LLC miss

[KKwanhee]

I found some model data which is 'not aligned' (maybe not aligned to ggml tensor alignment) and it can't use mmap so loads every model data on main memory.

My question is, compared to 'aligned' model data which use mmap, runing with 'not aligned' model data occurs around 3x LLC miss and why is that happening? I checked the LLC miss with Intel VTune and most LLC miss occured at 'ggml_vec_dit_f16' from 'ggml_compute_forward_mul_mat'.