CYBER WATCH / EQUINOXE INTELLIGENCE

← All articles

b12245: llama – apply the iGPU lazy-mode fallback only when mmap is off (#31957)

llama.cpp Releases · 2026-11-10 12:01:35

On a device lacking mmap support, --lazy-mode automatically resolved to off even when --load-mode mmap was used, preventing llama_prefetch_rows from running and causing large tensor rows to fault one page at a time. The fix enables the fallback only when mmap is off, leaving the default iGPU load path unchanged. Addresses issue #31943.

Read original source ↗