4 ms·MiniMax teased M3 Sparse Attention: 9.7x prefilling, 15.6x decoding at 1M9 points by rebekkamikkoa 4mo ago