Does DeepEP use SMs for its normal and low-latency kernels?

I saw in a DeepSeek-V3 paper (https://arxiv.org/pdf/2505.09343) that "during training, up to 20 of the SMs on the H800 GPU are allocated for communication-related operations, leaving fewer resources available for actual computation." and I also saw that in the issue (https://github.com/deepseek-ai/DeepEP/issues/61#issuecomment-2709205782) that with `return_recv_hook`, communication kernels does not use any SM, it just issue RDMA requests and just return.
Does that mean DeepSeek training uses normal kernel and the 20 SMs are used because of NVLink forwarding?
Any help is much appreciated. Thank you!

Start with the `return_recv_hook` behavior described in issue #61's linked comment, then inspect the normal and low-latency kernel entry points mentioned in this issue. Compare those findings with the cited DeepSeek-V3 paper. Done means documenting whether the kernels use SMs and how that relates to the reported 20 communication-reserved SMs and NVLink forwarding.

Written by the indexing model from the issue text.