holoscan-install-source

Build Holoscan SDK from source via the in-tree ./run script. Use only when published packages don't meet the user's needs.

2,779stars322forksUpdated 8/4/2026

DevOps & CI/CD #docker #nvidia #cmake #cuda #holoscan #build-from-source

Security Assessment

Safe(91/100)

Security Score91/100

About holoscan-install-source

holoscan-install-source is an official NVIDIA install skill for building the Holoscan SDK from its source tree using the in-tree ./run script, which builds inside a Docker container and produces a local install tree consumable as a CMake dependency. It solves the case where published packages (Conda, container, apt, wheel) do not meet a user's needs — for example when debug symbols, custom CMake options, or unsupported configurations are required — and is explicitly recommended only as a fallback for those situations.

The workflow consults the official 'Build from Source' docs first (authoritative on disagreement), verifies prerequisites (git, Docker with NVIDIA Container Toolkit GPU passthrough, buildx, ~20 GB disk), clones the nvidia-holoscan/holoscan-sdk repository, and optionally checks out a specific release tag for stability. It then runs ./run build (warning that a first build takes 10–30 minutes), documents build options such as --type debug/RelWithDebInfo, --arch aarch64 (needs qemu-user-static), --gpu igpu, and --dryrun, and covers cache clearing and locating the build/install directories. Verification runs the six required C++ and Python example tests (hello_world, tensor_interop, video_replayer) via ./run test. If the NVIDIA Container Toolkit is missing, the doc provides the official installation commands (adding NVIDIA's signed apt repo and configuring the Docker runtime). A skill-card and BENCHMARK.md record strong NVSkills-Eval results.

It targets developers and engineers who must compile the Holoscan SDK for custom build configurations, debug builds, or unsupported platform targets. Its operations — git clone, containerized builds, and toolkit setup that adds a GPG-verified NVIDIA apt source via sudo — are standard developer tooling with no destructive or exfiltrating behavior.

FAQ

When should I build from source instead of using a package?

Only when published packages (Conda, container, apt, wheel) don't fit your needs — for example debug symbols, custom CMake options, or unsupported configurations. Otherwise a prebuilt package is recommended.

What are the prerequisites?

A Linux host with an NVIDIA GPU and driver, git, Docker with the NVIDIA Container Toolkit (so docker run --gpus all works) and docker-buildx-plugin, about 20 GB free disk, and 10–30 minutes for a clean first build.

Is this a true bare-metal build?

No. The ./run script builds inside a Docker container, so Docker is still required; it is not a bare-metal build. Cross-compiling to aarch64 additionally needs qemu-user-static.

How do I build a specific version and verify it?

List tags with git tag and checkout a release tag (e.g. v4.1.0) for stability, run ./run build (optionally with flags like --type debug or --arch aarch64), then run ./run test for the six required hello_world, tensor_interop, and video_replayer C++ and Python example tests.

What does the NVIDIA Container Toolkit setup do?

If GPU passthrough fails, the skill provides NVIDIA's official commands to add a GPG-key-verified NVIDIA apt repository, install nvidia-container-toolkit, and configure the Docker runtime — standard, documented setup steps.