System update Problem

It’s a really small and foolish problem, but my friends and I have conflict.
When I use tensorflow with GPU, an error occurs, telling me that CUDA driver version is insufficient for CUDA runtime version.
So I asked my friends, who tell me that I cannot run the command ‘sudo apt-get update/upgrade’ after installing CUDA with JetPack3.2

Is it true that I cannot update/upgrade any more after JetPack? Personally, I don’t think so,still need double check.

Thanks.
Sincerely

Do you see these problems on Jetson or host ?

It is possible that some problems happen with apt after installing JetPack on host because it may have installed arm64 arch.

Some problems have also been seen with apt on Jetson after installing R28.2.

For these apt problems, see https://devtalk.nvidia.com/default/topic/1002140/jetson-tx2/apt-get-update-errors/post/5247634/#5247634.

The Problem is on my Jetson TX2.

Yesterday I flash the TX2 with Jetpack3.2 again. With no other operation, I install tensorflow1.7 by using whl file. And try the GPU at once, it still told me the same problem ‘CUDA driver version is insufficient for CUDA runtime version’. So apparently it’s not about update operation, it’s the JetPack I think.

I don’t know about tensorflow issues, but JetPack can act as a front end to both flashing and extra package installs. If you’ve flashed with JetPack3.2, then you should have L4T R28.2 (currently the most recent release). You can uncheck flash and manage packages at any time…this is a full Ubuntu 16.04 installation when you install L4T R28.2.

You should keep the Jetson updated. There can be an issue with the libglx.so library being overwritten by one of the standard packages, but this is not often and is easily fixed (especially if you fix before rebooting after a package update).

To test if packages are ok and current (not overwritten) you can run this command:

sha1sum -c /etc/nv_tegra_release

The only file I know of which might get overwritten (and would be listed as OK or not in the above command) is:

/usr/lib/xorg/modules/extensions/libglx.so

There is a copy of this file which would never be harmed by regular package updates (and the one which is is rare). This file can be used to restore the other:

/usr/lib/aarch64-linux-gnu/tegra/libglx.so

If you forget which, then just read “/etc/nv_tegra_release”, this is where the checksums are listed.

Example to restore:

sudo cp /usr/lib/aarch64-linux-gnu/tegra/libglx.so /usr/lib/xorg/modules/extensions/libglx.so

…use this to overwrite the bad copy with the good copy.

If the Jetson is rebooted with the wrong file, then the GUI will fail. Consoles and ssh will still work. The scp command (a subset of ssh) can still copy from one location to another remotely if you don’t have a console.

Someone else will have to comment on tensorflow, but keep your Jetson operating system updated with the usual package tools.

Also replied at https://devtalk.nvidia.com/default/topic/1031300/jetson-tx2/tensorflow-1-7-wheel-with-jetpack-3-2-/post/5256379/#5256379, but right thread may be https://devtalk.nvidia.com/default/topic/1033084/internal-cudagetdevice-failed-status-cuda-driver-version-is-insufficient-for-cuda-runtime-version/#5256252.

Hi,

Please check topic_1033084 for detail:
https://devtalk.nvidia.com/default/topic/1033084

Insufficient driver occurs if you try to install a wheel of CUDA8.0 on JetPack 3.2, which uses CUDA9.0.
Please make sure the wheel file you installed is built on the same environment.

More, apt-get works correctly on JetPack3.2.
Thanks.