# vLLM explains encoder disaggregation for scaling multimodal serving

> Original title: Encoder Disaggregation in vLLM: Scaling Multimodal Serving with EPD

- Source: vLLM Blog
- Published: 2026-11-11T00:00:00.000Z
- Bad Influence: https://badinfluence.news/items/epdoj4thfrglgn4c2z263x3vc
- Original: https://vllm.ai/blog/2026-11-11-encoder-disaggregation-epd

## Summary

vLLM's blog explains how separating the multimodal encoder into its own instance, using EPD disaggregation, changes serving. Metadata-only requests let downstream instances skip media preprocessing and receive embeddings through connectors. In a DeepSeek V4.1 Flash test on GB200, EPD delivered 6.6 to 9.5% higher throughput per GPU than a PD baseline at concurrency 64 to 128, with P90 TTFT 44 to 47% lower, counting encoder GPUs.

## Why it matters

The post explains how separating the encoder stage into metadata-only requests and dedicated execution changes multimodal serving, with measured throughput and latency results.