Then I went after long context, because my agent runs go past 50K tokens and it got slow there. With q8 or q4 KV on Ampere, llama.cpp runs decode attention in a vector kernel that reads every KV head…