Training-free KV cache compression for LLMs. 10-33x compression via E8 lattice quantization + attention-aware token eviction. One line of code.
