For others, Llama.cpp+Vulkan is not practically functional on this image.
Some research on this here:
TLDR:
Llama.cpp+Vulkan is causing the Panfrost/PanVK to consume massive amounts of memory internally leading to Memory Allocation failure. Unsure on exact root cause at this point, but suspect something to do with TLS (and, maybe specifically, lack of subgoup support in driver).
Note, I’ve had vulkan eat insane amounts of RAM on AMD as well. I don’t know if it’s vulkan or a recent llama.cpp change that causes this, but the totality of the model’s size is suddenly sucked from the CPU’s RAM at the moment the server starts the warmup phase (i.e. makes accesses). This makes no sense but that’s my experience with it.
Firefox in snap doesn’t 3D accelerate, but the deb package version does. To install it, you must add mozilla repository and prioritize it over the ubuntu one, which is a snap package in disguise.
Vesktop (Vencord in a electron package I suppose, which is a kind of fork of the discord client) isn’t able to get the microphone permissions from ubuntu in SNAP, but it is able to get them as a deb package.
All in all, it seems best to avoid snap on O6N for now.