thu-nics/C2C

Enables direct KV-Cache projection and fusion between LLMs using learnable transformer-based projectors, eliminating the need for text serialization. Supports multi-model setups where multiple "sharer" models can contribute cached representations to a single "receiver" model through layer-wise alignment. Provides the `RosettaModel` wrapper for seamless integration with any HuggingFace-compatible LLM, with pre-trained fusers available on HuggingFace Hub.

361 stars.