This issue applies to new installs of Holoscan SDK 0.1 after April 27, or existing installs which have never been switched to dGPU mode.
The issue: At this moment the nvgpuswitch.py script is broken when switching to dGPU mode. After running
sudo nvgpuswitch.py install dGPU and rebooting, the failures caused by using the old signing key will not be detected and will leave the system in an unusable state requiring a reset.
We are actively working on creating a fix for this issue. In the meanwhile, the preventative workaround for this issue is to run the following command on the developer kit before running nvgpuswitch.py to switch to dGPU mode:
$ sudo sed --in-place -e 's/7fa2af80\.pub/3bf863cc\.pub/' /usr/local/bin/nvgpuswitch.py
This issue started occurring on April 27 when the CUDA repository signing key was updated (the nvgpuswitch.py script from Holoscan SDK 0.1 installs the old signing key). Enabling dGPU prior to April 27 worked correctly, and would have installed the older signed package files in the local cache on the DevKit, so switching in and out of dGPU mode may continue to work on those systems. For those systems on which dGPU had been enabled prior to April 27, the nvgpuswitch.py script should not be patched unless enabling dGPU fails.