Memory-Efficient Visual Autoregressive Modeling with Scale-Aware KV Cache Compression | Read Paper on Bytez