New deepseek paper: Natively Trainable Sparse Attention mechanism

Authored and Uploaded by none others than Liang Wenfeng himself