I moved the topic to a separate thread.
GPU pass-through into docker is a complicated process.
I suggest you research the configuration for your explicit use case.
With certain specialised projects even small changes in applications or drivers may cause the project to malfunction.
I am guessing you are running an AI project in docker which depends on Nvidia GPU - perhaps the upstream project know which knobs to adjust.
An incident like you describe may be why certain projects tend to support Red Hat and Ubuntu - but not Arch Linux and derivatives like Manjaro - we are simply too unstable - even when using stable branch - in context of the project.