Inquiry About RDMA Utilization for EP128 in Lower-End GPU Clusters

Hi DeepSeek Team,

I came across your strategy of using RDMA to facilitate EP128 in large-scale GPU clusters, and I found it both impressive and thought-provoking. The approach of leveraging RDMA for overlapping computation and communication during decoding, while freeing up GPU SMs, is a brilliant optimization.

However, I was wondering whether this strategy could be applicable or scalable to lower-end GPU clusters. Given the significant costs associated with RDMA-capable networking hardware and infrastructure, **I am curious whether you have explored or considered similar optimization techniques on clusters with more modest hardware configurations.**

**Would it be possible to implement comparable overlapping strategies without RDMA**, perhaps using traditional high-speed Ethernet or software-based communication methods? Or do you see RDMA as a fundamental requirement to achieve the performance gains observed in EP128?

Any insights into the feasibility of adapting this method to less expensive setups would be greatly appreciated.

No file, test, or entry point is identified. Begin by checking whether this repository contains RDMA or EP128 implementation or documentation; the issue is done only when the feasibility question receives a maintainer-backed answer.

Written by the indexing model from the issue text.