vLLM documentation: Disaggregated Prefilling (experimental)

Fields

id

sources/pap-sys-vllm-disagg-prefill

URL

https://docs.vllm.ai/en/latest/features/disagg_prefill/

Publisher

vLLM project

Kind

repo

Accessed

2026-10-03

notes

States plainly that the feature is experimental, that 'Disaggregated prefill DOES NOT improve throughput', and that vLLM relies on third-party connectors (ExampleConnector, LMCacheConnectorV1, NixlConnector) for production-level disaggregated prefilling. Evidence that prefill/decode disaggregation is shipped but NOT a default or a throughput win.

Notes

States plainly that the feature is experimental, that 'Disaggregated prefill DOES NOT improve throughput', and that vLLM relies on third-party connectors (ExampleConnector, LMCacheConnectorV1, NixlConnector) for production-level disaggregated prefilling. Evidence that prefill/decode disaggregation is shipped but NOT a default or a throughput win.

Cross-references

This record references no other records.