Welch Labs 解析 DeepSeek 如何通过 MLA（Multi-head Latent Attention）技术重构 Transformer 架构，实现 KV 缓存减少 93.3% 和生成速度提升 6 倍以上。  

---

Url: https://www.youtube.com/watch?v=0VLAoVGf_74
Title: How DeepSeek Rewrote the Transformer [MLA]
Uploader: Welch Labs
Uploader ID: @WelchLabsVideo