Regarding the issue of CUDA Fabric Memory

[EmmaLee-c]

Hello, I have a question. For example, in a supernode environment, I execute osu_bw -d cuda D D for inter-node communication. However, the data buffer in osu_bw is allocated using cudaMalloc instead of using the cuMem* related interfaces. In this case, can UCX still use NVLink for communication instead of using RC for inter-node communication?
Thank you.