llama.cpp

e6ab7c1a - hex-dma: introduce direct-mapped DMA cache that is better suited for HVX FA mask handling (#29282)