DeltaLog: Deferred Materialization of Recurrent States for Linear Attention Decoding

TL;DR - DeltaLog speeds linear-attention decoding by storing recent recurrent-state updates in a bounded log and periodically merging them into a dense base state. This preserves model semantics while reducing memory traffic and improving serving performance.

• Replaces most full recurrent-state write-backs with compact update appends.
• Supports GDN, KDA, and RWKV6 in a prototype serving stack.
• Accelerates recurrent-state update kernels by up to 1.86× and cuts write traffic by up to 7.83×.
• Delivers 1.05–1.20× end-to-end speedups over dense recurrent baselines.