disaggregated inference thoughts: what are you seeing?

[vramport]

I'd approach it in a boring order. Before rebuilding anything, I'd test one clean baseline with Llama and compare from there. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.