vLLM adds metadata-only EPD for DeepSeek V4.1 Flash

vLLM developers explain the engineering behind encoder disaggregation (EPD) in multimodal serving, separating encoding from prefill and decode stages to scale resources independently. The updates include metadata-only requests to avoid duplicate media preprocessing, dedicated encoder execution runners, remote embedding transfer integrations like NIXL and Mooncake, and cross-encoder output reuse. Benchmarks on DeepSeek V4.1 Flash show higher throughput and reduced TTFT compared to traditional colocated setups. Open article →