The "--accelerator-trace" argument on nsys command line doesn't seem to work

I am profiling an application that utilizes DLA on my Jetson AGX Xavier (nsys version 2020.5.3.17-0256620).
I want to collect the traces of my DLA to see what it’s doing, using the command line interface of nsys.

The only command line argument that seem to do this is the --accelerator-trace option, which can only take none or nvmedia. I chose the latter since it’s the only one that looks right to me, show below:

However, once I finished profiling and opened the report, the report’s analysis option table shows that I have not enabled the Collect other accelerators trace, as shown below:

And I there is no “Other Accelerator Trace” timeline in the timeline view either.

However, all of the above works perfectly fine if I used the nsys-ui GUI program and profiled remotely. As long as I checked the Collect Other Accelerators Trace checkbox, I can get the timeline without any problem.

I am trying to automate things, and doing so on a GUI is tough. I need that cli interface to work.

Is there any fix to this?

Hello, you absolutely should be able to do that from the CLI.

Your Nsys version is pretty old, and we made some changes to --accelerator-trace with the last release, would you be willing to try with a newer version?

What’s the correct way of updating it?
I checked the Jetson SDK Manager, and the newest nsys still seem to be 2020.5.
I also installed the Nsight systems manually from the Download Center which provides an 2021.3, and I get this:

WARNING: GPU context switch tracing is not available on this system.                                                                                                                                        
    * The NVIDIA kernel module version must be 435.17 or greater.                                                                                                                                           
      Use 'nvidia-smi' to detect the kernel module version.
    * CUDA compatibility must not be enabled.
      It means that the version of 'libcuda.so' must match the version of the kernel module.

What should I do?

Hello, sir, could I get some further assistance? (above)

My apologies, Jetson versions update on a different schedule than the workstation versions, which is my usual focus. Let me pull in someone more comfortable on that platform.

The new Jetson release should go out in a week or two.

Thank you.
I’m in a hurry so I decided to automate the GUI version anyways. Tedious thing to do, but it works.