Presentation
Late Breaking Results: Efficient MoE-ViT Gating on Edge Accelerator via Dynamic Expert Routing and Quantized Router
DescriptionWe propose a dynamic routing strategy that adapts expert usage to input difficulty while reducing overall computation. We also quantize the gating weights using a multiplier-free scheme and apply
post-training calibration to preserve routing decisions. Compared with Top-K routing, dynamic routing improves multi-task performance by 5.07%. During inference on the ZCU102, it reduces MoE
expert FFN computation by 24.71% and lowers end-to-end latency by 4.3%, while gating quantization reduces DSP usage by 85%.
post-training calibration to preserve routing decisions. Compared with Top-K routing, dynamic routing improves multi-task performance by 5.07%. During inference on the ZCU102, it reduces MoE
expert FFN computation by 24.71% and lowers end-to-end latency by 4.3%, while gating quantization reduces DSP usage by 85%.
Event Type
Late Breaking Results
TimeMonday, July 275:45pm - 5:48pm PDT
LocationExhibit Hall
Similar Presentations