llama.cpp

vulkan: Use unclamped loads for flash attention mask

#12720

Merged