Segmentation fault (core dumped) on cuda 12.4 & cudnn 9.3.0

[mv2050arshad]

I build dlib (cuda-active) and i tried to build my application using dlib Face recognition.
Everything build successfully but i got run-time error :
Segmentation fault (core dumped)

It was detected faces and gpu , but i think the problem was when it try to load model in gpu for landmark extraction.

cout<< dlib::cuda::get_num_devices() ; // print 1
dlib::cuda::set_device(0); // without error run this also

Whats wrong? what is the last version of cuda/cudnn which supported by dlib?

I installed :
Cuda 12.4
Cudnn 9.3.0
Gpu 4070

[arrufat]

Could you try passing the following flag when building dlib? -DDLIB_USE_CUDA_COMPUTE_CAPABILITIES=89

[mv2050arshad]

I used qmake, is'nt work ?

[arrufat]

Then I'm not sure, CMake will set the flags for you, I don't know about qmake.

[mv2050arshad]

I build version 19.23 and it build with my application, but all the time got same error : Segmentation fault (core dumped).
i use ubuntu 22.04 and gcc 11.3, i changed CUDA_COMPUTE_CAPABILITIES to 89, 86 , 80, 50. No one work and same error.
i set_device(0) before make net, nothing change. Have you ever successful to use cuda on new versions of ubuntu or cuda or dlib?

[arrufat]

I always use dlib from master, mostly on Arch Linux. However, at work, I have one machine with Ubuntu and:

• nvidia drivers 560
• cuda 12.6
• cudnn 9.3

And everything works building with CMake.

[mv2050arshad]

Thank you so much for your answers, it means that there is a way and have to correct my bugs.