I wonder if turning on GGML_CUDA_NCCL would help reduce delay between cards