Yifei Zuo releases Parallax, a local linear attention architecture that beats standard attention when paired with the Muon optimizer

A custom decode kernel matches or exceeds FlashAttention performance.