Beyond Token Pruning: Operation Pruning in Vision-Language Models

Abstract: Prior Vision LLM (VLM) token pruning reduces computation by eliminating attention and feed-forward operations for pruned tokens while maintaining all operations for critical tokens. However, this binary approach conflates token/operation redundancy - critical operations may be removed along with discarded tokens, while preserved tokens retain all potentially redundant operations. To surgically eliminate redundant operations while preserving critical ones, we propose Greedily Sorted Operation Pruning (GSOP), a data-driven method that directly prunes operations rather than tokens. GSOP first decomposes a VLM decoder's computations into atomic operations along three dimensions: token groups, layer positions, and computation modules. GSOP determines the pruning order of operations through greedy sorting: GSOP iteratively selects the redundant operation that incurs minimal performance drop considering previously pruned operations. Different computational budgets can be accommodated without re-searching by simply pruning operations according to this order until the desired budget is met. GSOP enhances sorting efficiency through: a) leveraging historical operation rankings to avoid redundant evaluations; b) excluding the free-to-prune" anddanger-to-prune" operations from sorting. GSOP achieves compelling efficiency-performance tradeoffs, reducing computation by 70% with only 4% performance loss while maintaining up to 18% higher performance than state-of-the-art methods when transferred across diverse VLMs and tasks. Real GPU efficiency evaluations confirm its practical value. The code is in https://github.com/zxcvfd13502/GSOP.