RPC with tensor override

[yggdrasil75]

I have a system with 2 3090s and a system with a 395+. the 395+ doesnt have pcie slots but runs decent vulkan speeds.
I want to run a larger moe split across the 2 systems. glm air I am getting 191 t/s pp and 17 t/s tg. these numbers are quite good for a 106b at q6k.
but I figured I can do better by using the 3090s as cuda devices and ffn_up on the 395+. unfortunately that fails and I get 3 t/s pp and 4 t/s tg.

is it possible to offload the tensors by layer per device? ie: layers 1-14 on rpc0 with ffn_up from those layers on rpc0 system ram, layers 15-28 on rpc1 with those layers on rpc1 system ram, and the remaining layers on vulkan0 completely?