TL;DR: switching KV cache to f16 may give a boost in speed if using MTP and ngrams. I have a self-built "AI mega-cluster" with 2x P40s on a cheap Chinese motherboard and a Xeon CPU (around $1,100 to build, including water cooling for the GPUs). I was normally getting up to 15 tk/s with Qwen 3.8 27B
