question about sink attention

Author: hujiaxin0Created Nov 9, 2024Updated Nov 9, 2024

Thank you very much for your open source work abort streaming-llm!

Here I have a question about sink attention.

By debugging this program, I didn't find the process of generating the sink attention mask. Is casual attention still used to generate the first token?

Would you please provide a link to the code for generating the sink attention mask?

Source: mit-han-lab/streaming-llm