Questions regarding token calculation and FLOPs measurement

[zychen8]

Thank you for your excellent work on VisionZip. I have two questions:

1. FastV prunes R% of tokens from layer K onwards. Does the table below use the average number of tokens across all decoder layers, or is it calculated differently?

2. In addition to latency improvements, did you measure the theoretical FLOPs reduction for the different pruning approaches? The calculation process would be greatly appreciated.

Thank you for your time and for considering these questions.

[Yangsenqiao]

Thanks for your interest in VisionZip!

1. The table reports the average number of tokens across all decoder layers. Some of the reported results follow the SparseVLM.

2. Regarding FLOPs calculation, we provide a discussion in the Preliminary section. However, our main focus is on real running time, as we believe it better reflects actual efficiency in real-world scenarios compared to theoretical FLOPs.