llama.cpp

fix embeddings when using CUDA

#3657

Merged