Mixture-of-Experts (MoE) is widely adopted to deploy Large Language Models (LLMs) on edge devices with limited memory budgets. Although MoE is, in theory, an inborn memory-friendly architecture requiring only a few activated experts to reside in the memory for inference, current MoE architectures cannot effectively fulfill this advantage and will yield intolerable inference latencies of LLMs on memory-constrained devices. Our investigation pinpoints the essential cause as the remarkable temporal inconsistencies of inter-token expert activations, which generate overly frequent expert swapping demands dominating the latencies. To this end, we propose a novel MoE architecture, Oracle-MoE, to fulfill the real on-device potential of MoE-based LLMs. Oracle-MoE route tokens in a highly compact space suggested by attention scores, termed the oracle space, to effectively maintain the semantic locality across consecutive tokens to reduce expert activation variations, eliminating massive swapping demands. Theoretical analysis proves that Oracle-MoE is bound to provide routing decisions with better semantic locality and, therefore, better expert activation consistencies. Experiments on the pretrained GPT-2 architectures of different sizes (200M, 350M, 790M, and 2B) and downstream tasks demonstrate that without compromising task performance, our Oracle-MoE has achieved state-of-the-art inference speeds across varying memory budgets, revealing its substantial potential for LLM deployments in industry.

Language models (LMs) only pretrained on a general and massive corpus usually cannot attain satisfying performance on domain-specific downstream tasks, and hence, applying domain-specific pretraining to LMs is a common and indispensable practice. However, domain-specific pretraining can be costly and time-consuming, hindering LMs' deployment in real-world applications. In this work, we consider the incapability to memorize domain-specific knowledge embedded in the general corpus with rare occurrences and long-tail distributions as the leading cause for pretrained LMs' inferior downstream performance. Analysis of Neural Tangent Kernels (NTKs) reveals that those long-tail data are commonly overlooked in the model's gradient updates and, consequently, are not effectively memorized, leading to poor domain-specific downstream performance. Based on the intuition that data with similar semantic meaning are closer in the embedding space, we devise a Cluster-guided Sparse Expert (CSE) layer to actively learn long-tail domain knowledge typically neglected in previous pretrained LMs. During pretraining, a CSE layer efficiently clusters domain knowledge together and assigns long-tail knowledge to designate extra experts. CSE is also a lightweight structure that only needs to be incorporated in several deep layers. With our training strategy, we found that during pretraining, data of long-tail knowledge gradually formulate isolated, outlier clusters in an LM's representation spaces, especially in deeper layers. Our experimental results show that only pretraining CSE-based LMs is enough to achieve superior performance than regularly pretrained-finetuned LMs on various downstream tasks, implying the prospects of domain-specific-pretraining-free language models.