My friends keep asking for a workflow that just works on their normal GPUs and after revisiting F1.S and then then AIO models i get to this

this is the launch i suggest
--cuda-malloc --force-channels-last --dont-upcast-attention
with
--normalvram
or
--lowvram