3 ms·Attention Became Efficient and Scalable: KV Caching, MQA, GQA, MLA, and DSA2 points by ibobev 2mo ago