Joint Architecture-Token-Bitwidth Multi-Axis Optimization of Vision Transformers for Semiconductor IC Packaging
The study developed a multi-axis framework to optimize Vision Transformers for industrial tasks by jointly adjusting architecture, token strategy, and bit-width. By using Neural Architecture Search (AutoFormer) for compact backbones and Token Merging (ToMe) for sequence compression, the method achieved over 10x improvements in throughput while maintaining accuracy on a specific semiconductor defect classification task.