Surviving Partial Rank Failures in Wide Expert-Parallel MoE Inference

Xun Sun, Shaoyuan Chen, Pingchuan Ma +18

Mixture-of-Experts (MoE) serving relies on wide expert parallelism (EP) to aggregate the memory capacity and bandwidth of many GPUs within one inference instance. This efficiency c…