Compaction

Code for Fast KV Compaction via Attention Matching. Attention Matching (AM) compacts a KV cache in latent space by constructing a smaller set of keys and values that reproduce the original attention behavior.

Repository Layout

Try Compaction

python -m examples.qa_demo --model Qwen/Qwen3-4B --target-size 0.1

This prefills a short article, compacts its KV cache to 10% with Attention Matching, and compares QA accuracy before and after. See examples/qa_demo.py for details.

Evaluate QA Tasks

Run the evaluator with one or more compaction methods and datasets:

python -m evaluation.run_qa_evaluation \
  --algorithm-config default \
  --methods original AM-HighestAttnKeys \
  --dataset-name quality \
  --n-articles 1 \
  --compute-stats 1

For non-uniform budgets, point --precomputed-budget-path to one of the JSON files under head_budget_optimization/head_budgets/.

Development Status

We might continue developing this into a more polished, installable package and are open to contributions. If you're interested in collaborating, feel free to open an issue or PR. See TODO for current plans.

Citation

If you found this work useful, please cite:

@misc{zweiger2026fastkvcompactionattention,
      title={Fast {KV} Compaction via {Attention Matching}}, 
      author={Adam Zweiger and Xinghong Fu and Han Guo and Yoon Kim},
      year={2026},
      eprint={2602.16284},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2602.16284}, 
}