New O(NlogN) Attention System Retains 97% Long-Context Accuracy

A developer introduced ALHR, a static binary-tree attention system with learnable hierarchical routing that reduces the number of keys used. It cuts memory use and scales VRAM better with tokens while retaining 97% accuracy on the long-context MQAR task.