CoKV introduces a cooperative game theory framework to intelligently allocate memory resources in large language models, addressing the critical challenge of KV cache memory consumption.

• Treats attention heads as players in a cooperative game to determine optimal resource distribution
• Achieves up to 40% memory reduction with minimal performance impact
• Enables more efficient LLM deployment on resource-constrained devices
• Outperforms existing eviction-based methods by considering head interdependencies

This engineering advancement makes deploying large language models more practical and cost-effective across diverse hardware environments, potentially enabling wider adoption of LLM technology in production systems.

CoKV: Optimizing KV Cache Allocation via Cooperative Game