014dca49d6
* CUDA: manage NCCL communicators in context * add check that all backends are CUDA * remove unused vector, limit init to > 1 GPUs * fix warnings * fix cuda device, cache allreduce