Why Only Two Layers Use Sliding Window Attention? Architecture Detail Sparks Debate

stochasticchasm · x · 2026-09-11

A technical discussion on a model that uses sliding window attention (SWA) on only its first two of 20 layers. The author speculates the prefill FLOPs savings seem modest but notes less global attention is cheaper, possibly echoing first-k-dense designs in MoEs. Cross-attention ablations likely existed, but letting the decoder attend directly to encoder hidden states within its own sequence — similar to MoDA's allow-KVs-from-earlier-layers approach — seems a good design choice.

Related event: Inference-First Architecture Sparks Debate: FP4 KV Cache and Pure CSA2 Compression in Focus(11 posts)→

Original post →

More from Research

Research channel →