[SOLVED] Is there a pre-built wheel for PyTorch for Jetson Nano?

Is there a pre-built wheel for PyTorch for Jetson Nano?

I'm trying to install PyTorch with CUDA support on a Jetson Nano. It's running the Ubuntu 18.04 that the jetson-nano-jp461-sd-card-image image from Nvidia put on it with Python 3.6.9, CUDA 10.2, and JetPack 4.6.6. I tried:

pip3 install torch --index-url https://download.pytorch.org/whl/cpu

but this gives me the CPU-only version, and I need CUDA support for the Nano's GPU. I also tried

git clone --recursive --branch v1.8.0 https://github.com/pytorch/pytorch

and installing, but it is really, really, slow.

Is there a pre-built wheel available for this specific configuration? Many thanks in advance.

Solution

For those who may be searching, this worked for me:

sudo apt-get install python3-pip libopenblas-base libopenmpi-dev libomp-dev

pip3 install 'Cython<3'

pip3 install numpy

wget https://nvidia.box.com/shared/static/fjtbno0vpo676a25cgvuqc1wty0fkkg6.whl -O torch-1.10.0-cp36-cp36m-linux_aarch64.whl

sudo pip3 install torch-1.10.0-cp36-cp36m-linux_aarch64.whl