Evo 2 (7B) — User Guide

Evo 2 is a genomic foundation model developed by the Arc Institute and NVIDIA. It is trained on 9+ trillion DNA base pairs across all domains of life and can process sequences up to 1 million nucleotides. The 7B parameter version is available on our shared filesystem as an Apptainer container, ready to use with no installation required. This guide covers setup, basic usage, and common tasks such as scoring sequences, extracting embeddings, and generating DNA.

Where is it?

The Apptainer container image is stored on the shared filesystem:

Container: /mnt/dv/wid/projects1/WID-Software/rocky8/apptainer_images/evo2.sif

Setup

Add the following to your ~/.bashrc (then run source ~/.bashrc):

export EVO2_SIF=/mnt/dv/wid/projects1/WID-Software/rocky8/apptainer_images/evo2.sif
export EVO2_MODELS=$HOME/.cache/huggingface
mkdir -p $EVO2_MODELS

Note: The first time you run Evo 2, it will download model weights (~14GB) into $EVO2_MODELS. If your home directory has a tight quota, point EVO2_MODELS to a scratch or project directory instead.

Quick Start

1. Test that it works

This runs a short generation test to confirm everything is set up:

apptainer exec --nv \
  --bind $EVO2_MODELS:/root/.cache/huggingface \
  $EVO2_SIF \
  python3 -m evo2.test.test_evo2_generation --model_name evo2_7b

You should see generated DNA sequence output. If you get errors, see the Troubleshooting section below.

2. Interactive shell

This drops you into a shell inside the container where you can run Python interactively:

apptainer shell --nv \
  --bind $EVO2_MODELS:/root/.cache/huggingface \
  --bind $PWD:/work \
  $EVO2_SIF

Once inside (you'll see an Apptainer> prompt), start Python:

Apptainer> cd /work
Apptainer> python3

from evo2 import Evo2
model = Evo2('evo2_7b')
output = model.generate(prompt_seqs=["ACGT"], n_tokens=100, temperature=1.0, top_k=4)
print(output.sequences[0])

Type exit to leave Python, and exit again to leave the container.

3. Run your own script

If you have a Python script called my_analysis.py in your current directory:

apptainer exec --nv \
  --bind $EVO2_MODELS:/root/.cache/huggingface \
  --bind $PWD:/work \
  --pwd /work \
  $EVO2_SIF \
  python3 my_analysis.py

4. Select a specific GPU

To restrict the container to a single GPU (e.g. GPU 1), set CUDA_VISIBLE_DEVICES before launching:

CUDA_VISIBLE_DEVICES=1 apptainer exec --nv \
  --bind $EVO2_MODELS:/root/.cache/huggingface \
  --bind $PWD:/work \
  --pwd /work \
  $EVO2_SIF \
  python3 my_analysis.py

Understanding the Command

Here's what each piece means:

apptainer exec                                           # run a command inside the container
  --nv                                                   # enable GPU access
  --bind $EVO2_MODELS:/root/.cache/huggingface           # make model weights visible inside the container
  --bind $PWD:/work                                      # make your current directory visible as /work
  --pwd /work                                            # start in /work inside the container
  $EVO2_SIF                                              # the container file
  python3 my_script.py                                   # the command to run

Key concept — bind mounts: The container can't see your files by default. --bind /outside/path:/inside/path makes a directory on the host visible inside the container. You can add multiple --bind flags for different directories

Troubleshooting

fix table
Problem	Solution
`AssertionError: libcuda.so cannot found!`	Make sure you're using `--nv` and running on a node with a GPU.
`No space left on device`	Apptainer uses `/tmp` during runs. Set `export APPTAINER_TMPDIR=/scratch/$USER/tmp` to use a larger directory.
Permission denied on model weights	Check that your `$EVO2_MODELS` directory exists and is writable: `ls -la $EVO2_MODELS`
Script can't find your input files	Add another `--bind` for the directory containing your data.
`CUDA out of memory`	The 7B model needs ~16GB GPU memory. Request a GPU with enough VRAM.