A Replicate-and-Quantize Strategy for Plug-and-Play Load Balancing of Sparse Mixture-of-Experts LLMs

Zijie Liu, Jie Peng, Jinhao Duan +7

The paper proposes a training‑free method that replicates heavily used experts and quantizes less important ones to rebalance workload in sparse mixture‑of‑experts large language m…