disaggregated inference thoughts with llama.cpp

[kernelrunner]

I hit almost the same issue last night.

What exact setup are you running? What model, what runtime, and roughly what context size?

If you post your exact settings, people here can probably narrow it down pretty fast.