Yingxin Li

Zihui Zhao, Yingxin Li, Yang Li

Multimodal Large Language Models (MLLMs) have demonstrated exceptional success in various multimodal tasks, yet their deployment is frequently limited by substantial computational demands and prolonged inference times. Given that the vision modality typically contains more comprehensive information than the text modality, resulting in encoded representations comprising an extensive number of tokens, leading to significant computational overhead due to the quadratic complexity of the attention mechanism. Current token reduction methods are typically restricted to specific model architectures and often necessitate extensive retraining or fine-tuning, restricting their applicability to many state-of-the-art models. In this paper, we introduce a learning-free token reduction (LFTR) method designed for MLLMs. LFTR can be seamlessly integrated into most open-source MLLM architectures without requiring additional fine-tuning. By capitalizing on the redundancy in visual representations, our approach effectively reduces tokens while preserving the general inference performance of MLLMs. We conduct experiments on multiple MLLM architectures (LLaVA, MiniGPT, QwenVL), and our results show that LFTR achieves up to a $16\times$ reduction of visual tokens while maintaining or even enhancing performance on mainstream vision question-answering benchmarks, all in a learning-free setting. Additionally, LFTR is complementary to other acceleration techniques, such as vision encoder compression and post-training quantization, further promoting the efficient deployment of MLLMs. Our project is available at https://anonymous.4open.science/r/LFTR-AAAI-0528.

Mingzi Wang, Yuan Meng, Chen Tang et al.

The co-design of neural network architectures, quantization precisions, and hardware accelerators offers a promising approach to achieving an optimal balance between performance and efficiency, particularly for model deployment on resource-constrained edge devices. In this work, we propose the JAQ Framework, which jointly optimizes the three critical dimensions. However, effectively automating the design process across the vast search space of those three dimensions poses significant challenges, especially when pursuing extremely low-bit quantization. Specifical, the primary challenges include: (1) Memory overhead in software-side: Low-precision quantization-aware training can lead to significant memory usage due to storing large intermediate features and latent weights for back-propagation, potentially causing memory exhaustion. (2) Search time-consuming in hardware-side: The discrete nature of hardware parameters and the complex interplay between compiler optimizations and individual operators make the accelerator search time-consuming. To address these issues, JAQ mitigates the memory overhead through a channel-wise sparse quantization (CSQ) scheme, selectively applying quantization to the most sensitive components of the model during optimization. Additionally, JAQ designs BatchTile, which employs a hardware generation network to encode all possible tiling modes, thereby speeding up the search for the optimal compiler mapping strategy. Extensive experiments demonstrate the effectiveness of JAQ, achieving approximately 7% higher Top-1 accuracy on ImageNet compared to previous methods and reducing the hardware search time per iteration to 0.15 seconds.