CAOTE: KV Cache Selection for LLMs via Attention Output Error-Based Token Eviction
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2025
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866915533210779648 |
|---|---|
| author | Goel, Raghavv Park, Junyoung Gagrani, Mukul Jones, Dalton Morse, Matthew Langston, Harper Lee, Mingu Lott, Chris |
| author_facet | Goel, Raghavv Park, Junyoung Gagrani, Mukul Jones, Dalton Morse, Matthew Langston, Harper Lee, Mingu Lott, Chris |
| contents | While long context support of large language models has extended their abilities, it also incurs challenges in memory and compute which becomes crucial bottlenecks in resource-restricted devices. Token eviction, a widely adopted post-training methodology designed to alleviate the bottlenecks by evicting less important tokens from the cache, typically uses attention scores as proxy metrics for token importance. However, one major limitation of attention score as a token-wise importance metrics is that it lacks the information about contribution of tokens to the attention output. In this paper, we propose a simple eviction criterion based on the contribution of cached tokens to attention outputs. Our method, CAOTE, optimizes for eviction error due to token eviction, by seamlessly integrating attention scores and value vectors. This is the first method which uses value tokens on top of attention-based eviction scores in closed-form. Additionally, CAOTE can act as a meta-heuristic method with flexible usage with any token eviction method. We show that CAOTE, when combined with the state-of-the-art attention score-based methods, always improves accuracies on the downstream task, indicating the importance of leveraging information from values during token eviction process. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2504_14051 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | CAOTE: KV Cache Selection for LLMs via Attention Output Error-Based Token Eviction Goel, Raghavv Park, Junyoung Gagrani, Mukul Jones, Dalton Morse, Matthew Langston, Harper Lee, Mingu Lott, Chris Machine Learning Computation and Language While long context support of large language models has extended their abilities, it also incurs challenges in memory and compute which becomes crucial bottlenecks in resource-restricted devices. Token eviction, a widely adopted post-training methodology designed to alleviate the bottlenecks by evicting less important tokens from the cache, typically uses attention scores as proxy metrics for token importance. However, one major limitation of attention score as a token-wise importance metrics is that it lacks the information about contribution of tokens to the attention output. In this paper, we propose a simple eviction criterion based on the contribution of cached tokens to attention outputs. Our method, CAOTE, optimizes for eviction error due to token eviction, by seamlessly integrating attention scores and value vectors. This is the first method which uses value tokens on top of attention-based eviction scores in closed-form. Additionally, CAOTE can act as a meta-heuristic method with flexible usage with any token eviction method. We show that CAOTE, when combined with the state-of-the-art attention score-based methods, always improves accuracies on the downstream task, indicating the importance of leveraging information from values during token eviction process. |
| title | CAOTE: KV Cache Selection for LLMs via Attention Output Error-Based Token Eviction |
| topic | Machine Learning Computation and Language |
| url | https://arxiv.org/abs/2504.14051 |