Jetson Orin Wheels

Pre-built Python wheels for NVIDIA Jetson Orin devices on ARM64 / aarch64.

中文说明请见 README.zh-CN.md.

Releases

Download wheels from the GitHub Releases page:

https://github.com/Shattered217/Jetson-Orin-Wheels/releases

Do not mix wheels across JetPack or Python versions. The cp310 wheels are for Python 3.10, and the cp312 wheels are for Python 3.12.

Make sure uv is already installed. See the official installation guide: https://docs.astral.sh/uv/getting-started/installation/

Verified Environments

Field	JetPack 7.2.0	JetPack 6.2.1 rc1
Host used for verification	Jetson-Orin-NX	Jetson-Orin-Nano
Device	NVIDIA Jetson Orin NX Engineering Reference Developer Kit	NVIDIA Jetson Orin Nano Engineering Reference Developer Kit Super
OS	Ubuntu 24.04.4 LTS	Ubuntu 22.04.5 LTS
Jetson Linux / L4T	R39 revision 2.0 / nvidia-l4t-core 39.2.0	R36 revision 4.7 / nvidia-l4t-core 36.4.7
CUDA Toolkit	13.2.1	12.6.11
cuDNN	9.20.0.46 for CUDA 13	9.3.0.75 for CUDA 12
Python	3.12.3	3.10.12
System TensorRT	10.16.2.10	10.3.0
System OpenCV	5.0.0	4.11.0

JetPack 7.2.0

Release: https://github.com/Shattered217/Jetson-Orin-Wheels/releases/tag/7.2.0

Package Matrix

Package	Source	Version	Notes
PyTorch	torch-2.12.0-cp312-cp312-linux_aarch64.whl	2.12.0	Python 3.12 / aarch64 wheel.
TorchVision	torchvision-0.27.0+78839c2-cp312-cp312-linux_aarch64.whl	0.27.0	Matches the PyTorch wheel above.
ONNX Runtime GPU	onnxruntime_gpu-1.28.0-cp312-cp312-linux_aarch64.whl	1.28.0	Keep numpy<2 pinned.
TensorRT	System Python package	10.16.2.10	Reuse through --system-site-packages.
OpenCV	System Python package	5.0.0	Reuse through --system-site-packages.

Create Environment

JetPack 7 reuses the system TensorRT and OpenCV packages through --system-site-packages.

uv venv .venv --python 3.12 --system-site-packages
source .venv/bin/activate

Install Wheels

Install the JP7 wheels:

uv pip install \
  "https://github.com/Shattered217/Jetson-Orin-Wheels/releases/download/7.2.0/torch-2.12.0-cp312-cp312-linux_aarch64.whl" \
  "https://github.com/Shattered217/Jetson-Orin-Wheels/releases/download/7.2.0/torchvision-0.27.0%2B78839c2-cp312-cp312-linux_aarch64.whl" \
  "https://github.com/Shattered217/Jetson-Orin-Wheels/releases/download/7.2.0/onnxruntime_gpu-1.28.0-cp312-cp312-linux_aarch64.whl" \
  "numpy<2"

Verify

python - <<'PY'
import torch
import torchvision
import onnxruntime as ort
import tensorrt as trt
import cv2

print("torch:", torch.__version__)
print("torch cuda:", torch.cuda.is_available())
print("torchvision:", torchvision.__version__)
print("onnxruntime providers:", ort.get_available_providers())
print("tensorrt:", trt.__version__)
print("opencv:", cv2.__version__)
print("opencv cuda devices:", cv2.cuda.getCudaEnabledDeviceCount())
PY

JetPack 6.x

Release: https://github.com/Shattered217/Jetson-Orin-Wheels/releases/tag/6.2.1rc1

These wheels were verified on JetPack 6.2.1 rc1. Compatibility across other JetPack 6.x minor versions is not guaranteed.

Package Matrix

Package	Source	Version	Notes
PyTorch	torch-2.3.0a0+git97ff6cf-cp310-cp310-linux_aarch64.whl	2.3.0a0	Python 3.10 / aarch64 wheel.
TorchVision	torchvision-0.18.0-cp310-cp310-linux_aarch64.whl	0.18.0	Matches the PyTorch wheel above.
Torchaudio	torchaudio-0.18.0-cp310-cp310-linux_aarch64.whl	0.18.0	Minimal Jetson build.
ONNX Runtime GPU	onnxruntime_gpu-1.24.0-cp310-cp310-linux_aarch64.whl	1.24.0	CUDA / TensorRT-enabled GPU build.
TensorRT	System Python package	10.3.0	Reuse through --system-site-packages.
OpenCV	System Python package	4.11.0	Reuse through --system-site-packages.

Create Environment

For JetPack 6, using --system-site-packages to reuse the system TensorRT and OpenCV packages is recommended first. If you prefer a fully wheel-based setup, you can manually install the TensorRT and OpenCV wheels from the release page.

uv venv .venv --python 3.10 --system-site-packages
source .venv/bin/activate

Install Wheels

Install the JP6 wheels:

uv pip install \
  "https://github.com/Shattered217/Jetson-Orin-Wheels/releases/download/6.2.1rc1/torch-2.3.0a0%2Bgit97ff6cf-cp310-cp310-linux_aarch64.whl" \
  "https://github.com/Shattered217/Jetson-Orin-Wheels/releases/download/6.2.1rc1/torchvision-0.18.0-cp310-cp310-linux_aarch64.whl" \
  "https://github.com/Shattered217/Jetson-Orin-Wheels/releases/download/6.2.1rc1/torchaudio-0.18.0-cp310-cp310-linux_aarch64.whl" \
  "https://github.com/Shattered217/Jetson-Orin-Wheels/releases/download/6.2.1rc1/onnxruntime_gpu-1.24.0-cp310-cp310-linux_aarch64.whl" \
  "numpy<2"

Verify

python - <<'PY'
import torch
import torchvision
import torchaudio
import onnxruntime as ort
import tensorrt as trt
import cv2

print("torch:", torch.__version__)
print("torch cuda:", torch.cuda.is_available())
print("torchvision:", torchvision.__version__)
print("torchaudio:", torchaudio.__version__)
print("onnxruntime providers:", ort.get_available_providers())
print("tensorrt:", trt.__version__)
print("opencv:", cv2.__version__)
print("opencv cuda devices:", cv2.cuda.getCudaEnabledDeviceCount())
PY

YOLO Smoke Test

After installing PyTorch and OpenCV, Ultralytics YOLO can be checked with:

uv pip install ultralytics
yolo check

The output should show the Jetson GPU, CUDA availability, PyTorch, TorchVision, NumPy, and OpenCV versions.

Troubleshooting

• Wrong Python ABI: cp310 wheels require Python 3.10; cp312 wheels require Python 3.12.
• Wrong JetPack line: Install wheels from the release tag matching your JetPack version.
• NumPy compatibility: Keep numpy<2 pinned when using the ONNX Runtime GPU wheels.
• TensorRT / OpenCV imports fail: Recreate the environment with --system-site-packages.
• Multiple Python environments: Run which python, python --version, and python -m pip --version before installing.

References

• Jetson Orin Wheels releases
• Jetson Python environment: reuse system packages with uv --system-site-packages
• PyTorch for Jetson
• TensorRT Python package source