llama.cpp

CUDA: add dynamic shared mem to softmax, refactor general usage

#14497

Merged