Well it works for me on my local linux server with RTX3090 and 4090 using nVidia driver 575.51.03, and it also works on runpod as well using their 3090/4090 machines.
What’s lacking overall for linux is a common “baseline” of drivers, libraries and the TVAI stack to make this work for everyone. This is something TL could offer, by just posting a single message on this forum (or on the support page/online docs).
It would at least serve as reference of “what works” so beta-users can bisect what exactly differs on the specific machine they’re using where it doesn’t work, and once that’s understood, they’d be able to fix the problem there.
To that end, I created a Docker repo a while ago when I had this exact reproduction issue and the TL people only answered “works on my machine” / “in our snowflake environment inhouse”. You can find it here.
That repo standardizes everything but the nvidia driver on your machine (and the HW of course), since it has all the CUDA related stuff in the container as well. See if that works.
If that’s not sufficient, ping back and I can publish the TVAI runpod template (“VM” image) I’m using, which also standardizes the hardware and nVidia driver, so you can investigate what differs vs. your local machine.
