Zhi Chen

Yifei Yang, Zouying Cao, Qiguang Chen et al.

The development of large language models (LLMs) has significantly expanded model sizes, resulting in substantial GPU memory requirements during inference. The key and value storage of the attention map in the KV (key-value) cache accounts for more than 80\% of this memory consumption. Nowadays, most existing KV cache compression methods focus on intra-layer compression within a single Transformer layer but few works consider layer-wise compression. In this paper, we propose a plug-and-play method called \textit{KVSharer}, which shares the KV cache between layers to achieve layer-wise compression. Rather than intuitively sharing based on higher similarity, we discover a counterintuitive phenomenon: sharing dissimilar KV caches better preserves the model performance. Experiments show that \textit{KVSharer} can reduce KV cache computation by 30\%, thereby lowering memory consumption without significantly impacting model performance and it can also achieve at least 1.3 times generation acceleration. Additionally, we verify that \textit{KVSharer} is compatible with existing intra-layer KV cache compression methods, and combining both can further save memory.

Da Ma, Lu Chen, Situo Zhang et al.

The rapid expansion of context window sizes in Large Language Models~(LLMs) has enabled them to tackle increasingly complex tasks involving lengthy documents. However, this progress comes at the cost of a substantial increase in memory usage during inference, primarily due to the linear growth of the key-value~(KV) cache. Existing KV cache compression methods often discard less relevant tokens, which can lead to significant performance degradation when critical information is lost. In this paper, we propose \textsc{PoD}~(Proximal tokens over Distant tokens), a novel KV cache compression framework that allocates memory according to token importance, retaining less important tokens in a more compact, shared form rather than discarding them entirely. Our approach is motivated by two key observations: (1) proximal tokens -- those at the beginning and end of the context -- are significantly more important for next-token prediction, and (2) attention scores for distant tokens are highly redundant across consecutive layers. Leveraging these insights, \textsc{PoD} preserves the full KV cache for proximal tokens, while for distant tokens, it shares key states across layers. Since attention scores are determined by both queries and keys, sharing key states enables multiple layers to reuse a single set of keys for distant tokens, substantially reducing KV cache memory without discarding essential context. We further introduce a lightweight post-training adaptation to enable the model to adjust to this new attention-sharing structure. Extensive experiments on both synthetic~(Needle in a Haystack) and real-world long-context benchmarks demonstrate that \textsc{PoD} reduces KV cache memory usage by up to 35\% without compromising performance. Our method is orthogonal to existing token-selection-based techniques and can be combined with them for further KV cache compression.

Yifei Yang, Zouying Cao, Xinbei Ma et al.

Training Large Language Models (LLMs) from scratch requires immense computational resources, making it prohibitively expensive. Model scaling-up offers a promising solution by leveraging the parameters of smaller models to create larger ones. However, existing depth scaling-up methods rely on empirical heuristic rules for layer duplication, which result in poorer initialization and slower convergence during continual pre-training. We propose \textbf{LESA}, a novel learnable method for depth scaling-up. By concatenating parameters from each layer and applying Singular Value Decomposition, we uncover latent patterns between layers, suggesting that inter-layer parameters can be learned. LESA uses a neural network to predict the parameters inserted between adjacent layers, enabling better initialization and faster training. Experiments show that LESA outperforms existing baselines, achieving superior performance with less than half the computational cost during continual pre-training. Extensive analyses demonstrate its effectiveness across different model sizes and tasks.