GM-Skip: Metric-Guided Transformer Block Skipping for Efficient Vision-Language Models

TL;DR

GM-Skip introduces a metric-guided Transformer block skipping method that significantly accelerates vision-language model inference while maintaining high task performance, suitable for latency-sensitive applications.

Contribution

It proposes a novel, metric-adaptive framework for selectively skipping Transformer blocks in VLMs, balancing speed and accuracy with a greedy, feedback-driven approach.

Findings

Speeds up inference by over 40% on COCO tasks

Maintains high accuracy, e.g., 87.3% on Person classification

Reduces latency by up to 45.4% in autonomous vehicle deployment

Abstract

Transformer-based Vision-Language Models (VLMs) have achieved impressive performance on tasks such as image captioning, object recognition, and visual reasoning, but their high computational cost hinders deployment in latency-sensitive applications like autonomous driving. We introduce GM-Skip, a flexible and metric-adaptive framework for Transformer block skipping that accelerates VLM inference while preserving output quality. GM-Skip features a greedy, metric-guided block selection strategy that uses metric feedback (e.g., accuracy, CIDEr) to identify redundant layers, along with a reverse-order deletion mechanism that preserves early foundational blocks to avoid performance collapse. To support diverse deployment needs, it incorporates a tunable trade-off between sparsity and performance via a score-sparsity balance objective. Experiments across multiple tasks and datasets, including COCO and CODA, show that GM-Skip consistently improves inference speed while maintaining task performance. On the COCO dataset, GM-Skip improves single-object classification accuracy on the Person category from 19.1 percent to 87.3 percent while skipping more than 40 percent of Transformer blocks. In real-world deployment, it achieves up to 45.4 percent latency reduction on single-object detection when integrated into an autonomous vehicle running Autoware.Universe, validating the effectiveness of its skip configurations and confirming its practical value in accelerating real-world inference.