Fetching the paper…

TidalDecode: Fast and Accurate LLM Decoding with Position Persistent Sparse Attention · Around