ggml_repeat no cuda impl

[SpaceCowboy850]

I'm seeing a number of embedding models based off of

https://github.com/skeskinen/bert.cpp

Including this

https://github.com/xyzhang626/embeddings.cpp

Which implements BGE. But none of them are GPU accelerated. It seems that they all rely on ggml_repeat which doesn't have a cuda implementation, so we are stuck constantly going back and forth to the GPU if we were to put the other layers on the GPU. Do I have that right?

[FSSRepo]

It seems that those repositories have been without updates for quite some time and they use a rather old version of ggml. Many operations now have GPU acceleration implemented, but it requires those projects to implement ggml-backend.

[SpaceCowboy850]

Yes, exactly! I'm trying to update it, but ggml_repeat still has no gpu backend. It looks like it is a slow way of doing broadcasting. I've got the starts of a "BERTCPP" equivalency, with updated GGML, but I keep running into roadblocks like this that leave me scratching my head as to how to proceed.

Edit: Nevermind - I see a cuda repeat option has been added to latest, I need to update my LlamaCPP...