utulek

# This may be aliased as `python3` on your system.
python -m venv .venv

# .venv\Scripts\activate on Windows.
source .venv/bin/activate

# Common pre-install requirements. `python -m` here avoids permissions errors.
python -m pip install --upgrade -r requirements/pre.txt

# `gpu.txt` should be used instead when a suitable CUDA and cuDNN install is available.
python -m pip install --upgrade -r requirements/cpu.txt

utulek support is always pinned to a single version of all major dependencies. The table below describes the current pinned versions. These versions are guaranteed to be available on up-to-date Manjaro/ Arch.

Query the current CUDA version with nvcc -V or pacman -Q cuda. Query the current cuDNN version with pacman -Q cudnn. Query the current python version with python -V. Older python versions are usually available via the AUR on Arch and can be automatically built via pamac. It is recommended to not upgrade the full system via pacman -Syu while versions are correctly pinned.

Dependency	Version
python	3.11.9
pip	23.0
CUDA	12.4.1
cuDNN	8.9.7
tensorflow	2.16.1
torch	2.3.0
jax	0.4.28

When version pinning is updated, corresponding pins in requirements/* should also be updated. Upgrade packages on Arch with something like sudo pacman -Syu --ignore cuda --ignore cudnn to avoid losing the pin.

Python should be formatted using black.

Darwin systems are lagging in tensorflow, torch, and jax versions. We case for this via the platform_system marker. Darwin is not actively supported/tested.

Currently (2026_01_02), CUDA is only officially tested and supported on Linux. Windows and Darwin systems should use the CPU version of requirements.