torch-install

name	torch-install
description	Install NVIDIA-optimized PyTorch with CUDA support on reComputer Jetson devices, covering JetPack 5.x and JetPack 6.x workflows, cuSPARSELt setup, torch/torchvision pairing, probing multiple Python or conda targets, and managing multiple PyTorch versions with Miniconda.

Install PyTorch for reComputer Jetson

Execution model

Run one phase at a time. After each phase, verify the expected result before continuing.

• If a phase succeeds → print [OK] and move to the next phase.
• If a phase fails → print [STOP], consult the failure decision tree, and ask the user before retrying.

Phase 1 — Verify prerequisites

cat /etc/nv_tegra_release
dpkg -l | grep nvidia-jetpack
python3 --version
conda env list || true
ping -c 2 developer.download.nvidia.com

Expected: JetPack version identified; Python version noted; optional conda envs listed; internet reachable.

If multiple compatible Python targets exist, choose the install target first:

• system Python, if the wheel ABI matches exactly
• an existing conda env with matching Python version
• otherwise create a new conda env with the required Python version

Always install with the selected interpreter's python -m pip, not raw pip / pip3.

Phase 2A — Install PyTorch on JetPack 5.1.3

sudo apt-get -y update
sudo apt-get install -y python3-pip libopenblas-dev

wget https://developer.download.nvidia.com/compute/redist/jp/v512/pytorch/torch-2.1.0a0+41361538.nv23.06-cp38-cp38-linux_aarch64.whl
pip install torch-2.1.0a0+41361538.nv23.06-cp38-cp38-linux_aarch64.whl

Verify:

import torch
print(torch.cuda.is_available())  # Should print True
print(torch.__version__)

Expected: True and torch version displayed.

Phase 2B — Install PyTorch on JetPack 6.2

sudo apt-get -y update
sudo apt-get install -y python3-pip libopenblas-dev

Install cuSPARSELt (required for PyTorch 24.06+):

wget https://developer.download.nvidia.com/compute/cusparselt/0.7.1/local_installers/cusparselt-local-tegra-repo-ubuntu2204-0.7.1_1.0-1_arm64.deb
sudo dpkg -i cusparselt-local-tegra-repo-ubuntu2204-0.7.1_1.0-1_arm64.deb
sudo cp /var/cusparselt-local-tegra-repo-ubuntu2204-0.7.1/cusparselt-*-keyring.gpg /usr/share/keyrings/
sudo apt-get update
sudo apt-get -y install libcusparselt0 libcusparselt-dev

Install PyTorch and matching TorchVision with the selected interpreter:

wget https://developer.download.nvidia.cn/compute/redist/jp/v61/pytorch/torch-2.5.0a0+872d972e41.nv24.08.17622132-cp310-cp310-linux_aarch64.whl
<selected_python> -m pip install torch-2.5.0a0+872d972e41.nv24.08.17622132-cp310-cp310-linux_aarch64.whl
<selected_python> -m pip install torchvision-0.20.0a0+afc54f7-cp310-cp310-linux_aarch64.whl

Verify:

import torch
print(torch.cuda.is_available())  # Should print True

Expected: True

Phase 2C — Install PyTorch for other JetPack versions

sudo apt-get -y update
sudo apt-get install -y python3-pip libopenblas-dev

Browse NVIDIA PyTorch wheels and download the wheel matching your JetPack and Python version. If multiple supported torch versions are available for the current JetPack release, present those options to the user before installing:

wget <wheel_url>
<selected_python> -m pip install <wheel_filename>.whl
<selected_python> -m pip install <matching_torchvision>.whl

Verify:

import torch
print(torch.cuda.is_available())

Phase 3 — Multiple PyTorch versions with Conda (optional)

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-aarch64.sh
bash Miniconda3-latest-Linux-aarch64.sh

Follow prompts: agree to terms, use default path, enter "yes" to add to .bashrc.

source ~/.bashrc
conda --version
conda config --set auto_activate_base false

Create environment and install a specific PyTorch version:

conda create -n torch_2.0 python=3.8
conda activate torch_2.0
# Then download and install the appropriate wheel as in Phase 2

Failure decision tree

Symptom	Likely cause	Suggested fix
torch.cuda.is_available() returns False	Wrong wheel for JetPack version	Verify JetPack version and download matching wheel
pip install fails with version conflict	Python version mismatch	Check the selected interpreter version matches the wheel tag (cp38/cp310)
libopenblas not found	Missing dependency	sudo apt-get install libopenblas-dev
cuSPARSELt install fails (JP6)	Wrong Ubuntu version or arch	Verify Ubuntu 22.04 aarch64 on JetPack 6
wget download fails	Network issue or URL changed	Check connectivity; browse NVIDIA redist page for updated URL
Conda not found after install	.bashrc not sourced	Run source ~/.bashrc or open new terminal

Reference files

• references/source.body.md — Full original wiki content (reference only)