Hi, I've been trying out nccl tests on g6e nodes and it maxes out at 3GBps. I see that gpudirect rdma is not enabled on g6e nodes when trying to use aws nccl plugin, with efa enabled on 4 network interfaces. Nvidia states that those gpus do support gpudirect rdma.

Is it a matter of having a topology file or something more needed?

Would love to contribute support if given some pointers.