[ICML2025] SpargeAttention: A training-free sparse attention that accelerates any model inference. - View it on GitHub