llama.cpp

#1869 Fix null reference errors when training from scratch with CUDA

#1907

Merged