Search papers, labs, and topics across Lattice.
This paper surveys optimization techniques for accelerating the inference of Large Vision Language Models (LVLMs), which are bottlenecked by the quadratic complexity of attention over visual tokens. It categorizes existing approaches into visual token compression, memory management and serving, efficient architectural design, and advanced decoding strategies. The survey highlights limitations of current methods and suggests future research directions for efficient multimodal systems.
LVLM inference is ripe for optimization, but current acceleration techniques only scratch the surface.
Although Large Vision Language Models (LVLMs) have demonstrated impressive multimodal reasoning capabilities, their scalability and deployment are constrained by massive computational requirements. In particular, the massive amount of visual tokens from high-resolution input data aggravates the situation due to the quadratic complexity of attention mechanisms. To address these issues, the research community has developed several optimization frameworks. This paper presents a comprehensive survey of the current state-of-the-art techniques for accelerating LVLM inference. We introduce a systematic taxonomy that categorizes existing optimization frameworks into four primary dimensions: visual token compression, memory management and serving, efficient architectural design, and advanced decoding strategies. Furthermore, we critically examine the limitations of these current methodologies and identify critical open problems to inspire future research directions in efficient multimodal systems.
