vLLM AFD Plugin is a new experimental external plugin that introduces Attention-FFN Disaggregation (AFD) for Mixture-of-Experts (MoE) model serving. It separates Attention and FFN computation into independently deployable services, allowing each to scale with its own rank topology. The plugin supports NVIDIA GPUs and Ascend NPUs, synchronous and asynchronous connectors, DeepSeek V2/V3-family models, and multiple execution paths (eager, graph, dual-batch). Performance benchmarks on Ascend 910C show that a 64A16F AFD topology achieves up to +11.3% normalized decode throughput over a conventional EP64 baseline, while asynchronous prefill AFD reduces median TTFT by ~47% at 12 req/s. The plugin integrates via vLLM's plugin entry point without modifying the vLLM source tree, preserving the OpenAI-compatible serving interface. It is currently experimental and pinned to vLLM 0.19.1.
