llama.cpp

CUDA: better error for FA kernel with 0 occupancy

#16643

Merged