Debating AFD Disaggregation and Data Locality in Inference Architecture

GavinSBaker · x · 2026-08-26

In a technical debate, Gavin Baker discussed the compatibility of Attention-FFN disaggregation (AFD) with data locality, specifically the constraint of not moving the KV cache. He argued that OpenAI's Jalapeño project did not make a bet against disaggregation, but rather focused on a topology compatible with data locality: computing prefill and attention on Jalapeño while offloading FFN to another chip. He noted that OpenAI is obviously doing relatively crude disaggregated inference (PD) at scale today. AFD is considered a hard networking problem that offers potential superiority over simpler PD disaggregation approaches.

Related event: KV Cache Offloading Remains Unsolved for Large-Context Agents(2 posts)→

Original post →

More from Infra

Infra channel →