rocm is trash to apu. use vulkan or opengl (alternative backends)
find latest build in pytorch nightly repo
use sudo before invoking amdgpu related commands, otherwise unavailable
use directml (pytorch, tensorflow) instead. need windows.
you need to set the shared graphical memory to a larger value.
rocm does not work for integrated graphic cards (even if you build it and configured the card name), and supported amd cards are limited.