dusty-nv/jetson-containers

jetson‑containers – Ready‑to‑run AI/ML containers for NVIDIA Jetson

What it is

• A modular build system that creates Docker (or OCI) containers pre‑packed with the latest AI/ML libraries specifically for NVIDIA Jetson edge devices (JetPack 6.2, JetPack 7, Ubuntu 24.04, ARM SBSA, etc.).
• The repository ships a catalog of “packages” – each package is a Dockerfile fragment that adds a particular framework (PyTorch, TensorFlow, ONNX Runtime, DeepStream, ROS 2, LLM toolkits, vision models, speech engines, robotics simulators, etc.) and its dependencies compiled for the Jetson’s CUDA version.
• Helper scripts (jetson-containers) automate pulling, building, and running these images, handling NVIDIA runtime flags, device mounts, and cache servers automatically.

Key capabilities

Area	Example packages (shown in the README)
Core ML	pytorch, tensorflow, jax, onnxruntime
Large Language Models	sglang, vllm, llama.cpp, transformers, bitsandbytes, deepspeed
Vision / Multimodal	llava, NanoOWL, Segment Anything (SAM), clip_trt
Retrieval‑Augmented Generation	langchain, llama-index, FAISS, NanoDB
Robotics & Simulation	ROS, Isaac Sim, Habitat Sim, MuJoCo, OpenVLA
Diffusion & Graphics	Stable Diffusion WebUI, ComfyUI, nerfstudio, gsplat
Speech & Audio	whisper, faster-whisper, piper, riva-client
Home/IoT	homeassistant-core, wyoming-whisper
Miscellaneous	mamba, pykan, xltsm, agents (OpenClaw, OpenFang)

How you use it

1. Install the helper tools

   git clone https://github.com/dusty-nv/jetson-containers
   bash jetson-containers/install.sh   # add --pypi to start a local devpi cache
   

2. Pick the packages you need – the packages directory lists every available component.
3. Build a custom image (the tool resolves compatible base images, pulls or compiles wheels, and caches them):

   jetson-containers build --name=my_robot pytorch transformers ros:humble-desktop
   

4. Run it – a thin wrapper around docker run adds --runtime nvidia, mounts a /data cache, and forwards devices:

   jetson-containers run $(autotag l4t-pytorch)
   

   autotag automatically selects the correct image tag for the JetPack/L4T version on your board.
5. Swap CUDA versions – set CUDA_VERSION=12.6 (or 13.x) before building to target a different toolkit; you can also pin cuDNN, TensorRT, Python, or PyTorch versions via environment variables.

Why it matters

• Edge‑first: Jetson devices have limited compute and a unique ARM‑based CUDA stack. Compiling the latest AI frameworks from source on‑device is painful; these containers provide pre‑built, CUDA‑compatible wheels.
• Modular: You can mix‑and‑match any combination (e.g., ROS 2 + PyTorch + Transformers) without manually resolving binary compatibility.
• Cache‑aware: An optional local devpi/apt cache dramatically speeds up repeated builds and avoids hitting external mirrors.
• Broad coverage: From low‑level numeric libraries (cupy, numba) to full‑stack generative AI (LLM, VLM, diffusion) and robotics simulators, the catalog aims to be a one‑stop shop for Jetson AI development.

Who it’s for

• Robotics engineers building perception or planning pipelines on Jetson‑Orin, Nano, Xavier, etc.
• Researchers prototyping LLM/VLM inference at the edge.
• Developers who need a reproducible environment for computer‑vision, speech, or reinforcement‑learning workloads on ARM‑CUDA hardware.
• Educators & hobbyists who want a quick start without wrestling with cross‑compilation.

Where to learn more

• Full package list: https://github.com/dusty-nv/jetson-containers/tree/main/packages
• Docs: System Setup, Building Containers, Running Containers (linked in the README)
• Tutorials & demos: Jetson Generative AI Lab site – https://www.jetson-ai-lab.com
• Community: Discord invite in the README (discord.gg/BmqNSK4886).

---

All information above is taken directly from the repository’s README; no external assumptions have been added.