installation-linux

Install Sensory on Linux

Applies to: all

This guide applies to the Linux x64 distribution of Sensory 7.4.0 (Sensory-7.4.0-linux-x64.tar.gz, optionally Sensory-7.4.0-linux-x64.deb).

Before you begin

You need:

  • 64-bit x86-64 Linux with glibc 2.38 or newer, for example Ubuntu 24.04, Debian 13 or Fedora 40 (ldd --version shows the glibc version);
  • an NVIDIA GPU of the Turing generation or newer (GeForce RTX 20 series and later, Compute Capability 7.5+; the package contains kernels for 7.5, 8.6, 8.9 and 12.0 plus PTX) and the proprietary NVIDIA driver supporting CUDA 13.2. nvidia-smi must work. The driver library libcuda.so.1 comes from your system and is not part of the package;
  • for windows, an X11 or Wayland session. Without a display Sensory runs headless automatically;
  • for the camera a V4L2 device, for the microphone PipeWire, PulseAudio or ALSA. Both are optional.

Normal use does not require a compiler, CMake, OpenCV or the CUDA Toolkit.

1. Download and verify

Download the package together with SHA256SUMS.txt and SHA256SUMS.txt.asc from the download page, then verify the signature and the checksum:

gpg --verify SHA256SUMS.txt.asc SHA256SUMS.txt
sha256sum --ignore-missing -c SHA256SUMS.txt

sha256sum must print OK for the downloaded file. If it does not, do not run the package and download it again from cognia.cz.

2a. Portable installation (archive)

Extract the archive anywhere you have write access and keep the directory structure intact; programs find their resources relative to bin/:

tar -xzf Sensory-7.4.0-linux-x64.tar.gz
cd Sensory-7.4.0-linux-x64
./bin/Sensory --help

No root access and no installation step are needed. To use the tools from any directory, add bin/ to your PATH, for example in ~/.bashrc:

export PATH="$HOME/Sensory-7.4.0-linux-x64/bin:$PATH"

2b. Debian package

On Ubuntu 24.04+ and Debian 13+ the .deb installs the same tree into /opt/Sensory and lets apt check the system libraries:

sudo apt install ./Sensory-7.4.0-linux-x64.deb
/opt/Sensory/bin/Sensory --help

3. First run

./bin/Sensory                                              # default model
./bin/Sensory Cognia/examples/literal/literal.cognia Literal
./bin/Sensory --headless Cognia/examples/hierarchy/hierarchy.cognia

Press t to type text, s to save the network and l to load it; closing any window quits the application. Stop a headless run with Ctrl+C.

Check the installation with the command-line tools and a release gate:

./bin/cognia Cognia/examples/hierarchy/hierarchy.cognia
./bin/cognia_run Cognia/examples/hierarchy/hierarchy.cognia
./bin/cognia_validate_gpu

cognia_validate_gpu prints VERDICT: PASS on a supported GPU. On a machine without a usable CUDA device it reports the GPU checks as BLOCKED.

4. Optional: cuBLAS for the transformer CUDA tools

cognia_transformer_serve_cuda, cognia_transformer_train_cuda, cognia_transformer_benchmark and the cognia_validate_transformer_gpu and cognia_validate_placement_gpu gates need the cuBLAS 13 runtime, which is not bundled (about 560 MB). Install it from the NVIDIA CUDA repository or with the CUDA Toolkit 13.2:

sudo apt install libcublas-13-2
ldd bin/cognia_transformer_serve_cuda | grep cublas

If ldd still reports libcublas.so.13 => not found, add the library directory to the loader path, for example export LD_LIBRARY_PATH=/usr/local/cuda/lib64. All other programs, including Sensory, run without cuBLAS.

5. Optional: monitoring desktop

./bin/cognia-desktop.sh                 # start the server and open the browser
./bin/cognia-desktop.sh --install-menu  # add "Cognia Desktop" to the menu

The server listens only on 127.0.0.1; COGNIA_DESKTOP_PORT selects a port other than 8091.

Data and saved networks

User data is stored under $XDG_DATA_HOME/sensory (by default ~/.local/share/sensory). The networks/ directory of the package is reserved for your network files.

Uninstall

  • Archive: delete the extracted directory.
  • Debian package: sudo apt remove sensory.
  • Optionally delete ~/.local/share/sensory and ~/.local/share/applications/cognia-desktop.desktop.

Troubleshooting

  • error while loading shared libraries: libcublas.so.13 — only the transformer CUDA tools need cuBLAS; see step 4.
  • version 'GLIBC_2.38' not found — the distribution is too old; use Ubuntu 24.04, Debian 13, Fedora 40 or newer.
  • No windows appear — check that $DISPLAY or $WAYLAND_DISPLAY is set. Over SSH or in a container Sensory runs headless; use --headless explicitly to skip the check.
  • no usable CUDA device — the NVIDIA driver is missing or the GPU is older than Turing (Compute Capability below 7.5); check nvidia-smi.
  • Camera not found — check that /dev/video0 exists and that your user can read it (usually the video group).