>>106337644
They aren't that competent
Otherwise we would have already had accelerated transformer specific pipelines on Nvidia GPUs. Instead they push this FP16->FP8->FP4 "gains"