vLLM Workshop
AFD Plugin: Attention–FFN Disaggregation in vLLM
Date, time, and room will be added once confirmed.
Talk overview
This workshop introduces AFD Plugin, an external vLLM plugin that separates Attention and FFN or MoE execution into independently configurable roles. It explains activation exchange, synchronous decode, asynchronous prefill, microbatching, communication–computation overlap, and performance tradeoffs demonstrated with DeepSeek-V3.2 on Ascend 910C.
