llama.cpp

df1d0641 - reuse ggml_backend_cuda_comm_allreduce_nccl for per-rank communicators. Remove the helper function backend_comm_proc_address