llama.cpp

8ad92dc1 - ggml : switch to padded F16 mask for ggml_soft_max, ggml_flash_attn_ext