Again, you're using some 3rd party quantisations, not the weights supplied by Nvidia (which don't fit in 24GB).