Tools
llama.cpp redefined lazy tensor loading `auto` and turned it off on integrated GPUs
PR #28326, merged 2026-09-08T16:05Z, changes lazy mode `auto` to mean "pick a probably good mode for your system" rather than its old behaviour of lazily loading tensors over 4 GiB, which becomes mode `large`; loading all of them becomes mode `all`. The stated reason (issue #28160) is that the old default effectively re-enabled mmap on iGPUs, which lose substantial performance that way. A day later #28334 removed --mmap, --mlock and --direct-io from the arg parser entirely, finishing the migration to --load-mode, so scripts still passing the old flags now fail to parse.
Source
↳ Follow the thread