EpiKV Scores KV Cache Tokens via Representation Change, No Attention Matrix Needed

EpiKV replaces attention-weight-based token scoring with an epiphany score — the change in internal model representations read directly from the forward pass — enabling KV cache eviction compatible with FlashAttention without materializing the attention matrix. The method requires no training or custom kernels and scales to 16x longer feasible context at a 4096-token cache budget.