# vLLM separates multimodal inference stages so each can scale independently

> Original title: Multimodal workloads don’t scale uniformly. More images shouldn’t mean scaling the entire language-model tier.

- Source: vLLM
- Published: 2026-11-11T10:03:44.000Z
- Bad Influence: https://badinfluence.news/items/u2j60hdzxv641tk471y6qmmh5
- Original: https://x.com/vllm_project/status/2120457439090753959

## Summary

vLLM's EPD design splits multimodal inference into encoding, prefill, and decode stages that scale independently, so adding images does not require scaling the entire language-model tier. The post says DeepInfer also uses EPD disaggregated deployment to allocate resources flexibly across those stages.

## Post

Multimodal workloads don’t scale uniformly. More images shouldn’t mean scaling the entire language-model tier.

EPD in vLLM separates encoding, prefill, and decode so each stage can scale independently. DeepInfer also uses EPD disaggregated deployment to allocate resources flexibly across stages.