evo2-metal

Run ARC Institute's Evo 2 DNA language model on Apple Silicon (Metal GPU) — no CUDA required.

Evo 2 is a 7B–40B parameter genomic foundation model trained on 9.3 trillion DNA tokens. This package makes the 7B models fully usable on a Mac with M-series chips.

How it works

Evo 2 depends on flash-attn, which requires NVIDIA CUDA and does not run on Mac. This package provides:

• A FlashAttention v2 kernel written in Metal Shading Language (Apple's GPU compute API), dispatched via PyObjC.
• A set of monkey-patches applied before importing evo2 that redirect all CUDA-dependent calls to CPU/Metal equivalents — without modifying the original Evo 2 source code.

import evo2_metal   ←  patches flash_attn, vortex, and evo2.scoring
from evo2 import Evo2

Installation

Requires: macOS with Apple Silicon (M1/M2/M3/M4/M5), Python 3.11 or 3.12.

# 1. Create a Python 3.12 environment (evo2 requires <3.13)
conda create -n evo2-mac python=3.12
conda activate evo2-mac

# 2. Install PyTorch (CPU build for Mac)
pip install torch --index-url https://download.pytorch.org/whl/cpu

# 3. Install evo2-metal (PyPI)
pip install evo2-metal

# Alternative: install directly from GitHub
# pip install git+https://github.com/lemonardo1/evo2-metal.git

Usage

Sequence scoring

import evo2_metal          # must be first
from evo2 import Evo2

model = Evo2('evo2_7b_base')

scores = model.score_sequences([
    "ATGAAAGCAATTTTCGTACTGAAAGGTTCAGGT",   # wildtype
    "ATGAAAGCAATTTTCGTACTGAAAGGTTCAGGA",   # variant
])
print(scores)  # [np.float32(-1.27), np.float32(-1.43)]

DNA sequence generation

import evo2_metal
from evo2 import Evo2

model = Evo2('evo2_7b_base')

output = model.generate(
    prompt_seqs=["ATGAAAGCAATTTTCGTACTGAAAGGTTCAGGT"],
    n_tokens=200,
    temperature=1.0,
    top_k=4,
)
print(output.sequences[0])

Supported models

Patches applied

Verified on

• Apple M5 Max, macOS 15 (Sequoia)
• Python 3.12, PyTorch 2.11, evo2 0.5.3

License

Apache 2.0 — see LICENSE.

Evo 2 model weights are released under their own license by ARC Institute. See the Evo 2 repository for details.