Jetpack 7.2. Jetson Orin Nano Super (8gb) Pytorch for Compute Capability 8.7

ivan_zzz · August 10, 2026, 5:06pm

Hi.

I have Jetson Orin Nano Reference Developer Kit Super (8gb)

Installed Jetpack 7.2.

Tried to compile ENGINE file from .pt and checked versions of installed software:

python3 -c “import torch; print(torch.version, torch.version.cuda, torch.cuda.is_available(), torch.cuda.get_device_name(0))”

I got gollowing message with WARNING:
/home/jetson1/.local/lib/python3.12/site-packages/torch/cuda/init.py:422: UserWarning: Found GPU0 Orin which is of compute capability (CC) 8.7.
The following list shows the CCs this version of PyTorch was built for and the hardware CCs it supports:

8.0 which supports hardware CC >=8.0,<9.0 except {8.7}
9.0 which supports hardware CC >=9.0,<10.0
10.0 which supports hardware CC >=10.0,<11.0 except {10.1}
11.0 which supports hardware CC >=11.0,<12.0
12.0 which supports hardware CC >=12.0,<13.0
No published PyTorch CUDA builds for release 2.13.0+cu132 support this GPU. Visit https://pytorch.org/get-started/locally/ to find a compatible release.

So I cant install correct version of pytorch for Jetpack 7.2 (jetson orin nano super).

Where to find correct version and how to upgrade it properly?

Tried to find here: Start Locally | PyTorch (<https://docs.pytorch.org/get-started/locally/>)

But still not sure which one to install to get full compatibility with Compute Capability 8.7.

Checked current libs:

python3 -c “import sys, cv2, torch, torchvision, ultralytics; print(f’Python: {sys.version.split()[0]}\nPyTorch: {torch.version}\nTorchvision: {torchvision.version}\nUltralytics (YOLO): {ultralytics.version}\nOpenCV: {cv2.version}')” && gst-inspect-1.0 --version | head -n 1

python3 -c “import cv2; print(‘Версия OpenCV:’, cv2.version); print(cv2.getBuildInformation())” | grep -E “GStreamer|NVIDIA|CUDA”

pip list | grep -E “opencv|ultralytics|torch|gpio”

/usr/local/cuda/bin/nvcc --version

dpkg -l | grep -E “opencv|gstreamer|v4l2|gpiod”

ii gir1.2-gstreamer-1.0:arm64 ii gpiod ii gstreamer1.0-alsa:arm64 ii gstreamer1.0-gl:arm64 ii gstreamer1.0-gtk3:arm64 ii gstreamer1.0-libcamera:arm64 ii gstreamer1.0-packagekit ii gstreamer1.0-pipewire:arm64 ii gstreamer1.0-plugins-bad:arm64 ii gstreamer1.0-plugins-base:arm64 ii gstreamer1.0-plugins-base-apps ii gstreamer1.0-plugins-good:arm64 ii gstreamer1.0-plugins-ugly:arm64 ii gstreamer1.0-tools ii gstreamer1.0-x:arm64 ii libgpiod-dev:arm64 ii libgpiod2t64:arm64 ii libgstreamer-gl1.0-0:arm64 ii libgstreamer-opencv1.0-0:arm64 ii libgstreamer-plugins-bad1.0-0:arm64 ii libgstreamer-plugins-base1.0-0:arm64 ii libgstreamer-plugins-base1.0-dev ii libgstreamer-plugins-good1.0-0:arm64 ii libgstreamer-plugins-good1.0-dev ii libgstreamer1.0-0:arm64 ii libgstreamer1.0-dev:arm64 ii libgtk-4-media-gstreamer ii libopencv ii libopencv-calib3d406t64:arm64 ii libopencv-contrib406t64:arm64 ii libopencv-core406t64:arm64 ii libopencv-dev ii libopencv-dnn406t64:arm64 ii libopencv-features2d406t64:arm64 ii libopencv-flann406t64:arm64 ii libopencv-highgui406t64:arm64 ii libopencv-imgcodecs406t64:arm64 ii libopencv-imgproc406t64:arm64 ii libopencv-ml406t64:arm64 ii libopencv-objdetect406t64:arm64 ii libopencv-photo406t64:arm64 ii libopencv-python ii libopencv-samples ii libopencv-shape406t64:arm64 ii libopencv-stitching406t64:arm64 ii libopencv-video406t64:arm64 ii libopencv-videoio406t64:arm64 ii libopencv-viz406t64:arm64 ii libv4l2rds0t64:arm64 ii nvidia-l4t-gstreamer ii nvidia-opencv ii nvidia-opencv-dev ii opencv-licenses ii opencv-samples-data ii python3-opencv:arm64 1.24.2-1ubuntu0.1 arm64 GObject introspection data for the GStreamer library
1.6.3-1.1build1 arm64 Tools for interacting with Linux GPIO character device - binary
1.24.2-1ubuntu0.4 arm64 GStreamer plugin for ALSA
1.24.2-1ubuntu0.4 arm64 GStreamer plugins for GL
1.24.2-1ubuntu1.5 arm64 GStreamer plugin for GTK+3
0.2.0-3fakesync1build6 arm64 complex camera support library (GStreamer plugin)
1.2.8-2ubuntu1.5 arm64 GStreamer plugin to install codecs using PackageKit
1.0.5-1ubuntu3.3 arm64 GStreamer 1.0 plugin for the PipeWire multimedia server
1.24.2-1ubuntu4 arm64 GStreamer plugins from the “bad” set
1.24.2-1ubuntu0.4 arm64 GStreamer plugins from the “base” set
1.24.2-1ubuntu0.4 arm64 GStreamer helper programs from the “base” set
1.24.2-1ubuntu1.5 arm64 GStreamer plugins from the “good” set
1.24.1-1build1 arm64 GStreamer plugins from the “ugly” set
1.24.2-1ubuntu0.1 arm64 Tools for use with GStreamer
1.24.2-1ubuntu0.4 arm64 GStreamer plugins for X11 and Pango
1.6.3-1.1build1 arm64 C library for interacting with Linux GPIO device - static libraries and headers
1.6.3-1.1build1 arm64 C library for interacting with Linux GPIO device - shared libraries
1.24.2-1ubuntu0.4 arm64 GStreamer GL libraries
1.24.2-1ubuntu4 arm64 GStreamer OpenCV libraries
1.24.2-1ubuntu4 arm64 GStreamer libraries from the “bad” set
1.24.2-1ubuntu0.4 arm64 GStreamer libraries from the “base” set
1.24.2-1ubuntu0.4 arm64 GStreamer development files for libraries from the “base” set
1.24.2-1ubuntu1.5 arm64 GStreamer development files for libraries from the “good” set
1.24.2-1ubuntu1.5 arm64 GStreamer development files for libraries from the “good” set
1.24.2-1ubuntu0.1 arm64 Core GStreamer libraries and elements
1.24.2-1ubuntu0.1 arm64 GStreamer core development files
4.14.5+ds-0ubuntu0.10 arm64 GStreamer media backend for the GTK graphical user interface library
4.8.0-3-g6ef37b4 arm64 Open Computer Vision Library
4.6.0+dfsg-13.1ubuntu1 arm64 computer vision Camera Calibration library
4.6.0+dfsg-13.1ubuntu1 arm64 computer vision contrlib library
4.6.0+dfsg-13.1ubuntu1 arm64 computer vision core library
4.8.0-3-g6ef37b4 arm64 Open Computer Vision Library
4.6.0+dfsg-13.1ubuntu1 arm64 computer vision Deep neural network module
4.6.0+dfsg-13.1ubuntu1 arm64 computer vision Feature Detection and Descriptor Extraction library
4.6.0+dfsg-13.1ubuntu1 arm64 computer vision Clustering and Search in Multi-Dimensional spaces library
4.6.0+dfsg-13.1ubuntu1 arm64 computer vision High-level GUI and Media I/O library
4.6.0+dfsg-13.1ubuntu1 arm64 computer vision Image Codecs library
4.6.0+dfsg-13.1ubuntu1 arm64 computer vision Image Processing library
4.6.0+dfsg-13.1ubuntu1 arm64 computer vision Machine Learning library
4.6.0+dfsg-13.1ubuntu1 arm64 computer vision Object Detection library
4.6.0+dfsg-13.1ubuntu1 arm64 computer vision computational photography library
4.8.0-3-g6ef37b4 arm64 Open Computer Vision Library
4.8.0-3-g6ef37b4 arm64 Open Computer Vision Library
4.6.0+dfsg-13.1ubuntu1 arm64 computer vision shape descriptors and matchers library
4.6.0+dfsg-13.1ubuntu1 arm64 computer vision image stitching library
4.6.0+dfsg-13.1ubuntu1 arm64 computer vision Video analysis library
4.6.0+dfsg-13.1ubuntu1 arm64 computer vision Video I/O library
4.6.0+dfsg-13.1ubuntu1 arm64 computer vision 3D data visualization library
1.26.1-4build3 arm64 Video4Linux Radio Data System (RDS) decoding library
39.2.0-20260601141651 arm64 NVIDIA GST Application files
7.2-b187 arm64 NVIDIA OpenCV Meta Package
7.2-b187 arm64 NVIDIA OpenCV dev Meta Package
4.8.0-3-g6ef37b4 arm64 Open Computer Vision Library
4.8.0-3-g6ef37b4 arm64 Open Computer Vision Library
4.6.0+dfsg-13.1ubuntu1 arm64 Python 3 bindings for the computer vision library

Now I cant even uninstall current version (not mentioning about lack of knowledge which correct link or command to use for install…

ivan_zzz · August 10, 2026, 6:09pm

Found some link with versions, but maximum Jetpack 7.1 supported for now… https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform-release-notes/pytorch-jetson-rel.html>

AastaLLL · August 11, 2026, 2:58am

Hi,

This message is a harmless warning:

/home/jetson1/.local/lib/python3.12/site-packages/torch/cuda/init.py:422: UserWarning: Found GPU0 Orin which is of compute capability (CC) 8.7.
The following list shows the CCs this version of PyTorch was built for and the hardware CCs it supports:

You can install upstream PyTorch on JetPack 7.2+Orin with the below command:

$ pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu132

Below is the command for ignoring the warning for your reference:

import warnings
warnings.filterwarnings("ignore", message=".*Found GPU.*compute capability.*")

Please find the comment below for details:

Thanks.

ivan_zzz · August 12, 2026, 9:03pm

ISNt it for x86_64 on that link ? I have Jetson Orin Nano board. Will it fit?
And also which Numpy and Gstreamer versions work together?

OpenCV 4.8.0 ↔ NumPy 1.26.4 ↔ CuPy 13.x ↔ TensorRT 10.16 ↔ PyTorch 2.5/2.6 ?

OpenCV 4.8.0 ↔ NumPy ??? ↔ CuPy 13.x ↔ TensorRT 10.16 ↔ PyTorch ???

AastaLLL · August 13, 2026, 4:55am

Hi,

It will install the aarch64 packages from the https://download.pytorch.org/whl/cu132.

And also which Numpy and Gstreamer versions work together?
Is this question for Deepstream? There is no dependency between TensorRT and PyTorch.

Thanks.

ivan_zzz · August 13, 2026, 2:40pm

Thanks. Just trying to restore original Jetpack 7.2 versions after occasional updating libraries in main system with flag “–break-system-packages” :) Want to restore original list, to install the rest (non-standard) into the venv. Want to use native NVIDIA utils for the jetpack 7.2. is there any easy way to restore the package without flashing all SSD?

ivan_zzz · August 13, 2026, 5:07pm

Why I asked about NumPy - there is a message shown in my script running inside the venv. Tthe same message when I check by the command python3 -c “import cv2; print(cv2.version)” outside the venv…

even when I run such script inside the venv (with checked OpenCV = 5.0.0)

So maybe I compiled some .Engine file on the old 1.xx version (apparently manually installed into the main system), and after System updating so many messages appeared… ? Is it NORMAL situation with the libraries as I revised them below:

What should I do to bring things into order?
thx

ivan_zzz · August 15, 2026, 11:34am

hi again. I have Jetson Orin Nano Super (8gb) board,

I installed pytorch to VENV from there: $ pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu132

BUT should I install it with some flag for Orin (sm_87)?

Many times to compile .engine file inside .venv
resulted :

ONNX: starting export with onnx 1.22.0 opset 18…
ONNX: slimming with onnxslim 0.1.95…
ONNX: export success ✅ 7.7s, saved as ‘yolo11n.onnx’ (10.2 MB)

TensorRT: starting export with TensorRT 10.16.2.10…

TensorRT: input “images” with shape(1, 3, 640, 640) DataType.FLOAT
TensorRT: output “output0” with shape(1, 84, 8400) DataType.FLOAT
TensorRT: building FP16 engine as yolo11n.engine

TensorRT: export success ✅ 418.5s, saved as ‘yolo11n.engine’ (8.1 MB)

Export complete (419.8s)
Results saved to /home/jetson1/Desktop/yolo11/yolo11n.engine

(Warm-up)…
Loading yolo11n.engine for TensorRT inference…

[TRT] [I] Loaded engine size: 8 MiB

[TRT] [I] [MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +0, GPU +9, now: CPU 0, GPU 14 (MiB)

(FPS): 86.06

Looks OK right? BUT NO ! if I try to get metadata from that .engine file or open it with

/usr/src/tensorrt/bin/trtexec --loadEngine=yolo11n.engine
&&& RUNNING TensorRT.trtexec [TensorRT v101602] [b10]

&&&& FAILED TensorRT.trtexec [TensorRT v101602] [b10] # /usr/src/tensorrt/bin/trtexec --loadEngine=yolo11n.engine

Where to get Pytorch for Jetson orin nano board (Jetpack 7.2) (sm_87)?

Or I need to use that WHL from https://download.pytorch.org/whl/torch/

?

Or I need to make adopted lib from “torch-2.5.1-cp312-cp312-linux_aarch64.whl” directly for my board ?

Or how to compile .engine files to keep correct working through native trtexec ?

thx

ivan_zzz · August 16, 2026, 3:32pm

I reinstalled jetpack completely, created venv environment and tried to install pytorch there according to your instructions: run from inside

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu132

Some packages installed but errors in the end:

Requirement already satisfied: six in /usr/lib/python3/dist-packages (from torchvision) (1.16.0)
Downloading https://download-r2.pytorch.org/whl/torchvision-0.1.6-py3-none-any.whl (16 kB)
ERROR: Cannot install torch and torchaudio==2.2.0 because these package versions have conflicting dependencies.

The conflict is caused by:
The user requested torch
torchaudio 2.2.0 depends on torch==2.2.0

To fix this you could try to:

loosen the range of package versions you’ve specified
remove package versions to allow pip attempt to solve the dependency conflict

ERROR: ResolutionImpossible: for help visit Dependency Resolution - pip documentation v26.3.dev0

So do I need to find another version matched with my JetPack 7.2 (Python 3.12, CUDA 13.2)?

maybe

torch-2.5.1+nv24.11-cp312-cp312-linux_aarch64.whl ?
torchvision-0.20.1+nv24.11-cp312-cp312-linux_aarch64.whl ?
torchaudio-2.5.1+nv24.11-cp312-cp312-linux_aarch64.whl ?

Is there a correct link there or need some flags? please send full command to be sure.

thx

ivan_zzz · August 16, 2026, 4:53pm

there is temporary solution found without making own wheels.

pip install torch==2.12.1+cu132 torchvision==0.27.1+cu132 torchaudio --pre --index-url https://download.pytorch.org/whl/cu132

uses pre-release version, installs with flags (inside VENV just not to break native system packages). Kind of that…

AastaLLL · August 17, 2026, 6:45am

Hi,

Sorry for the late update.

BUT should I install it with some flag for Orin (sm_87)?

No, you don’t need to. The upstream pacakge contains sm_80 that can work on Orin already.

torchaudio 2.2.0 depends on torch==2.2.0

It looks like the installer is finding an old package so trigger this error (there is no 2.2.0+python3.12 package).
Is there any package insider your virtual environment depends on torch 2.2.0?
Is the virtual environment created with Python 3.12?

Thanks.

ivan_zzz · August 17, 2026, 4:19pm

Here is full review what the CLEAN os Ubuntu 24.04 +jetpack 7.2 from the USB contained initially and what had been installed into the VENV.

I spent several days jut to wind how to get it works! Why native jetpack downt contain the right versions out of the box? And why no proper wheels without breaking system flags usage?

This time I just reinstalled full OS+jetpack from scratch, than was trying to install the rest libs I need (pytorch, torchvision, Simpy, Onnx, ultralytics, Opencv).
Created VENV (because as you know, its impossible to install all these libs without breack-system-package flags into the main system), so I decided to keep it safe and install all the rest into the venv. For normal use NATIVE OPENCV and GSTREAMER with hardware GPU support.

If I used standard link you gave me (just pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu132 without flags), some issues appeared with coordination between NATIVE system versions and installed pytorch components versions. Thats why I installed with that flag, to avoid conflict of the libs, and than also downgraded numpy from 2.5.2 to 1.26.4 + opencv-python from 5.0.0.93 to 4.8.0

There is the list what was initially in system and after installed VENV with another versions.

	OS initial	installed into VENV (works)
OS	L4T R39.2.1	L4T R39.2.1
L4T	Nvidia-l4t-core 39.2.1-20260806224157	Nvidia-l4t-core 39.2.1-20260806224157
Jetpack	7.2.1-b49	7.2.1-b49

Machine	aarch64	aarch64
System	Linux	Linux
Distribution	Ubuntu 24.04	Ubuntu 24.04
Release	6.8.12-1021 tegra	6.8.12-1021 tegra
CUDA	13.2.86	13.2.86
cuDNN	_9.20.0	_9.20.0
TensorRT	10.16.2.10	10.16.2.10
VPI	4.1.4.0	4.1.4.0
Vulkan	1.4.321	1.4.321
OpenCV	4.8.0. with cuda NO	4.8.0. with cuda NO
P-number	p3767-0005	p3767-0005
Soc	Tegra 234	Tegra 234
L4T	L4T R39.2.1	L4T R39.2.1
Jetpack	7.2	7.2
Cuda Arch Bin	8.7	8.7
python –version	Python 3.12.3	Python 3.12.3
Nvidia-smi	NVIDIA-SMI 595.78	NVIDIA-SMI 595.78
	Driver Version: 595.78	Driver Version: 595.78
	CUDA Version: 13.2	CUDA Version: 13.2
pip3 –version	pip 24.0 from /usr/lib/python3/dist-packages/pip (python 3.12)	pip 26.2.1 from /home/jetson1/venv1/lib/python3.12/site-packages/pip (python 3.12)
libjpeg-dev:arm64	8c-2ubuntu11	0.3.26+ds-1ubuntu0.1
libopenblas-dev:arm64	0.3.26+ds-1ubuntu0.1	0.3.26+ds-1ubuntu0.1
python3-pip	24.0+dfsg-1ubuntu1.3	24.0+dfsg-1ubuntu1.3
zlib1g:arm64	1:1.3.dfsg-3.1ubuntu2.1	1:1.3.dfsg-3.1ubuntu2.1
zlib1g-dev:arm64	1:1.3.dfsg-3.1ubuntu2.1	1:1.3.dfsg-3.1ubuntu2.1
Libopenblas-dev	not installed	0.3.26+ds-1ubuntu0.1
numpy	1.26.4	numpy (2.5.2) — versions conflict possible (2.5.2) downgraded to native 1.26.4
SciPy	1.11.4	1.11.4
simpy	not installed	4.1.2
onnx	not installed	1.22.0
ultralytics	not installed	8.4.120
Ultralytics-platform	not installed	0.1.5
Ultralytics-platform	not installed	2.1.6
opencv-python	not installed	Opencv-python (5.0.0.93) — performance issue possible 5.0.0.93 downgraded to 4.8.0
	4.8.0	4.8.0
gstreamer	1.24.2	1.24.2
tensorrt	10.16.2.10-1+cuda13.2	10.16.2.10-1+cuda13.2
opencv	4.8.0-4-g18251aa	4.8.0-4-g18251aa
numpy	1.26.4	uses global ?
scipy	1.11.4	uses global ?
PyTorch	not installed	2.12.1+cu132
Torchvision	not installed	2.12.1+cu132
torchaudio	not installed	2.11.0+cpu

Please tell me, is it OK? WHICH IS THE RIGHT WAY to get a proper versions of working Pytorch + ultralytics package without docker and without breaking system versions of the libs?

in curernt state all works without errors. But Im still thinking whether it worth using OpenCV 5.0 and numpy 2.5.2 versions (without downgrading) or maybe some MAIN os upgrade should be dome just to fit newer versions of the native wheels for Jetson Orin Nano (jetpack 7.2) ?

ivan_zzz · August 17, 2026, 7:01pm

! significant note. If you run heavy engine/opencv tasks like creating .engine files with multi-batch parameters, ultralyticsupdates automatically and dragged with it NumPy 2.5.2 back !!!

native PyTorch for Jetson (NVIDIA) included into OpenCV (cv2) compiled by old good NumPy 1.x. [1, 2, 3]

Running pencv-python you leaned Python- OpenCV to overwrite NumPy up to 2.5.2.

It breakes the harmony and crashes your inference…

To avoid that just LOCK NumPy at stable 1.x. version:

pip uninstall -y opencv-python opencv-python-headless
pip install “numpy==1.26.4” --force-reinstall
python3 -m venv /home/jetson1/venv1 --system-site-packages
pip freeze > /home/jetson1/Desktop/models/requirements.txt

in the future install applications throught PIP with –no-deps
like this: pip install blabla-something --no-deps

It might help.

AastaLLL · August 28, 2026, 6:03am

Hi,

Do you get a working environment now?

To run Ultralytics software, you can check their installation steps (native/container) below:

Thanks.

Y-T-G · August 28, 2026, 6:18pm

Ultralytics has guide for Ultralytics installation and PyTorch installation on JetPack 7.2 linked above.

You just need to follow that. Except don’t use [export]. Just pip install ultralytics

Topic		Replies	Views
Install Pytorch with cuda on Jetson Orin nano Devloper Kit Jetson Orin Nano cuda , pytorch	12	4863	July 8, 2024
Pytorch & torchversion compatible issue on L4T35.5.0 Jetson Orin Nano pytorch	19	1757	November 7, 2024
Pytorch compatibility issues (torch 2.0.0+nv23.5 && torchvision 0.15.1) Jetson Orin NX pytorch	9	19094	June 13, 2023
Hello AI World for Jetpack 6.0 DP - Pytorch 2.1.0 Installed, Torchvision Did Not Install Jetson Orin Nano pytorch	23	2521	January 15, 2024
Cannot get Torch met CUDA to work on Jetson Orin Jetson Orin NX pytorch	9	3087	February 26, 2024
[REQUEST] build script for pytorch or up to date pytorh binary release supporting jetson boards running L4T35.6(ubuntu20.04) Jetson Orin Nano pytorch	19	1008	January 5, 2025
Help me with Correct Pytorch and Torchvision versions requirement for Jetpack 6.2.1 orin super Jetson Orin Nano pytorch , cusparse , cudss , jetson-orin	59	4424	October 17, 2025
PyTorch for Jetson Announcements pytorch , python , kb	1366	845810	May 14, 2024
NO WAY TO USE GPU - TORCH NOT DETECTING GPU - CUDA 12.6 - JP 6.1 ORIN NANO Jetson Orin Nano pytorch	4	201	March 9, 2026
One way to setup Jetson Orin Nano for OpenCV and Yolo with Cuda May 2026 Jetson Orin Nano opencv , cuda , yolo , cudnn , cublas , cufft , jetson , cudss	1	433	May 22, 2026

Jetpack 7.2. Jetson Orin Nano Super (8gb) Pytorch for Compute Capability 8.7

Related topics