Mixture-of-Experts (MoE) architectures have recently gained increasing popularity within the domain of large language models (LLMs) due to their ability to significantly reduce training and inference overhead. However, MoE architectures face challenges, such as significant disparities in the number of tokens assigned to each expert and a tendency toward homogenization among experts, which adversely affects the model's semantic generation capabilities. In this paper, we introduce LocMoE+, a refined version of the low-overhead LocMoE, incorporating the following enhancements: (1) Quantification and definition of the affinity between experts and tokens. (2) Implementation of a global-level adaptive routing strategy to rearrange tokens based on their affinity scores. (3) Reestimation of the lower bound for expert capacity, which has been shown to progressively decrease as the token feature distribution evolves. Experimental results demonstrate that, without compromising model convergence or efficacy, the number of tokens each expert processes can be reduced by over 60%. Combined with communication optimizations, this leads to an average improvement in training efficiency ranging from 5.4% to 46.6%. After fine-tuning, LocMoE+ exhibits a performance improvement of 9.7% to 14.1% across the GDAD, C-Eval, and TeleQnA datasets.

The Mixtures-of-Experts (MoE) model is a widespread distributed and integrated learning method for large language models (LLM), which is favored due to its ability to sparsify and expand models efficiently. However, the performance of MoE is limited by load imbalance and high latency of All-To-All communication, along with relatively redundant computation owing to large expert capacity. Load imbalance may result from existing routing policies that consistently tend to select certain experts. The frequent inter-node communication in the All-To-All procedure also significantly prolongs the training time. To alleviate the above performance problems, we propose a novel routing strategy that combines load balance and locality by converting partial inter-node communication to that of intra-node. Notably, we elucidate that there is a minimum threshold for expert capacity, calculated through the maximal angular deviation between the gating weights of the experts and the assigned tokens. We port these modifications on the PanGu-Sigma model based on the MindSpore framework with multi-level routing and conduct experiments on Ascend clusters. The experiment results demonstrate that the proposed LocMoE reduces training time per epoch by 12.68% to 22.24% compared to classical routers, such as hash router and switch router, without impacting the model accuracy.

Training large-scale mixture of experts models efficiently on modern hardware requires assigning datapoints in a batch to different experts, each with a limited capacity. Recently proposed assignment procedures lack a probabilistic interpretation and use biased estimators for training. As an alternative, we propose two unbiased estimators based on principled stochastic assignment procedures: one that skips datapoints which exceed expert capacity, and one that samples perfectly balanced assignments using an extension of the Gumbel-Matching distribution [29]. Both estimators are unbiased, as they correct for the used sampling procedure. On a toy experiment, we find the `skip'-estimator is more effective than the balanced sampling one, and both are more robust in solving the task than biased alternatives.

In deep learning, models typically reuse the same parameters for all inputs. Mixture of Experts (MoE) defies this and instead selects different parameters for each incoming example. The result is a sparsely-activated model -- with outrageous numbers of parameters -- but a constant computational cost. However, despite several notable successes of MoE, widespread adoption has been hindered by complexity, communication costs and training instability -- we address these with the Switch Transformer. We simplify the MoE routing algorithm and design intuitive improved models with reduced communication and computational costs. Our proposed training techniques help wrangle the instabilities and we show large sparse models may be trained, for the first time, with lower precision (bfloat16) formats. We design models based off T5-Base and T5-Large to obtain up to 7x increases in pre-training speed with the same computational resources. These improvements extend into multilingual settings where we measure gains over the mT5-Base version across all 101 languages. Finally, we advance the current scale of language models by pre-training up to trillion parameter models on the "Colossal Clean Crawled Corpus" and achieve a 4x speedup over the T5-XXL model.