holoscan-install-container

This skill installs the NVIDIA Holoscan SDK via its official NGC Docker container and verifies the installation. It solves the problem of getting a correct, GPU-matched Holoscan environment running without native apt/pip/Conda installs: it pulls the official image from nvcr.io/nvidia/clara-holoscan/holoscan, selects the right CUDA/architecture tag for the host GPU, and validates the result against the bundled example programs. It is explicitly scoped to container-based installs and points to the canonical NVIDIA documentation as the source of truth if anything disagrees.

The workflow proceeds in ordered steps: choose the tag by reading the SDK version from the docs and matching a suffix (cuda13, cuda12-dgpu, or cuda12-igpu) to the host's nvidia-smi CUDA version; verify GPU passthrough with a throwaway ubuntu container running nvidia-smi and then pull the image (warning about the 10-20 GB download); verify with six bundled examples spanning bare Python and C++ runtimes and Holoviz/Vulkan rendering; and finally hand off a documented launch command. Holoviz examples are run headless by injecting a headless flag into the example YAML so they work over SSH without X11. The skill documents prerequisites (NVIDIA driver, Docker, NVIDIA Container Toolkit), limitations, and troubleshooting for common driver/toolkit and CUDA-mismatch errors.

It targets developers and researchers deploying Holoscan for medical/edge AI and sensor-processing workloads who prefer containers. The Docker commands use elevated runtime options (CAP_SYS_PTRACE, --ipc=host, raised ulimits, --gpus all) that are standard and necessary for GPU compute containers, and all commands pull from and reference official NVIDIA sources rather than untrusted ones.