llama.cpp

0324696b - fit : count nextn (MTP) blocks in n_gpu_layers so front layers stay on GPU (#26177)