Beyond Homogeneous Attention: Memory-Efficient LLMs via Fourier-Approximated KV Cache

VLMRe-assign community