DraftExpert: Expansion-Aware Self-Speculative Decoding for End-Device MoE Inference

Large Mixture-of-Experts (MoE) language models are attractive for end-device deployment because only a small subset of experts is active per token, but their routed expert weights often exceed accelerator memory. We target latency-critical single-user settings where routed experts are staged on demand from CPU memory to a GPU or from Flash to a mobile NPU. In this setting, self-speculative decoding faces a new bottleneck: increasing the draft expert set improves accuracy but triggers extra expert loading, while cheap small-footprint drafts have low acceptance; moreover, verifying a multi-token block activates the union of target experts and is no longer close to one target step. We propose DraftExpert, an expansion-aware self-speculative decoding framework for expert-offloaded MoE inference. DraftExpert trains one lightweight accelerator-resident draft expert per layer by self-distilling residual, logit/token, and router-agreement signals from the frozen target MoE. At inference time, it uses a fixed-footprint shared+top-1+draft-expert drafter together with confidence--expansion truncation and target-expert prefetching, while final tokens are still exactly verified by the target model. On DeepSeek-V2-Lite and Moonlight-16B-A3B across CPU-GPU and Flash-NPU offload, DraftExpert improves decode throughput by 1.45x on average, raises draft acceptance to 84~87%, and achieves 86~88% prefetch hit rates.

SiHGNN: Leveraging Properties of Semantic Graphs for Efficient HGNN Acceleration

Heterogeneous Graph Neural Networks (HGNNs) have expanded graph representation learning to heterogeneous graph fields. Recent studies have demonstrated their superior performance across various applications, including medical analysis and recommendation systems, often surpassing existing methods. However, GPUs often experience inefficiencies when executing HGNNs due to their unique and complex execution patterns. Compared to traditional Graph Neural Networks, these patterns further exacerbate irregularities in memory access. To tackle these challenges, recent studies have focused on developing domain-specific accelerators for HGNNs. Nonetheless, most of these efforts have concentrated on optimizing the datapath or scheduling data accesses, while largely overlooking the potential benefits that could be gained from leveraging the inherent properties of the semantic graph, such as its topology, layout, and generation. In this work, we focus on leveraging the properties of semantic graphs to enhance HGNN performance. First, we analyze the Semantic Graph Build (SGB) stage and identify significant opportunities for data reuse during semantic graph generation. Next, we uncover the phenomenon of buffer thrashing during the Graph Feature Processing (GFP) stage, revealing potential optimization opportunities in semantic graph layout. Furthermore, we propose a lightweight hardware accelerator frontend for HGNNs, called SiHGNN. This accelerator frontend incorporates a tree-based Semantic Graph Builder for efficient semantic graph generation and features a novel Graph Restructurer for optimizing semantic graph layouts. Experimental results show that SiHGNN enables the state-of-the-art HGNN accelerator to achieve an average performance improvement of 2.95$\times$.

Characterizing and Understanding HGNN Training on GPUs

Owing to their remarkable representation capabilities for heterogeneous graph data, Heterogeneous Graph Neural Networks (HGNNs) have been widely adopted in many critical real-world domains such as recommendation systems and medical analysis. Prior to their practical application, identifying the optimal HGNN model parameters tailored to specific tasks through extensive training is a time-consuming and costly process. To enhance the efficiency of HGNN training, it is essential to characterize and analyze the execution semantics and patterns within the training process to identify performance bottlenecks. In this study, we conduct an in-depth quantification and analysis of two mainstream HGNN training scenarios, including single-GPU and multi-GPU distributed training. Based on the characterization results, we disclose the performance bottlenecks and their underlying causes in different HGNN training scenarios and provide optimization guidelines from both software and hardware perspectives.