Generation process very slow!

Hello, thank you for your excellent research work.
The results of this repo sound very impressive. The only problem is that it consumes a large amount of GPU memory—generating text with 200 characters already takes up 11GB of VRAM, which makes the generation process very slow. Is there any way to solve this problem?
Looking forward to your reply. ^^ ^^

The issue names no file, test, or generation entry point. Reproduce text generation at around 200 characters while measuring VRAM and runtime, then trace the generation path to identify the memory bottleneck; done means materially lower GPU memory use and faster generation without changing the expected output.

Written by the indexing model from the issue text.