Holoscan NGC Container Installer
Install, validate, and launch NVIDIA Holoscan SDK through an NGC Docker container.
The skill identifies an official image source, prerequisites, disk/network impact, and warns before the large pull. However, docker-group/sudo use, --ipc=host, CAP_SYS_PTRACE, full GPU access, and external network access are not paired with explicit per-action confirmation, digest/signature verification, or cleanup/rollback guidance, so points are deducted.
The ordered workflow, expected outputs, and troubleshooting notes are useful, and the benchmark report claims evaluation success. Static review did not execute anything, however. Commands depend on image layout, brittle YAML text substitutions, and tag assumptions; edge cases and version-drift feedback are limited, so the score is capped and reduced.
The audience and scenario are clear: Linux hosts with NVIDIA GPUs using Docker/NGC, explicitly excluding native apt/pip/Conda installs. GPU/architecture tag guidance and headless SSH handling improve fit. Unsupported drivers/platforms and network boundaries are not fully defined, and Chinese-language support or mainland-China reachability is not addressed despite reliance on overseas NGC/documentation services, so points are deducted.
The skill has structured frontmatter, version, Apache-2.0 licensing, ownership, scope, limitations, commands, troubleshooting, and benchmark documentation. It lacks a dedicated Examples section, changelog, explicit maintenance/update path, pinned dependency information, and image lifecycle guidance, so points are deducted.
The core task is covered with direct commands for image retrieval, GPU passthrough, six example validations, and launch. The included benchmark reports positive-task results, but users still need to adapt the tag, driver, and environment, and no execution was performed in this review; the static ceiling therefore applies.
The material names NVIDIA documentation, the NGC container, and the product repository, and includes expected outputs plus benchmark tasks/results. It lacks a skill-specific executable test suite, CI coverage, or independently reviewable execution artifacts; the benchmark is only an in-file assertion, so points are deducted.
- Before execution, confirm Docker, GPU, IPC, and CAP_SYS_PTRACE access, external image retrieval, and approximately 10–20 GB of disk usage.
- Do not interpret “verify” as image-integrity verification; the workflow mainly runs examples and lacks digest, signature, or vulnerability-scanning steps.
- The CUDA tag table, image paths, and sed substitutions are version-sensitive; recheck the canonical NVIDIA material when the image or SDK changes.
- NGC and documentation access may be constrained from mainland China; no mirror, proxy, or offline installation path is provided.
What does this skill do, and when should you use it?
This skill guides Linux users with NVIDIA GPUs through a container-based Holoscan SDK installation from NGC. It selects a CUDA and architecture tag from the host's nvidia-smi output, checks GPU passthrough, pulls the image, and validates it with six Python and C++ examples. It configures Holoviz examples for headless execution and provides a launch command plus troubleshooting guidance. It does not cover native apt, pip, or Conda installations.
Reads the CUDA Version field from nvidia-smi; selects the cuda13, cuda12-dgpu, or cuda12-igpu suffix; runs a Docker GPU passthrough check; pulls nvcr.io/nvidia/clara-holoscan/holoscan:<TAG>; runs hello_world, tensor_interop, and video_replayer examples in Python and C++; patches YAML files with sed to enable Holoviz headless mode; sets HOLOSCAN_INPUT_PATH; explains Docker runtime flags and supplies a container launch command.
- A Linux NVIDIA GPU user wants to install Holoscan SDK from the official NGC container.
- A developer needs to validate Holoviz examples over SSH or on a headless host.
- A Holoscan developer wants one guided check covering Python bindings, the C++ runtime, and video examples.
- A user needs concrete diagnosis for CUDA initialization, NVIDIA runtime, stack-limit, display, or video-path failures after pulling the container.
What are this skill's strengths and limitations?
- Covers the workflow from GPU passthrough and image pull through six example validations.
- Includes specific remedies for tag mismatch, stack size, Holoviz display, and video data-path failures.
- Runs Holoviz validation headlessly, which suits SSH and display-less environments.
- Published within NVIDIA's Agent Skills catalog, which is installed through the npx skills CLI and maintained through ongoing synchronization.
- Supports only the container tag matrix described in the skill and provides no Conda or pip environment inside the container.
- Requires a Linux NVIDIA GPU, working driver, Docker, and NVIDIA Container Toolkit.
- The image pull requires approximately 10–20 GB of disk space and network access to nvcr.io.
- Not intended for native apt, pip, Conda, or source installations of Holoscan SDK.
How do you install this skill?
Provide a Linux host with an NVIDIA GPU and working driver, Docker, NVIDIA Container Toolkit, network access to NGC, and approximately 10–20 GB of free disk space. Install the specific skill with the NVIDIA skills CLI: npx skills add nvidia/skills --skill holoscan-install-container --yes. The README does not specify a fixed manual folder path; the CLI prompts for the installation destination.
How do you use this skill?
Ask a compatible Agent Skills client something like: “Install Holoscan SDK on my NVIDIA Linux host using the NGC Docker container and validate all six examples.” The skill checks nvidia-smi, chooses a <version>-<suffix> tag, verifies docker run --rm --gpus all ubuntu:22.04 nvidia-smi, pulls nvcr.io/nvidia/clara-holoscan/holoscan:<TAG>, runs the six examples, and hands off a launch command using --runtime=nvidia, --gpus all, --ipc=host, and the specified ulimits. The current SDK version should be taken from the referenced NVIDIA documentation page.
How does this skill compare with similar options?
Compared with native apt, pip, Conda, or source installation, this skill specifically targets the NGC Docker-container route; those alternative installation methods are outside its scope.