llama.cpp

CUDA: use registers instead of smem in topk-moe

#16647

Merged