ReMoE: Boosting Expert Reuse through Router Fine-Tuning in Memory-Constrained MoE LLM Inference

Fine-grained Mixture-of-Experts (MoE) models sparsely activate only a subset of experts per token, reducing activated computation while maintaining high model capacity. However, in memory-constrained inference scenarios, only a small set of experts can be cached. Experts not in the cache must be fetched from slow external storage (e.g., UFS), leading to frequent evictions and substantial I/O overhead. We propose ReMoE, a router fine-tuning framework designed to boost token-wise expert reuse. ReMoE biases the router toward recently selected experts, producing temporally stable routing that better matches cache locality constraints. By increasing short-horizon expert reuse, ReMoE reduces expert fetches from storage without adding inference-time computation. Experiments on DeepSeek and Qwen models show that ReMoE improves expert reuse by 26\% while maintaining downstream task performance. Real-system evaluations further confirm these benefits, improving output throughput by 8.4\% under vLLM GPU--CPU expert offloading and reducing TPOT by 43.6-49.8\% under llama.cpp on Jetson Orin NX, corresponding to a 1.77-1.99$\times$ decode speedup across diverse workloads. Checkpoints and usage instructions are available at https://github.com/BUAA-OSCAR/ReMoE.

Large AI models are starting to move from cloud servers to personal and edge devices, such as phones, laptops, and small computers. Some of these models save computation by splitting their knowledge into many specialized parts and using only a few parts for each word they generate. This design is efficient, but it creates a memory problem: when the model keeps switching between different parts, the device must repeatedly load data from slower storage, which makes generation slower. We propose ReMoE, a lightweight way to teach the model to reuse recently used parts when this is appropriate. Instead of rebuilding the model, changing the hardware, or adding extra work during use, ReMoE adjusts only the small decision-making component that chooses which parts of the model to activate. It also keeps this component close to its original behavior, so the model can preserve its language, reasoning, and coding abilities. Across several models and devices, ReMoE reduces unnecessary switching and makes generation faster. On an edge device, it reduces per-word generation delay by about 44% to 50%, while keeping standard task performance largely unchanged.