Linux prerequisites

Each llama.cpp backend module the app bundles needs its own system libraries on Linux. Install them on every machine that runs the app. Choosing modules: Native runtime configuration.

Dependency mapping#

• cpu: OpenMP runtime (libgomp.so.1). libggml-base.so links it too, so every llama.cpp load needs it whichever backend modules are selected; without it libllamadart.so fails to load with libgomp.so.1: cannot open shared object file.
• vulkan: Vulkan loader and valid GPU driver/ICD.
• blas: OpenBLAS runtime (libopenblas.so.0).
• cuda: NVIDIA driver plus the CUDA 12 runtime libraries. libggml-cuda.so links libcudart.so.12 and libcublas.so.12, and llamadart does not ship them on Linux. Without them the CUDA module fails to load and llama.cpp runs on CPU. A GPU-ready cloud image can ship the driver without them.
• hip: ROCm runtime libs (for example libhipblas.so.2).

Package examples#

Ubuntu/Debian:

sudo apt-get update
sudo apt-get install -y libgomp1 libvulkan1 vulkan-tools libopenblas0

Fedora/RHEL/CentOS:

sudo dnf install -y libgomp vulkan-loader vulkan-tools openblas

Arch Linux:

sudo pacman -S --needed libgomp vulkan-icd-loader vulkan-tools openblas

Minimal cloud and container images, such as the stock GCE Ubuntu 24.04 accelerator image, may lack libgomp1.

Quick link check#

for f in .dart_tool/lib/libggml-*.so; do
  LD_LIBRARY_PATH=.dart_tool/lib ldd "$f" | grep "not found" || true
done

Any not found line names a missing system library for that module.