jetson-package

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Jetson Package & Environment

Jetson 包与环境

Agents often suggest

docker pull

images or

pip install

wheels that claim aarch64 support but were never built for Jetson’s GPU streaming multiprocessor (SM) targets. On Jetson, default to NVIDIA-curated artifacts unless the user explicitly opts out.

智能体常推荐声称支持aarch64但并未针对Jetson GPU的**流多处理器（SM）**目标构建的

docker pull

镜像或

pip install

安装包。在Jetson上，除非用户明确选择其他方案，否则默认使用NVIDIA官方维护的制品。

Purpose

用途

Choose Jetson-compatible containers and Python package indexes before installing GPU-native ML stacks. This skill prevents agents from recommending generic ARM wheels or stale container tags that do not include the right CUDA, JetPack, or SM target for the device.

在安装GPU原生机器学习栈之前，选择兼容Jetson的容器和Python包索引。该技能可避免智能体推荐未适配设备所需的CUDA、JetPack或SM目标的通用ARM安装包或过时容器标签。

When to use

使用场景

"Which Docker image / container should I use on this Jetson?"
"Where do I get PyTorch / vLLM / CUDA wheels for Jetson?"
"
```
pip install
```
failed" or "wrong CUDA / SM" after installing a generic ARM wheel.
Before
```
docker run
```
or
```
pip install
```
for ML stacks on Orin or Thor.
User or agent looks for
```
l4t-cuda
```
containers on NGC — redirect to
```
nvcr.io/nvidia/cuda
```
(multi-arch).
"Which PyTorch container should I use on Jetson?" — answer depends on Thor vs Orin and JetPack version.

“在这台Jetson上应该使用哪个Docker镜像/容器？”
“在哪里可以获取适用于Jetson的PyTorch / vLLM / CUDA安装包？”
安装通用ARM安装包后出现
```
pip install
```
失败或“CUDA/SM版本不匹配”的问题。
在Orin或Thor上运行
```
docker run
```
或
```
pip install
```
安装机器学习栈之前。
用户或智能体在NGC上查找
```
l4t-cuda
```
容器时，引导至
```
nvcr.io/nvidia/cuda
```
（多架构版本）。
“在Jetson上应该使用哪个PyTorch容器？”——答案取决于Thor或Orin型号以及JetPack版本。

Canonical sources (use these first)

官方推荐源（优先使用）

Prebuilt containers (GHCR) — NVIDIA-AI-IOT packages:
```
llama_cpp
```
,
```
ollama
```
,
```
live-vlm-webui
```
, older-Orin
```
vllm
```
, and related images built for Jetson JetPack stacks. Prefer these over random
```
arm64
```
images on Docker Hub. For vLLM, use upstream
```
vllm/vllm-openai
```
on Thor and Orin JetPack 7.2 / L4T r39+.

NGC CUDA / PyTorch containers — Tag selection depends on Jetson generation. Do not treat example PyTorch tag shapes as pinned recommendations; look up the current tag in the NGC PyTorch catalog before giving a command.

Jetson	CUDA base	PyTorch
Thor	`nvcr.io/nvidia/cuda:<ver>-devel-ubuntu<ver>` (multi-arch, arm64 included)	`nvcr.io/nvidia/pytorch:<current-tag>-py3` (main multi-arch tag; verify current NGC tag)
Orin + r36 / JetPack 6	same multi-arch CUDA base	`nvcr.io/nvidia/pytorch:<current-tag>-py3-igpu` — verify the current NGC tag and use the `-igpu` suffix for Orin iGPU (SM 8.7) when NGC publishes it
Orin + r39+ (future)	same	likely main multi-arch tag once Orin becomes SBSA; verify when r39 ships

l4t-cuda
is the legacy Orin-era CUDA container line. If a user cannot find

l4t-cuda

on NGC, redirect them to the current multi-arch

nvcr.io/nvidia/cuda

image instead of third-party images. 3. Python package indexes (devpi) — Jetson AI Lab PyPI: browse the tree (for example

jp6/cu126

jp6/cu128

) and pick the index that matches your JetPack / CUDA userland. Prefer these over PyPI-only wheels for GPU-native stacks.

预构建容器（GHCR） — NVIDIA-AI-IOT 包：包含为Jetson JetPack栈构建的
```
llama_cpp
```
、
```
ollama
```
、
```
live-vlm-webui
```
、旧款Orin适配的
```
vllm
```
及相关镜像。优先选择这些镜像，而非Docker Hub上随机的
```
arm64
```
镜像。对于vLLM，在Thor和JetPack 7.2/L4T r39+版本的Orin上使用上游的
```
vllm/vllm-openai
```
镜像。

NGC CUDA / PyTorch容器 — 标签选择取决于Jetson世代。请勿将示例中的PyTorch标签格式视为固定推荐，在给出命令前请先在NGC PyTorch目录中查找当前可用的标签。

Jetson型号	CUDA基础镜像	PyTorch镜像
Thor	`nvcr.io/nvidia/cuda:<ver>-devel-ubuntu<ver>` （多架构，包含arm64）	`nvcr.io/nvidia/pytorch:<current-tag>-py3` （主多架构标签；请确认NGC上的当前标签）
Orin + r36 / JetPack 6	相同的多架构CUDA基础镜像	`nvcr.io/nvidia/pytorch:<current-tag>-py3-igpu` — 请确认NGC上的当前标签，当NGC发布该版本后，针对Orin集成GPU（SM 8.7）使用带 `-igpu` 后缀的标签
Orin + r39+（未来版本）	相同	一旦Orin支持SBSA，将使用主多架构标签；请在r39发布时确认

l4t-cuda
是Orin时代的旧版CUDA容器系列。如果用户在NGC上找不到

l4t-cuda

，请引导他们使用当前的多架构

nvcr.io/nvidia/cuda

镜像，而非第三方镜像。 3. Python包索引（devpi） — Jetson AI Lab PyPI：浏览索引目录（例如

jp6/cu126

、

jp6/cu128

），选择与你的JetPack / CUDA用户态匹配的索引。对于GPU原生栈，优先选择这些索引而非仅使用PyPI的安装包。

GPU architecture reminder (why generic ARM fails)

GPU架构说明（通用ARM包失效的原因）

Jetson family	CUDA compute capability	Build target	Note
Orin (AGX / NX / Nano)	8.7	`sm_87`	Many desktop `aarch64` wheels omit Jetson Orin kernels.
Thor (T5000 / T4000)	11.0	`sm_110`	Requires CUDA / wheels / containers that include Blackwell Jetson support.

A wheel or container may install on ARM64 Linux and still be unusable or slow if CUDA kernels were not compiled for your Jetson’s SM.

Use CUDA build target names when discussing wheel compatibility:

sm_87

for Jetson Orin and

sm_110

for Jetson Thor. Do not infer the generation from a prompt or a hostname — run

scripts/artifact_hints.sh

and use its detected

generation

variant

l4t

, and

cuda_sm_hint

fields before recommending wheels or container tags.

Jetson系列	CUDA计算能力	构建目标	说明
Orin（AGX / NX / Nano）	8.7	`sm_87`	许多桌面端 `aarch64` 安装包未包含Jetson Orin的内核。
Thor（T5000 / T4000）	11.0	`sm_110`	需要包含Blackwell Jetson支持的CUDA/安装包/容器。

如果CUDA内核未针对你的Jetson的SM编译，即使安装包或容器能在ARM64 Linux上安装，也可能无法使用或运行缓慢。

讨论安装包兼容性时，请使用CUDA构建目标名称：Jetson Orin对应

sm_87

，Jetson Thor对应

sm_110

。请勿通过提示词或主机名推断设备世代——在推荐安装包或容器标签前，运行

scripts/artifact_hints.sh

并使用其检测到的

generation

、

variant

、

l4t

和

cuda_sm_hint

字段。

GPU Python wheels on Jetson

Jetson上的GPU Python安装包

Default PyPI wheels for GPU-native packages are usually not the right answer on Jetson, even when they claim

aarch64

support. For

onnxruntime-gpu

, PyTorch, vLLM, and similar packages, use the Jetson AI Lab package index as the canonical source and choose the subtree that matches the device's JetPack / CUDA userland.

For

onnxruntime-gpu

, lead with Jetson AI Lab rather than plain PyPI:

bash

pip install --extra-index-url https://pypi.jetson-ai-lab.io/jp6/cu126/+simple/ onnxruntime-gpu

Adjust the

jp6/cu126

portion to match the detected JetPack / CUDA line. Do not present

pip install onnxruntime-gpu

from default PyPI as an equivalent Jetson GPU option.

GPU原生包的默认PyPI安装包通常不适用于Jetson，即使它们声称支持

aarch64

。对于

onnxruntime-gpu

、PyTorch、vLLM及类似包，请将Jetson AI Lab包索引作为官方源，并选择与设备JetPack/CUDA用户态匹配的子目录。

对于

onnxruntime-gpu

，优先使用Jetson AI Lab源而非默认PyPI：

bash

pip install --extra-index-url https://pypi.jetson-ai-lab.io/jp6/cu126/+simple/ onnxruntime-gpu

调整

jp6/cu126

部分以匹配检测到的JetPack/CUDA版本。请勿将默认PyPI的

pip install onnxruntime-gpu

作为Jetson GPU的等效方案。

Do not fabricate device facts

请勿编造设备信息

Do not invent SKU names, RAM sizes, JetPack versions, CUDA versions, or GPU SM targets. Quote only what

scripts/artifact_hints.sh

or the user's supplied environment reports. If a field is unavailable, omit it or say it is unknown.

请勿编造SKU名称、RAM大小、JetPack版本、CUDA版本或GPU SM目标。仅引用

scripts/artifact_hints.sh

或用户提供的环境报告中的信息。如果某字段不可用，请省略或说明其未知。

Prerequisites

前提条件

Run package-detection scripts on a Jetson target, not on the host workstation.
Network access is needed to inspect GHCR, NGC, or Jetson AI Lab package indexes.
Source device facts from
```
scripts/artifact_hints.sh
```
,
```
jetson-diagnostic
```
, or user-provided environment output before recommending tags or wheels.

在Jetson设备上运行包检测脚本，而非主机工作站。
需要网络访问权限以查看GHCR、NGC或Jetson AI Lab包索引。
在推荐标签或安装包前，从
```
scripts/artifact_hints.sh
```
、
```
jetson-diagnostic
```
或用户提供的环境输出中获取设备信息。

Available Scripts

可用脚本

Script	Purpose	Arguments
`scripts/artifact_hints.sh`	Emits detected Jetson SKU/generation, CUDA SM hint, canonical package URLs, and a preferred vLLM image hint.	`--human` for a readable summary; no argument for JSON.

If your agent runtime supports

run_script

, use it to run

scripts/artifact_hints.sh

and read the JSON output. Otherwise run the script with

bash

from the repository root.

脚本	用途	参数
`scripts/artifact_hints.sh`	输出检测到的Jetson SKU/世代、CUDA SM提示、官方包URL以及推荐的vLLM镜像提示。	使用 `--human` 参数可获取可读摘要；无参数则输出JSON格式。

如果你的智能体运行时支持

run_script

，请使用它运行

scripts/artifact_hints.sh

并读取JSON输出。否则，从仓库根目录使用

bash

运行该脚本。

Instructions

操作步骤

Run

scripts/artifact_hints.sh

(JSON on stdout). It sources

skills/jetson-diagnostic/scripts/detect_jetson.sh

and returns

sku

generation

product_line

variant

l4t

, a preferred vLLM image,

cuda_sm_hint

, and canonical URLs.

For pip, open the devpi root in a browser, pick the jp6 subtree that matches your CUDA line, and set
```
--extra-index-url
```
/
```
PIP_EXTRA_INDEX_URL
```
— see
```
references/pypi-jetson-ai-lab.md
```
.

For containers, see

references/ghcr-images.md

and

jetson-llm-serve

for vLLM.

运行
```
scripts/artifact_hints.sh
```
（标准输出为JSON格式）。该脚本会调用
```
skills/jetson-diagnostic/scripts/detect_jetson.sh
```
，返回
```
sku
```
、
```
generation
```
、
```
product_line
```
、
```
variant
```
、
```
l4t
```
、推荐的vLLM镜像、
```
cuda_sm_hint
```
以及官方URL。
对于pip，在浏览器中打开devpi根目录，选择与你的CUDA版本匹配的jp6子目录，并设置
```
--extra-index-url
```
/
```
PIP_EXTRA_INDEX_URL
```
——详情请参考
```
references/pypi-jetson-ai-lab.md
```
。
对于容器，请参考
```
references/ghcr-images.md
```
；vLLM相关内容请参考
```
jetson-llm-serve
```
。

Limitations

局限性

This skill points to package catalogs and emits compatibility hints; it does not verify that a specific model checkpoint fits in memory.
NGC and GHCR tags change. Treat placeholder tag shapes such as
```
<current-tag>-py3
```
as lookup instructions, not literal tags.
If
```
generation
```
or
```
cuda_sm_hint
```
is unknown, do not guess a container tag.

该技能仅指向包目录并提供兼容性提示，不验证特定模型 checkpoint 是否能装入内存。
NGC和GHCR的标签会更新。请将
```
<current-tag>-py3
```
这类占位符标签格式视为查找指引，而非字面标签。
如果
```
generation
```
或
```
cuda_sm_hint
```
未知，请勿猜测容器标签。

Hand off to

后续关联技能

```
jetson-llm-serve
```
— run upstream/native vLLM 0.20+ on Thor and Orin JetPack 7.2 / L4T r39+, or
```
vllm:latest-jetson-orin
```
on older Orin.
```
jetson-llm-benchmark
```
— measure after the stack is installed.
```
jetson-diagnostic
```
— if installs succeed but runtime fails, snapshot first.

```
jetson-llm-serve
```
— 在Thor和JetPack 7.2/L4T r39+版本的Orin上运行上游/原生vLLM 0.20+，在旧款Orin上运行
```
vllm:latest-jetson-orin
```
。
```
jetson-llm-benchmark
```
— 在栈安装完成后进行性能测试。
```
jetson-diagnostic
```
— 如果安装成功但运行时失败，请先进行快照诊断。

Safety

安全性

Read-only: points to catalogs and emits hints; does not install or pull.

只读操作：仅指向目录并提供提示，不执行安装或拉取操作。

Sources

参考源

NVIDIA-AI-IOT GitHub Packages, pypi.jetson-ai-lab.io.