llama.cpp redefined lazy tensor loading `auto` and turned it off on integrated GPUs.

auto now means "pick a probably good mode for your system"; its old behavior of lazily loading tensors over 4 GiB becomes mode large, and loading all becomes all. The old default effectively re-enabled mmap on iGPUs, which lose substantial performance that way. A day later, --mmap, --mlock and --direct-io were removed from the arg parser entirely, so scripts passing the old flags now fail to parse. GitHub