llama.cpp

d52056be - reduce occupancy on RDNA4 for large attention head sizes