aws/aws-ofi-nccl ? reverse-engineered prompt

Build me the AWS OFI NCCL plugin so I can use libfabric as the network layer for NCCL applications on EC2.

I want a setup that can compile cleanly on a supported Linux machine, install the plugin, and make it easy to run a basic test to confirm it works with GPU collective jobs. If there are any missing prerequisites like libfabric or hwloc, please point that out and handle the setup steps clearly. I’d also like it to work for both NVIDIA NCCL and AMD RCCL if possible, and to follow the normal build and install flow from the repo.

Please keep it practical, use the existing project files, and if you need current dependency or build details, check the latest docs online. After that, make sure the result looks like something a developer could build and use without a lot of manual cleanup.