Running MnemoVi on Linux

If your distribution is older than the floor, MnemoVi will not start — that is a hard limit, not a setting. Everything newer works, including releases that came out after this page was written.

Worth checking before you buy: Ubuntu 22.04 LTS, Debian 12 and RHEL 9 are below the floor. They are common and they are still supported by their vendors, but their C library is older than MnemoVi needs, and there is no workaround short of moving to a newer release.

The AppImage is 430.6 MiB. That is the cost of everything being included — and since there is no updater, it is what each new release downloads.

It has to be the 12 series. CUDA 13 will not do instead: the libraries carry the major version in their filename, and MnemoVi looks for the 12 ones by name — a 13 install leaves it reporting the same thing as no install at all. The two can sit side by side, so installing 12 does not disturb a 13 you already have. Within the 12 series pick 12.8 or newer: cards from the RTX 50 series onwards are not known to earlier releases. The driver is almost never the problem — drivers are backward compatible, so a current one already runs CUDA 12, and you should never downgrade a working driver.

Get them from NVIDIA. We do not mirror their software and we do not write install steps for it — their download pages are the only place either should come from:

Start here, before downloading anything. The version nvidia-smi prints is the newest CUDA your driver could support — not what is installed. Reading it as "CUDA is present" is the single most common reason people are sure they have it while MnemoVi says they do not. These two commands answer the real question:

nvcc --versionldconfig -p | grep -E 'libcudart\.so\.12|libcudnn\.so\.9'

Read this before following NVIDIA’s download page. That page serves the newest release, and its copy-paste block currently finishes with cuda-toolkit-13-3 — CUDA 13, the one series that cannot work here. The 12 series sits in the same repository; it only has to be asked for by name. This installs the keyring once, then names both packages explicitly:

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.debsudo dpkg -i cuda-keyring_1.1-1_all.debsudo apt-get updatesudo apt-get -y install cuda-toolkit-12-9 cudnn9-cuda-12

You may not need the multi-gigabyte toolkit at all. NVIDIA publishes the same libraries as Python packages, and MnemoVi already looks inside them — no root, no repository, no system changes, and they are found automatically after a restart:

pip install --user nvidia-cuda-runtime-cu12 nvidia-cublas-cu12 \    nvidia-cudnn-cu12 nvidia-cufft-cu12 nvidia-curand-cu12 \    nvidia-cuda-nvrtc-cu12

MnemoVi searches the usual places on its own — /usr/local/cuda, /opt/cuda, your distribution’s library directories, and Python package directories. If yours live somewhere it would not think to look, name the directory and it will look there first. It also respects CUDA_PATH, CUDA_HOME, CUDNN_PATH and LD_LIBRARY_PATH if you already set those:

export MN_CUDA_LIB_DIR=/path/to/your/cuda/lib64

You do not have to guess, and you do not have to reinstall the app to find out. When MnemoVi reports the components as missing it offers a "Check again" button that re-examines the machine on the spot. Once it reports them found, restart the app — the engines are chosen at startup.