NACL: A General and Effective KV Cache Eviction Framework for LLMs at Inference Time

2024/08/07 by Yilong Chen, Guoxia Wang, Chen, Yilong +17 · 12 citations
Computer Science · Social Sciences · Business, Management and Accounting · #Digital Rights Management and Security #Artificial Intelligence in Law #Financial Distress and Bankruptcy Prediction

paper · pdf · doi:10.48550/arxiv.2408.03675

Abstract

Large Language Models (LLMs) have ignited an innovative surge of AI applications, marking a new era of exciting possibilities equipped with extended context windows. However, hosting these models is cost-prohibitive mainly due to the extensive memory consumption of KV Cache involving long-context modeling. Despite several works proposing to evict unnecessary tokens from the KV Cache, most of them rely on the biased local statistics of accumulated attention scores and report performance using unconvincing metric like perplexity on inadequate short-text evaluation. In this paper, we propose NACL, a general framework for long-context KV cache eviction that achieves more optimal and efficient eviction in a single operation during the encoding phase. Due to NACL's efficiency, we combine more accurate attention score statistics in PROXY TOKENS EVICTION with the diversified random eviction strategy of RANDOM EVICTION, aiming to alleviate the issue of attention bias and enhance the robustness in maintaining pivotal tokens for long-context modeling tasks. Notably, our method significantly improves the performance on short- and long-text tasks by 80% and 76% respectively, reducing KV Cache by up to 50% with over 95% performance maintenance. The code is available at https://github.com/PaddlePaddle/Research/tree/master/NLP/ACL2024-NACL.

Cited by

• Seen, Said, or Forgotten? A Causal Audit of Visual KV Memory Across Dialog Turns
• KV Admission: Learning What to Write for Efficient Long-Context Inference
• Blink: Dynamic Visual Token Resolution for Enhanced Multimodal Understanding
• Cache What Lasts: Token Retention for Memory-Bounded KV Cache in LLMs
• Scalable Processing-Near-Memory for 1M-Token LLM Inference: CXL-Enabled KV-Cache Management Beyond GPU Limits
• Alleviating Forgetfulness of Linear Attention by Hybrid Sparse Attention and Contextualized Learnable Token Eviction
• MemDecay: Region-Aware KV Cache Eviction for Efficient LLM Agent Inference
• ThinKV: Thought-Adaptive KV Cache Compression for Efficient Reasoning Models
• AQUA: Attention via QUery mAgnitudes for Memory and Compute Efficient Inference in LLMs
• Spotlight Attention: Towards Efficient LLM Generation via Non-linear Hashing-based KV Cache Retrieval
• KV Cache Compression for Inference Efficiency in LLMs: A Review
• SmallKV: Small Model Assisted Compensation of KV Cache Compression for Efficient LLM Inference

Related

• LLMs Know What to Drop: Self-Attention Guided KV Cache Eviction for Efficient Long-Context Inference
• In-context KV-Cache Eviction for LLMs via Attention-Gate
• Stateful KV Cache Management for LLMs: Balancing Space, Time, Accuracy, and Positional Fidelity
• Efficient LLM Inference with Kcache
• Scissorhands: Exploiting the Persistence of Importance Hypothesis for LLM KV Cache Compression at Test Time