llama.cpp

ggml : fix unnecessary f32 -> f16 -> f32 casts (mmla)

#5951

Merged