Can llama.cpp support run different attentions layers on different devices?

[chosen-ox]

For example, llama3.2 1b have 16 layers. Can I run the first 8 layers on one machine and then sent kv cache to another machine to run another 8 layers. This is simliar to idea of pipeline parallelism. If not supported now, I wonder can llama.cpp run only partial layers of model or not? If so, I can support pipeline parallelism manually. If llama.cpp can only run the inference as a whole, then I cannot do that.

[egebeysel]

I was also looking if llama.cpp, I think this #6017 added pipeline parallelism support.

[chosen-ox]

Thank you for your reply. I appreciate it.