Jetson Xavier NX storage problem

Hi there!

A time ago i got a Jetson Xavier NX (reComputer Jetson 20-1-H2 J202 Carrier Board). and installed JetPack 4.6.1 along with some useful libraries like pip3, docker-compose and jtop. My purpose is to run a container using a PyTorch script i have and allow the GPU of the device.

After JetPack installation, i remained with 500MBs of free space, so i followed the post (clean storage jetson nano), dropping to 3,8GBs of free space available. Finally, i tried to pull a image from l4t pytorch containers, but this image takes more than the free space i had.

I searched on forums and find: jetson jetpack installation memory issue, but the hardware im using only supports USB (J202 doesnt have a SD card slot). I bought a 32GBs pendrive and managed to boot from there, however, USB device is too slow to use.

Also i found a post regarding saving docker images on USB device (post), but i still want to only use eMMC storage because of his speed.

My questions are:

  1. Is it necessary to install JetPack to use GPU through containers and run a PyTorch script?
  2. There is a way to only install some JetPack libraries to clear more space, without PyTorch. (My goal is to run this script, maybe isn’t necessary CUDA)
  3. It is possible to use images from USB without speed issues if i have limited storage? Is this the way it was intended?
  4. It is necessary to use the image showed in l4t pytorch containers to run what i need?

Regards and happy coding.

Hi,

1. It depends.
On JetPack 4, the container mounts libraries from the Jetson natively.
So you will need to install the JetPack components to get these libraries.

But from JetPack 5, we include all the required libraries into the container directly.

2. Yes. you can.
You can install it via the OTA feature, which allows you to install the specific component based on your requirement.
https://docs.nvidia.com/jetson/archives/r35.1/DeveloperGuide/text/SD/SoftwarePackagesAndTheUpdateMechanism.html#over-the-air-update

3. It is bounded by the bandwidth and speed of the USB and external storage.

4. Yes.

Thanks.

Thanks @AastaLLL for the quick reply.

I have a few questions about what you said.

Just to make it clear… In order to do what i explain before, i could use a Pytorch based container on Jetson using Nvidia GPU without fall into storage problems by using the OTA feature. Am i wrong with this?

  1. If i customized the JetPack SDK, Which packages should i removed in order to run pytorch with minor dependencies?
  2. Should i used image-based OTA or Deb package management-based OTA?

Furthermore, im confused with the procedure of the OTA feature. Its supposed to update to new minor releases, not for uninstalling SDK packages.

Regards.

Hi,

1. Do you run PyTorch with CPU only?
If GPU mode is required, you will need to install CUDA, cuDNN and maybe TensorRT (based on your use case).

2. For example, if you want to run the PyTorch nvcr.io/nvidia/l4t-pytorch:l4t-pytorch:r32.7.1-pth1.10-py3 which supports GPU mode.

  1. Reflash your device with JetPack 4.6.2 without installing the componets.

  2. Install CUDA, cuDNN and TensorRT via OTA.

$ sudo apt update
$ sudo apt install nvidia-tensorrt
  1. Run the docker which has PyTorch preinstalled.
sudo docker run -it --rm --runtime nvidia --network host nvcr.io/nvidia/l4t-pytorch:l4t-pytorch:r32.7.1-pth1.10-py3

Thanks.

Thanks again @AastaLLL.

That make more sense, i’ll do what you suggest.

Regards.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.