Small thing that I have found missing in the current discussion around the xz debacle:
If a project that you have an interest in has already a problem with maintainer capacity, probably the worst solution is to just add a maintainer (and even worse doing it in a rush).
You don't need to arm-twist the overworked maintainer in sharing power, you can just start doing non-binding reviews, answering questions and offloading other work from the maintainer(s).
@mntmn I think it should work well, but I have looked at the model and it uses lots of operations that aren't supported at the moment.
I would need to take a closer look, but because there are so many operations that aren't convolutions, I think it is possible that this model would be more efficiently run on a GPU.
@mntmn Actually, I just ran the model and not a single operation was run on the systolic arrays. Only a few were run in the (underpowered) programmable core and the rest were rejected by Vivante's driver.
The @collabora guys are working on open-source TTS models, maybe they have something that maps better to this HW.