本文へ移動
cccskills
無料GitHub で公開

jetson-init-image

Extract Jetson Linux + sample-rootfs tarballs and run apply_binaries.sh for the active target, then record bsp_image in the profile. Use after jetson-init-target; not for source-tree setup.

インストール方法を見る

含まれるファイル(5)

  • SKILL.md7.1 KB
  • BENCHMARK.md3.7 KB
  • evals/evals.json3.9 KB
  • skill-card.md4.0 KB
  • skill.oms.sig4.5 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Initialize BSP Image

Output is only bsp_image: in the active profile: derived version plus root_path only when overriding <workspace>/Image.

When to invoke

  • The user asks to extract, prepare, or initialize the BSP image.
  • A downstream skill reports missing <bsp_image.root_path>/Linux_for_Tegra/.
  • The active profile has no bsp_image: block yet.

Procedure

Resolve target and image path

Resolve the active profile per ../../context/target-platform-contract.md. Refuse if there is no active profile or reference_devkit: is missing.

<workspace> is the parent of the active profile's target-platform/ directory. <bsp_image.root_path> defaults to <workspace>/Image.

Profile stateAction
bsp_image.root_path existsUse it without prompting.
bsp_image: exists without root_pathUse <workspace>/Image.
No bsp_image: blockAsk once: Enter for default, or absolute override path.

For an override, validate that the closest existing parent is writable. Omit root_path when the default is used.

Determine GPU stack

Use the shared GPU-driver invariant from ../../context/target-platform-contract.md. Derive the expected stack from reference_devkit.module.id and the catalogue:

Chip familyModule IDsStackapply_binaries.sh flag
T234 / Orinp3701, p3767nvgpunone
T264 and later / Thor+p3834OpenRM--openrm

Refuse unknown module IDs. The --openrm flag is only valid on BSP releases that ship the OpenRM stack; if the active BSP doesn't expose the flag, omit it regardless of what the target wants.

Reuse or extract

If <bsp_image.root_path>/Linux_for_Tegra/ already exists:

  1. Do not extract over it unless the user explicitly requested re-extraction and accepted the overwrite risk.

  2. Derive the on-disk version from Linux_for_Tegra/nv_tegra_release. Ask before replacing a different recorded bsp_image.version.

  3. Verify the installed GPU stack against the platform-derived expectation when possible. Detection precedence (first probe that yields a definitive answer wins):

    1. Linux_for_Tegra/rootfs/etc/nv_tegra_release carries an INSTALL_TYPE= token on BSP releases that expose it (newer lines). Read and compare directly.
    2. Otherwise, find Linux_for_Tegra -name nvgpu.ko: present → nvgpu, absent → OpenRM.
    3. If the chip family has only ever shipped one stack (e.g. T234 / Orin is always nvgpu in current BSPs), fall back to the catalogue-derived expectation without disk probing.

If the installed stack conflicts with the active target, refuse and ask the user to re-extract with the correct stack or fix the target profile. Otherwise skip extraction and update the profile.

Locate tarballs

When extraction is needed, search:

  1. <bsp_image.root_path>/
  2. <workspace>/
  3. current working directory

Prompt for absolute paths for anything missing. Required filenames:

  • Jetson_Linux_R<ver>_aarch64.tbz2
  • Tegra_Linux_Sample-Root-Filesystem_R<ver>_aarch64.tbz2

Both filenames must contain the same R<ver> token. Refuse mismatches and record <ver> as bsp_image.version. Do not download tarballs.

Extract and apply binaries

Use absolute tarball paths; they may live outside <bsp_image.root_path>.

ROOT="<bsp_image.root_path>"
BSP_TARBALL="<absolute path to Jetson_Linux_R<ver>_aarch64.tbz2>"
ROOTFS_TARBALL="<absolute path to Tegra_Linux_Sample-Root-Filesystem_R<ver>_aarch64.tbz2>"

mkdir -p "$ROOT"
tar xjf "$BSP_TARBALL" -C "$ROOT"
sudo tar xpjf "$ROOTFS_TARBALL" -C "$ROOT/Linux_for_Tegra/rootfs"

cd "$ROOT/Linux_for_Tegra"
if [ "$GPU_STACK" = "openrm" ]; then
  sudo ./apply_binaries.sh --openrm
else
  sudo ./apply_binaries.sh
fi

Set GPU_STACK from the "Determine GPU stack" step above. Abort on the first failing command and surface the failed command.

Update the active profile

Persist the resolved BSP image metadata in the active target profile so later skills can find the BSP without re-prompting. Preserve existing blocks, comments, and quoted SKU values; use a round-tripping YAML writer such as ruamel.yaml.

bsp_image:
  root_path: <absolute override path>  # omit for <workspace>/Image
  version: "<derived version>"

Rules:

  • Same version and same root path: no rewrite.
  • Different version: ask before updating.
  • Different recorded root_path: refuse automatic rewrite.
  • Always quote version.

Finish

Report the image path, extracted vs reused state, GPU stack, derived version, and profile update status. Then suggest /jetson-init-source.

Purpose

Materialize Linux_for_Tegra/ on disk by extracting the right Jetson Linux + sample-rootfs tarballs and running apply_binaries.sh with the GPU-stack flag derived from the active target (nvgpu for T234, OpenRM for T264+). Then commit the derived BSP version into the profile's bsp_image: block.

Prerequisites

  • Active target profile resolved per ../../context/target-platform-contract.md.
  • Jetson Linux BSP tarball and matching sample-rootfs tarball staged on disk (e.g. by /jetson-download-bsp or hand-placed).
  • Write access to the workspace Image/ root (or the override bsp_image.root_path).
  • sudo available for apply_binaries.sh.

Limitations

  • Writes only the bsp_image: block; source tree, documents, and carrier profile are owned by sibling skills.
  • Refuses to overwrite an existing Linux_for_Tegra/ without explicit user direction.
  • Does not download tarballs; rely on /jetson-download-bsp or hand-stage the inputs.

Troubleshooting

  • apply_binaries.sh exits non-zero — re-read its console output; most failures are missing sudo, missing rootfs tarball, or wrong GPU stack flag for the SoC generation.
  • nv_tegra_release absent after extract — extraction stopped early; verify tarball integrity and rerun.
  • Recorded version disagrees with the tarball filename — the tarball was renamed; trust the value parsed from Linux_for_Tegra/nv_tegra_release over filenames.
  • Different recorded root_path already in profile — refuse and ask the user to confirm before overwriting.

References

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Official NVIDIA-authored guidance for NVIDIA cuDF GPU DataFrames, pandas acceleration, dask-cuDF, ETL, joins, groupby, CSV/Parquet I/O, nullable semantics, and multi-GPU DataFrame workloads.

日本語の概要は準備中です。原文の説明を表示しています。

NVIDIA/skills3,5612026年10月10日 更新

Customize NVIDIA Nemotron Voice Agent's Generic Pipecat example for healthcare appointment, five-field patient intake, or custom tool-calling workflows without a separate backend.

日本語の概要は準備中です。原文の説明を表示しています。

NVIDIA/skills3,5612026年10月10日 更新

Calibrate a new dataset from live RTSP camera streams via the AutoMagicCalib REST API. Use when the user provides RTSP URLs or asks to calibrate live cameras; VIOS records clips, AMC ingests them, then runs calibration.

日本語の概要は準備中です。原文の説明を表示しています。

NVIDIA/skills3,5612026年10月10日 更新

Run end-to-end calibration on the shipped sample dataset (sdg_08_2_sample_data_010926.zip) against a running AMC microservice. Use when user says 'test sample dataset', 'run sample calibration', 'verify AMC install', or 'launch and test'.

日本語の概要は準備中です。原文の説明を表示しています。

NVIDIA/skills3,5612026年10月10日 更新

Calibrates pre-recorded `cam_*.mp4` datasets through the AutoMagicCalib REST API. Use for user-supplied local MP4s; route live RTSP streams to `amc-run-rtsp-calibration`.

日本語の概要は準備中です。原文の説明を表示しています。

NVIDIA/skills3,5612026年10月10日 更新

Launch AutoMagicCalib microservice and web UI from NGC release images via Docker Compose. Use when user says 'deploy auto calibration', 'launch auto calibration', 'launch AMC', 'start MS+UI', or 'set up auto-magic-calib'. Requires NGC API key.

日本語の概要は準備中です。原文の説明を表示しています。

NVIDIA/skills3,5612026年10月10日 更新

NVIDIA のスキルをすべて見る

このスキルの問題を報告する