Long-context Multimodal Large Language Models (MLLMs) that incorporate long text-image and text-video modalities, demand substantial computational resources as their multimodal Key-Value (KV) cache grows with increasing input lengths, challenging memory and time efficiency. For multimodal scenarios, the cross-modal interactions inevitablely increase complexity, and prior methods for KV cache compression, in both text-only and multimodal LLMs, have neglected attention density variations across layers, often adopting uniform or progressive reduction strategis for layer-wise cache allocation. This results in precision loss and suboptimal performance. We propose MEDA, a novel approach specifically designed for the complexities of multimodal settings, dynamically allocating KV cache sizes based on attention entropy to better adapt to multimodal interactions.Through a dynamic multimodal KV cache allocation strategy, MEDA compresses the KV cache, adaptively retains sufficient multimodal information at each layer. Meanwhile, to mitigate the degradation of contextual information due to cache compression, we also integrate KV pairs merging techniques to maintain coherence. MEDA achieves up to 72% KV cache memory reduction and 2.82 faster decoding speeds in some cases, while maintaining or enhancing performance on various multimodal tasks in a long context, including multi-image and long video scenarios.

Despite significant advancements, the practical deployment of Large Language Models (LLMs) is often hampered by their immense sizes, highlighting the need for effective compression techniques. Singular Value Decomposition (SVD) emerges as a promising method for compressing LLMs. However, existing SVD-based compression approaches suffer from substantial truncation losses, leading to severe performance degradation in compressed models. In this work, we introduce , a novel SVD-based LLM compression method that optimizes singular value truncation in SVD compression with two key strategies. First, employs dynamic compression ratio allocation to effectively balance the extremely large truncation loss across different layers. Second, it implements loss-optimized weight truncation to ensure that the truncated singular values result in a lower and more stable truncation loss in practice. We evaluate on ten datasets and five models on various scales and demonstrated that outperforms current state-of-the-art methods. The source code is available at https://github.com/AIoT-MLSys-Lab/SVD-LLM.

For pretrained language models such as Google’s BERT, recent research designs several input-adaptive inference mechanisms to improve the efficiency on cloud and edge devices. In this paper, we reveal a new attack surface on input-adaptive multi-exit BERT, where the adversary imperceptibly modifies the input texts to drastically increase the average inference cost. Our proposed slow-down attack called SlowBERT integrates a new rank-and-substitute adversarial text generation algorithm to efficiently search for the perturbation which maximally delays the exiting time. With no direct access to the model internals, we further devise a time-based approximation algorithm to infer the exit position as the loss oracle. Our extensive evaluation on two popular instances of multi-exit BERT for GLUE classification tasks validates the effectiveness of SlowBERT. In the worst case, SlowBERT increases the inference cost by 4.57×, which would strongly hurt the service quality of multi-exit BERT in practice, e.g., increasing the real-time cloud services’ response times for online users.

The state-of-the-art methods in aspect-level sentiment classification have leveraged the graph based models to incorporate the syntactic structure of a sentence. While being effective, these methods ignore the corpus level word co-occurrence information, which reflect the collocations in linguistics like “nothing special”. Moreover, they do not distinguish the different types of syntactic dependency, e.g., a nominal subject relation “food-was” is treated equally as an adjectival complement relation “was-okay” in “food was okay”. To tackle the above two limitations, we propose a novel architecture which convolutes over hierarchical syntactic and lexical graphs. Specifically, we employ a global lexical graph to encode the corpus level word co-occurrence information. Moreover, we build a concept hierarchy on both the syntactic and lexical graphs for differentiating various types of dependency relations or lexical word pairs. Finally, we design a bi-level interactive graph convolution network to fully exploit these two graphs. Extensive experiments on five bench- mark datasets show that our method outperforms the state-of-the-art baselines.