llama: add default load-mode auto, which avoids mmap on iGPUs - #26081

Merged

ggerganov merged 9 commits into

0cc4m/mmap-auto

Aug 11, 2026

Commits on Aug 5, 2026

llama: add new default load-mode auto which picks mmap unless a non-Metal iGPU is used
0cc4m
committed
01649e6
Update ggml/src/ggml-hexagon/ggml-hexagon.cpp

authored and
0cc4m
committed
9d8bd14
set mmap_support to false on OpenCL backend
0cc4m
committed
4b4fada
fix order of load modes
0cc4m
committed
2f2d246
use -1 for auto
0cc4m
committed
9391a09
resolve load mode auto earlier to correctly pick gpu host or cpu memory
0cc4m
committed
18c2399
add load mode auto to llama-bench
0cc4m
committed
f00aaaa
bump virtgpu api version, regenerate docs
0cc4m
committed
125b6c6

Commits on Aug 11, 2026

Merge branch 'master' into 0cc4m/mmap-auto
ggerganov
committed
233ec61