What does this skill do, and when should you use it?
This skill helps agents choose compatible containers, vLLM runtime images, and Python package indexes for NVIDIA Jetson Orin and Thor devices. It uses device generation, JetPack, CUDA userland, and GPU SM targets to avoid generic ARM64 artifacts that lack the required Jetson kernels. It prioritizes NVIDIA-maintained GHCR, NGC, and Jetson AI Lab sources. It provides catalog guidance and compatibility hints but does not install packages or pull images.
Runs scripts/artifact_hints.sh and reads its detected SKU, generation, product line, variant, L4T, CUDA SM hint, canonical URLs, and preferred vLLM image; directs the agent to NVIDIA-AI-IOT GHCR packages, NGC CUDA/PyTorch catalogs, and the Jetson AI Lab PyPI tree; maps Orin to sm_87 and Thor to sm_110; provides a matching extra-index example for GPU packages; and forbids guessing when device facts are unavailable.
- A Jetson user needs a Docker image matching an Orin or Thor device and its JetPack version before running a container.
- A developer is installing PyTorch, vLLM, or onnxruntime-gpu and wants to avoid generic ARM64 wheels without Jetson GPU kernels.
- A user cannot find an l4t-cuda image on NGC and needs the current multi-architecture CUDA alternative.
- A user has CUDA, SM, or stale-container-tag errors and needs device-specific artifact hints first.
How do you install this skill?
- Before execution, run the detector on the actual Jetson and verify JetPack, L4T, CUDA, and current catalog tags; do not execute placeholder tags literally.
- GHCR, NGC, and Jetson AI Lab indexes require network access; confirm mainland-China reachability and enterprise network policy.
- Before installation, independently verify image and wheel provenance, digest, or signature, and obtain user confirmation for external downloads.
- Shell / CLI
- Network access
- Local filesystem
Bash
Install the individual skill with the CLI command documented in the repository README: npx skills add nvidia/skills --skill jetson-package --yes. The README does not document another skill-specific installation method.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/jetson-package" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
On the Jetson target, run bash scripts/artifact_hints.sh from the repository root and read its JSON output. Then inspect the Jetson AI Lab PyPI index, GHCR packages, or NGC catalog using the detected JetPack/CUDA line. For example, onnxruntime-gpu can use: pip install --extra-index-url https://pypi.jetson-ai-lab.io/jp6/cu126/+simple/ onnxruntime-gpu. Adjust jp6/cu126 to the matching line; placeholder tags must be looked up rather than copied literally.
What are this skill's strengths and limitations?
- Clearly distinguishes Orin sm_87 from Thor sm_110.
- Prioritizes NVIDIA GHCR, NGC, and Jetson AI Lab as canonical sources.
- Requires device facts before recommending architecture- or version-specific artifacts.
- Defines a read-only scope and does not install software or pull images.
- Does not verify whether a specific model checkpoint fits in device memory.
- NGC and GHCR tags change, so current catalog lookup requires network access.
- Detection must run on the Jetson target; unknown generation or SM prevents specific tag recommendations.
- Provides catalog and compatibility guidance rather than a complete installation or runtime troubleshooting workflow.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| Jetson Package & Environment Advisor this page ✓ NVIDIA · Official | 56 · Use with care | ★ 3.8k | today | Apache-2.0 |
| Jetson LLM Serving ✓ NVIDIA · Official | 54 · Use with care | ★ 3.8k | today | Apache-2.0 |
| Jetson vLLM Speculative Decoding ✓ NVIDIA · Official | 54 · Use with care | ★ 3.8k | today | Apache-2.0 |
| Jetson LLM Benchmark ✓ NVIDIA · Official | 53 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO NVIDIA GPU Host Setup ✓ NVIDIA · Official | 55 · Use with care | ★ 3.8k | today | Apache-2.0 |
Compared with installing GPU wheels from default PyPI or using arbitrary arm64 images from Docker Hub, this skill prioritizes Jetson AI Lab, NVIDIA GHCR, and NGC while checking JetPack, CUDA, and SM targets. When legacy l4t-cuda images are unavailable, it points to the multi-architecture nvcr.io/nvidia/cuda line.
How did FollowSkills review this skill?
The skill explicitly operates read-only and does not install or pull artifacts; it requires detected device facts rather than guessing, and shows no credential theft, covert exfiltration, or destructive default. Points are deducted because external registries and package indexes are trusted without documented signature/hash verification, explicit user confirmation, or rollback guidance.
The script uses strict Bash settings, emits intelligible errors for unknown Jetson generations, and refuses to guess; the instructions, script, and references are largely consistent. Static review did not execute it, and its detector dependency is not included in the supplied material, so points are deducted; the static ceiling also limits this dimension to 10.
Audience, trigger scenarios, Orin/Thor branches, unknown-device limits, and handoffs are clearly stated. Points are deducted because it is narrowly Jetson-specific, depends on network catalogs and device-side detection, provides no Chinese-language guidance, and does not establish mainland-China network reachability.
The skill has frontmatter, layered sections, prerequisites, script arguments, references, limitations, version, license, and a skill card. Points are deducted for the missing Examples section noted by validation, nonconforming author metadata, absent changelog and explicit maintenance/update path, and hidden assumptions around changing tags.
It connects device detection, SM targets, JetPack/CUDA indexes, and container selection, while discouraging generic ARM wheels; the supplied evaluation report shows positive results on core metrics. Points are deducted because the dataset is unavailable, coverage is limited, and static review cannot confirm that commands and artifacts work directly in practice.
The supplied revision context, script, evaluation summary, and assertions provide partial traceability for key claims. Points are deducted because reproducible evaluation data, execution logs, and skill-specific committed tests are absent; static calibration therefore limits this dimension to 5.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →