Offloading with prefetch-aware memory optimization: enabling efficient single-GPU inference for large language models