App compiled with cubin.h slow start

Hi,

I have an app that’s compiled using cubin.h, installed in a docker container.
A second app uses ptx files - also in a docker continer.

When the first app (cubin based) runs after the second (ptx app), it takes a long time to start.
If we run the cubun based app right after the cubin based app (twice in a row), it starts ok.

Ubuntu 20.04
docker containers
nvidia driver version 560.35.03
NVIDIA GeForce RTX 3050 Ti Laptop GPU
CUDA Version: 12.6
Server Vendor Version: 1.20.13 (Can this be relevant?)

Does not reproduce on pc with driver models GTX 1650

Any idea?

Thanks