Global-batch load balance almost free lunch to improve your MoE LLM training

GITHUB HUGGING FACE MODELSCOPE DISCORD Background The Mixture-of-Experts (MoEs) architecture has become a popular model-parameter-scale-up technique. Typically, one MoE layer consists of a router (often parameterized as one single Linear layer) and a group of experts (for transformer-based models, each expert is one feedforward layer). Given an input, only a subset of experts will be activated,…