CYBER WATCH / EQUINOXE INTELLIGENCE

← All articles

b12331: Vulkan – Intel FA kernel for prefill (#29357)

llama.cpp Releases · 2026-11-13 02:02:43

The update introduces an Intel FA kernel optimized for prefill, featuring a custom shader, bounds checks (max_bias/logit_softcap), SIMD32 and M16N16K16 support, and fixes for alignment and precision. It limits SLM usage to 48 KiB, improves handling of large K dimensions, unaligned strides, and multi‑head masks, and strengthens host‑side robustness by removing unnecessary syncs and adding SLM size checks.

Read original source ↗