CYBER WATCH / EQUINOXE INTELLIGENCE

← All articles

b12144: CUDA – Linearize MoE data layout for MMQ/MMF (#31047)

llama.cpp Releases · 2026-11-06 04:01:10

This update linearizes the MoE data layout for MMQ/MMF, applies MMQ tile size optimization more broadly, and fixes an F16 performance regression.

Read original source ↗