Workbench is dead on DGX Spark

AI & Data Science NVIDIA AI Workbench
1

I got this DGX Spark and after the latest system update the workbench is dead! I did some trobleshooting and from GPT I got this: Bottom Line - Your system is ARM. The AppImage is x86_64. It will never execute. Nothing is broken.
is there a way to fix the problem and getting the workbench working again. It was running before, why not now?

Please tick the appropriate box to help us categorize your post
Bug or Error
Feature Request
Documentation Issue
Other

2

That is very strange because Workbench on Ubuntu moved away from AppImage a while ago, and I’m pretty sure that the install in the DGX OS is using the apt package.

I’m going to loop in an engineer, but a few questions first:

  1. Did you do any sort of manual uninstall/reinstall of Workbench on your Spark?
  2. Can you run cat /etc/dgx-release in a terminal on your Spark and paste the output here?
3

I got a similar error message about my platform being ARM64 rather than AMD64. In general I haven’t been able to get any of the NVidia examples’ containers to build. I was able to select a base container in Workbench, add a few packages and an HF key and got a GPT2 1.5G model from Raschka’s book to run on my DGX Spark. Initially inference took 22 sec but after a system reboot it dropped to under a second. When I try to add missing packages in a terminal I always get “this environment is externally managed message” so I am not able to create my own environment the way I did on my X86 Ubuntu PC. I guess I have to use the Workbench which did work for my own code development. I think the examples are just old, looks like 3 years and just haven’t been updated. correct me if I am wrong.

4

There’s a backlog item to go through the example projects and get them running on Sparks.

However, you should be able to start net-new projects on the spark using the appropriate based images in the new-project creation flow.

See here: Create, Clone, Publish and Delete Projects — NVIDIA AI Workbench User Guide

This CUDA 13 container w/pytorch is available as a default starting point: python-cuda130 | NVIDIA NGC

In addition, if you are trying to add packages to a Workbench container, you need to do it in the container, not on the actual system. The Ubuntu distro on the system is locked down in a variety of ways, and that’s why Workbench uses a container.

You can see how to add and manage packages in the Workbench project here ( Environments and AI Workbench — NVIDIA AI Workbench User Guide) and here ( Manage Packages — NVIDIA AI Workbench User Guide) and here ( Use the postBuild.bash Script — NVIDIA AI Workbench User Guide.

Let me know if this helps.