Reproduce “Communicate the Change: Efficient Expert Parallelism via Activation Compression” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Fengyuan Liu, Jay Gala, Siva Reddy, Dzmitry Bahdanau
EchoMoE reduces communication in mixture-of-experts serving by caching a token’s earlier hidden state and sending compressed changes for dispatch, while also compressing expert outputs for combine. On Qwen3-30B-A3B, one adapted configuration cuts estimated BF16 inter-node RDMA bytes by 58% and raises throughput by 22.6% in a specified serving setup, while the 27-pair mean score falls 3.8 points, from 66.9% to 63.1%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.