DeaMoE is proposed, a decoding-efficient MoE architecture, in which the experts are grouped into several departments, and the experts belonging to the same department share most parameters since they come from the same professional field, and additionally each expert contains a few private parameters to reflect its uniqueness.
Abstract
Mixture-of-Experts (MoE) models have been widely adopted in real-time interactive applications such as coding assistants, real-time audio-video interaction systems. To meet the extremely low response latency requirements of these scenarios, practitioners commonly employ small-batch decoding, under which MoE inference becomes memory-bound and is severely bottlenecked by expert weight loading. However, this bottleneck has received limited attention, and existing solutions such as post-training weight compression or fine-grained expert design during pre-training either degrade model accuracy or introduce additional computation and communication overhead. To tackle this issue, we propose DeaMoE, a decoding-efficient MoE architecture, in which the experts are grouped into several departments, and the experts belonging to the same department share most parameters since they come from the same professional field, and additionally each expert contains a few private parameters to reflect its uniqueness. Moreover, we design customized two-stage routing strategy for DeaMoE to avoid redundant loading, under which DeaMoE greatly improves the efficiency during LLM decoding. Compared with vanilla MoE, DeaMoE reduces per-step loaded weights by up to 50.9% and achieves up to 1.33 end-to-end TPOT speedup for the pre-trained 7B model on A40, and up to 2.00x and 1.97x peak speedup for DeepSeek-V3 on A40 and H100 in microbenchmarks.
AE-PSL is proposed, a communication-efficient PSL framework that compresses intermediate activations and gradients using a lightweight AutoEncoder placed at the split layer and introduces a novel two-stage alignment mechanism, which adapts the AE to the pre-trained model's feature manifold and client-specific feature distributions before DFT.
Bas Meuwissen, Vasileios Tsouvalas, N. Meratnia· 1 citation
ExactMoE, an inference design that applies symmetric group-128 four-bit weight quantization only to routed experts, stores those experts in kernel-native MARLIN form in pinned host memory, and executes all selected experts through a configurable GPU-resident slot cache and fused grouped MoE kernels, identifies a practical memory-transfer-throughput frontier for complete-expert MoE inference.
EdgeXpert is proposed, a software-hardware co-designed LLM accelerator that resolves this incompatibility and achieves up to 56.3% latency reduction and 44.1% energy reduction compared to prior works, while maintaining near-baseline accuracy.
Sangwoo Ha, Hyunwoo Seo, Yurim Jo et al.· 0 citations
DraftExpert is proposed, an expansion-aware self-speculative decoding framework for expert-offloaded MoE inference that improves decode throughput by 1.45x on average, raises draft acceptance to 84~87%, and achieves 86~88% prefetch hit rates.
Diffusion transformers are essential for high-fidelity video generation, but long token sequences make attention a dominant inference bottleneck. Training-free dynamic sparse attention alleviates this bottleneck by computing only selected key-value blocks, yet existing methods struggle to sparsify attention both efficiently and accurately for two reasons: (1) Rigid, unpredictable, and costly routing: selecting a fixed fraction of top-ranked blocks by proxy score imposes fixed budgets, whereas retaining blocks to reach a target cumulative proxy probability mass yields dynamic but potentially imbalanced budgets; both incur non-negligible overhead from computing and materializing proxy scores. (2) Lossy keep-or-drop sparsification: unselected blocks are discarded entirely, degrading accuracy under aggressive sparsity. These limitations motivate cheaper dynamic-budget routing while limiting accuracy degradation. In this paper, we introduce training-free Sol-Attn (Sparsifying online attention), which unifies dynamic routing, sparse computation, and approximation correction in a single online-softmax pass, achieving a better accuracy-efficiency trade-off in sparse attention. The core of Sol-Attn is on-the-fly block thresholding with proxy-score reuse, which selects critical blocks by comparing block proxy scores against a threshold during online softmax. This design enables dynamic yet controllable block budgets without materializing the proxy map, while directly reusing the proxy scores of unselected blocks to approximate their contribution. Experiments across image and video generation tasks show that Sol-Attn advances the quality-efficiency frontier of training-free sparse attention, delivering 2.1 times and 2.3 times end-to-end speedups for video generation and editing, respectively, while preserving visual quality.
Haopeng Li, Yitong Li, Junsong Chen et al.· 2 citations
Ultra-lightweight models are essential for the deployment of deep learning-based speech enhancement algorithms on edge devices. Although recent approaches have achieved a certain balance between computational complexity and performance, pushing the complexity limits further demands more sophisticated designs. In this letter, we propose CoFi-Lite, a highly efficient model that decouples spectral modeling into coarse- and fine-grained streams. By leveraging two parallel and symmetric encoder-decoder paths, it simultaneously extracts full-band envelopes and low-frequency details for complementary enhancement. In addition, a novel Cross-Path Fusion (CPF) module is introduced to bridge the distinct paths, facilitating efficient feature interaction. Remarkably, CoFi-Lite requires extremely low computational resources, featuring only 12.87 M MACs/s and 83.12 k parameters. Experimental results demonstrate that our proposed model outperforms the ultra-lightweight baseline GTCRN while requiring only 40.26% of its computational complexity. Its scaled-up variant also delivers performance on par with that of the SOTA ultra-lightweight model AdaptCRN alongside a 19.34% reduction in computational cost.
Leyan Yang, Dahan Wang, Xiaobin Rong et al.· IEEE Signal Processing Lette...· 0 citations