak.git

Then modify your .bashrc to add:

conda activate universal_vlm_jailbreak_env
cd PerezAstraFellowship-Universal-VLM-Jailbreak
  1. Make SECRETS"
  2. export LFS_HOME=/workspace

Setup Method 1 - install from env file

  • We have a conda environment file to pin package versions and make setup faster.
  1. (Optional) Update conda:

conda update -n base -c defaults conda -y

  1. Create a conda environment from the env file and activate it (will take a while to build the env):

conda env create -f environment.yml -n vlm -y && conda activate vlm

  1. Grab the git submodules:

git submodule update --init --recursive

  1. Install Prismatic and Deepseek

Adding --config-settings editable_mode=compat is optionable - its for your vscode language to recognize the packages cd submodules/prismatic-vlms && pip install -e . --config-settings editable_mode=compat && cd ../.. cd submodules/DeepSeek-VL && pip install -e . --config-settings editable_mode=compat && cd ../..

  1. Make sure to log in to W&B by running wandb login
  2. Login to Huggingface with huggingface-cli login

Setup Method 2 - Manual

  1. (Optional) Update conda:

conda update -n base -c defaults conda -y

  1. Create and activate the conda environment:

conda create -n universal_vlm_jailbreak_env python=3.11 -y && conda activate universal_vlm_jailbreak_env

  1. Update pip in preparation of the next step.

pip install --upgrade pip

  1. Install Pytorch:

conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia -y

  1. Install Lightning:

conda install lightning -c conda-forge -y

  1. Grab the git submodules:

git submodule update --init --recursive

  1. Install Prismatic and Deepseek

Adding --config-settings editable_mode=compat is optionable - its for your vscode language to recognize the packages cd submodules/prismatic-vlms && pip install -e . --config-settings editable_mode=compat && cd ../.. cd submodules/DeepSeek-VL && pip install -e . --config-settings editable_mode=compat && cd ../..

  1. Then follow their instructions:

pip install packaging ninja && pip install flash-attn --no-build-isolation

  1. Manually install a few additional packages:

conda install joblib pandas matplotlib seaborn black tiktoken -y

10a. For running salesforce's xgen

# Needs the latest dev version("4.41.0.dev0") 
pip uninstall -y transformers && pip install git+https://github.com/huggingface/pip
pip install open_clip_torch==2.24.0
pip install einops
pip install einops-exts
  1. Install more stuff.

pip install nvidia-htop sentencepiece hf_transfer anthropic termcolor

  1. Make sure to log in to W&B by running wandb login
  2. Login to Huggingface with huggingface-cli login

Note: To run on a CPU-only machine (e.g., for eval), use conda install pytorch torchvision torchaudio cpuonly -c pytorch

Note: You might need to subsequently run conda install conda-forge::charset-normalizer -y

Additional Modifications

  • I am using prismatic 41a93d1295ca18e593223f6cde50506db7728ec7
  • Prismatic VLMs also disables gradients for the vision backbone. Comment out https://github.com/TRI-ML/prismatic-vlms/blob/main/prismatic/models/vlms/prismatic.py#L308 to optimize attacks.
  • Prismatic VLMs also assumes a default HF_HUB_REPO=TRI-ML/prismatic-vlms. Until Sidd adds our models to his HF repo, we need to add a conditional to prismatic/models/load.py to handle our models. Go to line 58:
        if model_id_or_path not in GLOBAL_REGISTRY:
            raise ValueError(f"Couldn't find `{model_id_or_path = }; check `prismatic.available_model_names()`")

        overwatch.info(f"Downloading `{(model_id := GLOBAL_REGISTRY[model_id_or_path]['model_id'])} from HF Hub")
        config_json = hf_hub_download(repo_id=HF_HUB_REPO, filename=f"{model_id}/config.json", cache_dir=cache_dir)
        checkpoint_pt = hf_hub_download(
            repo_id=HF_HUB_REPO, filename=f"{model_id}/checkpoints/latest-checkpoint.pt", cache_dir=cache_dir
        )

Replace the above python code with the below python code:

        rylan_trained_models = {
            "prism-gemma-instruct+2b+clip",
            "prism-gemma-instruct+2b+siglip",
            "prism-gemma-instruct+2b+dinosiglip",
            "prism-gemma-instruct+8b+clip",
            "prism-gemma-instruct+8b+siglip",
            "prism-gemma-instruct+8b+dinosiglip",
            "prism-llama2-chat+7b+clip",
            "prism-llama2-chat+7b+siglip",
            "prism-llama2-chat+7b+dinosiglip",
            "prism-llama3-instruct+8b+clip",
            "prism-llama3-instruct+8b+siglip",
            "prism-llama3-instruct+8b+dinosiglip",
            "prism-mistral-instruct-v0.2+7b+clip",
            "prism-mistral-instruct-v0.2+7b+siglip",
            "prism-mistral-instruct-v0.2+7b+dinosiglip",
            "prism-phi-instruct-3+4b+clip",
            "prism-phi-instruct-3+4b+siglip",
            "prism-phi-instruct-3+4b+dinosiglip",
        }
        
        if (model_id_or_path not in rylan_trained_models) and (model_id_or_path not in GLOBAL_REGISTRY):
            raise ValueError(f"Couldn't find `{model_id_or_path = }; check `prismatic.available_model_names()`")

        if model_id_or_path in rylan_trained_models:
            HF_HUB_REPO = "RylanSchaeffer/prismatic-vlms"
        else:
            HF_HUB_REPO = "TRI-ML/prismatic-vlms"

        overwatch.info(f"Downloading `{(model_id := GLOBAL_REGISTRY[model_id_or_path]['model_id'])} from HF Hub")
        config_json = hf_hub_download(repo_id=HF_HUB_REPO, filename=f"{model_id}/config.json", cache_dir=cache_dir)
        checkpoint_pt = hf_hub_download(
            repo_id=HF_HUB_REPO, filename=f"{model_id}/checkpoints/latest-checkpoint.pt", cache_dir=cache_dir
        )

Optimizing Jailbreaks

Create a sweep using wandb sweep <path to W&B sweep e.g., sweeps/attack/...>. This will give you a W&B sweep ID. Then:

cd PerezAstraFellowship-Universal-VLM-Jailbreak
conda activate universal_vlm_jailbreak_env
export PYTHONPATH=.
export CUDA_VISIBLE_DEVICES=7
wandb agent rylan/universal-vlm-jailbreak/srvmywo4

Evaluating Jailbreaks

cd PerezAstraFellowship-Universal-VLM-Jailbreak
conda activate universal_vlm_jailbreak_env
export PYTHONPATH=.
export CUDA_VISIBLE_DEVICES=1
wandb agent rylan/universal-vlm-jailbreak-eval/jb02fx4o