KV cache middleware for 1M context on 12GB VRAM. Uses K vectors as a retrieval index to fetch V on-demand from system RAM. No model modification, no retraining. Drop-in HuggingFace cache replacement.

工欲善其事，必先利其器。给自己写的编辑器。