Nvidia releases 8B model with learned 8x KV cache compression