Salvato in:
| Autori principali: | , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2026
|
| Soggetti: | |
| Accesso online: | https://arxiv.org/abs/2603.14224 |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866911517621878784 |
|---|---|
| author | Yang, Xu Zhang, Jiapeng Zhao, Dongyang Chen, Guo Tang, Zhuo |
| author_facet | Yang, Xu Zhang, Jiapeng Zhao, Dongyang Chen, Guo Tang, Zhuo |
| contents | The KV cache in self-attention has emerged as a major bottleneck in long-context and large-batch inference for LLMs. Existing approaches often treat sparsity prediction and compression as separate modules, relying on auxiliary index structures to select relevant tokens, and on complex quantization schemes to reduce memory usage. This fragmented design introduces redundant overhead and limits scalability.
In this paper, we propose a novel paradigm: treating the compressed key representation not merely as storage, but as a self-indexing structure that directly enables efficient sparse attention. By designing a sign-based 1-bit vector quantization (VQ) scheme, our method unifies compression and retrieval in a single, hardware-friendly format. This approach eliminates the need for external indices or learning-based predictors, offering a lightweight yet robust solution for memory-constrained inference. All components are designed to be hardware-efficient and easy to implement. By implementing custom CUDA kernels, our method integrates seamlessly with FlashAttention, minimizing additional runtime and memory overhead. Experimental results demonstrate that our approach delivers both effectiveness and efficiency. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_14224 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Self-Indexing KVCache: Predicting Sparse Attention from Compressed Keys Yang, Xu Zhang, Jiapeng Zhao, Dongyang Chen, Guo Tang, Zhuo Machine Learning Artificial Intelligence The KV cache in self-attention has emerged as a major bottleneck in long-context and large-batch inference for LLMs. Existing approaches often treat sparsity prediction and compression as separate modules, relying on auxiliary index structures to select relevant tokens, and on complex quantization schemes to reduce memory usage. This fragmented design introduces redundant overhead and limits scalability. In this paper, we propose a novel paradigm: treating the compressed key representation not merely as storage, but as a self-indexing structure that directly enables efficient sparse attention. By designing a sign-based 1-bit vector quantization (VQ) scheme, our method unifies compression and retrieval in a single, hardware-friendly format. This approach eliminates the need for external indices or learning-based predictors, offering a lightweight yet robust solution for memory-constrained inference. All components are designed to be hardware-efficient and easy to implement. By implementing custom CUDA kernels, our method integrates seamlessly with FlashAttention, minimizing additional runtime and memory overhead. Experimental results demonstrate that our approach delivers both effectiveness and efficiency. |
| title | Self-Indexing KVCache: Predicting Sparse Attention from Compressed Keys |
| topic | Machine Learning Artificial Intelligence |
| url | https://arxiv.org/abs/2603.14224 |