The repack cache is 1.9 per cent larger than the model file where it was ten
per cent, so an 8B holds 4781 MiB of weights against 5151 and a 4090 holds
5540 MiB of them at a context of 2048 against 5910. That is 1.066 times what
llama.cpp holds on the card, and it costs 2.5 per cent of a decode.

There is no longer a lossy step anywhere in the repack. All eight quantized
types round trip bit for bit and the logit corpus agrees with llama.cpp on
all thirteen device cases on both backends.