RestoreKV

RestoreKV Checkpoints

Trained restore-token embeddings + LoRA adapters for RestoreKV (paper | project page), a budget-matched, single-pass plug-in that recovers full-cache behavior under aggressive query-agnostic KV cache eviction (built on KVzip).

Each checkpoint stores only the 8 restore-token embeddings and the rank-8 LoRA adapters (~0.4% of the backbone; ~41 MB). The base model weights are not included — load them from their original repositories.

Usage (KVPress)

from kvpress import RestoreKVPress
press = RestoreKVPress(
    compression_ratio=0.9375,                       # 16x
    checkpoint_path="llama3.1-8b_restorekv.pt",     # downloaded from this repo
    budget_matched=True,
)

Evaluate with the official KVPress protocol (RULER-4096, Llama-3.1-8B):

python evaluate.py --dataset ruler --data_dir 4096 \
  --model meta-llama/Llama-3.1-8B-Instruct \
  --press_name RestoreKV --compression_ratio 0.875 \
  --restore_checkpoint llama3.1-8b_restorekv.pt

For the *_plus.pt checkpoints, use the KVzip+ variant (--press_name RestoreKV_plus).

License

Released under CC BY-NC 4.0 (research / non-commercial use only). These checkpoints are distilled using a training mixture that contains LongAlpaca-derived self-study data, so they inherit its most restrictive component license (CC BY-NC 4.0).