Streaming-LLM?

[regularfry]

https://github.com/mit-han-lab/streaming-llm is a neat-looking approach for extending the apparent context length of an LLM, but the way it does it made me think that it might be a better approach to streaming whisper than what's currently in examples/stream. It's a moderately invasive update to the model itself, but it looks fairly formulaic.

Has anyone got any intuition (or, better, any code) that might inform whether a) it could work for whisper.cpp; and b) might reduce latency overall?