Installation

Installation#

Prerequisites#

GATAC requires Python ≥ 3.13, Linux, and a CUDA-capable GPU with drivers compatible with CUDA 12 or 13.

The following system libraries must be available:

  • CUDA Toolkit 12.x or 13.x

Install from PyPI#

Choose the extra of the CUDA version that nvidia-smi reports; it installs RAPIDS (cuDF, cuML, cuGraph) and CuPy for that version:

pip install 'gatac[rapids-cu12]'
# or
pip install 'gatac[rapids-cu13]'

With uv: uv pip install 'gatac[rapids-cu12]', or uv add 'gatac[rapids-cu12]' in a project. Without a CUDA extra, and unless RAPIDS is installed otherwise (see Conda below), the first use of a GATAC module (ga.pp, gatac tile, …) fails with a message naming the extras.

After installation the gatac command is available:

gatac --version
gatac --help

Optional extras#

Extra

For

archr-compat

Seurat’s approximate neighbour search (Annoy), to reproduce ArchR’s clustering

bigwig

Writing footprint tracks as bigWig (FootprintResult.to_bigwig)

Combine them with the CUDA extra: pip install 'gatac[rapids-cu12,bigwig]'.

Install from source#

git clone https://github.com/scverse/gatac.git
cd gatac

# With CUDA 12 support
uv sync --extra rapids-cu12

# Or with CUDA 13 support
uv sync --extra rapids-cu13

Optional: documentation dependencies#

To build the docs locally, install the docs dependency group:

uv sync --group docs

Then build with:

cd docs
make html

Verifying your installation#

import gatac as ga

print(ga.__version__)

# Check that GPU is available
import cudf

print(cudf.get_device_info())

Hardware requirements#

Component

Minimum

Recommended

GPU VRAM

8 GB

24–80 GB

RAM

32 GB

128+ GB

Storage

—

NVMe SSD

Note

GATAC needs an NVIDIA GPU: it imports cuDF and CuPy, which the rapids-cu12/rapids-cu13 extras install.

Conda / Mamba#

There is no conda package of GATAC yet. To use RAPIDS from conda, create an environment with RAPIDS 26.02 or newer (the RAPIDS install selector gives the command for your CUDA version), then install GATAC into it from PyPI, without a CUDA extra so that pip does not install a second RAPIDS:

conda create -n gatac -c rapidsai -c conda-forge \
    rapids=26.10 python=3.13 cuda-version=12.9
conda activate gatac
pip install gatac