vLLM gains an experimental Attention-FFN disaggregation plugin for MoE

StepFun's post is a short thank-you crediting the vLLM team, Ant Group, and "FastAFD" for bringing Attention-FFN Disaggregation, or AFD, to the open-source community, but it does not explain what AFD does. That detail comes from vLLM's own plugin announcement, published as an experimental release.

AFD separates a mixture-of-experts model's attention step, which is mostly limited by memory access speed, from its expert, or feed-forward, step, which is limited by raw compute, letting each run as an independently scaled service instead of sharing one GPU that cannot be tuned well for both at once. vLLM's plugin currently supports DeepSeek's V2 and V3 model family, including V3.2, and GLM's MoE DSA architecture.