kv-distill

Trainable Compression For LLMs

This is the repository for the work KV-Distill: Nearly Lossless Learnable Context Compression for LLMs. The code to reproduce all experiments and checkpoints will be made available by the end of March.