Tag
HeadWiseKV is a training-free framework that compresses KV caches in hybrid long-context language models, reducing GPU memory usage and extending context lengths while maintaining quality.