First of all, just wonder how many memory does your system have? 32GB?

Not sure https://github.com/ROCm/ROCm/issues/1732#issuecomment-1911876812 could help you. Or you can try to run "sudo dmesg | grep -E 'amdgpu.*VRAM'" to see if the setting work.

If the kernel message told you that you that your VRAM is 4GB then pytorch would not give you more than 4GB. Better check this first.

If you see 4GB then it might be wrong with the functionalities of the UniversalAMDFormBrowser.

amdgpu could not build the latest dkms on the Zabbly kernel but it would use the dkms created on the ubuntu factory kernel. You can run 'dkms status' to check. But I guess if torch.cuda already give you 4GB ram then you should be fine.