Building a PyTorch wheel on the Jetson Nano takes forever, while setting up a cross-compilation build system is prone to many failures. Moreover the official wheels (right now) are built only for specific python versions (e.g. python3.6). This creates a bottleneck of workarounds, beginning with simple python environment management and ending with overly complicated containers deployed on the Jetson Nano.
As such, the target is to create a simple and manageable build environment to compile and natively deploy PyTorch for the python version of choice.
The approach taken here is to use a docker container for the desired target Jetson Nano OS running via QEMU that builds the actual package wheel on the host system. Thereby, avoiding cross-compilation package management.
-
Install docker by following this guide.
-
Install QEMU packages:
sudo apt-get install qemu binfmt-support qemu-user-static
- Execute the registering scripts:
docker run --rm --privileged multiarch/qemu-user-static --reset -p yes
- Test the QEMU emulation environment
sudo docker run --platform linux/arm64/v8 --rm -t arm64v8/ubuntu uname -m
the output should be # aarch64.
Checkout the References section for more information.
To run the build, input the following command:
docker buildx build --platform linux/arm/v8 --rm -t jetson-pythorch-build -o wheels .
BSD 3-Clause License