There are some transformation approaches to re-use the kv cache across inferences, but none are in wide use due to accuracy concerns following the transformation.

When your abstract was clearly generated by an LLM and not curated to at least make it sound human, it does not make me want to read your paper.