When you should NOT use prefill/decode disaggregation @redhat
When you should NOT use prefill/decode disaggregation  @redhat
Uploaded September 2026 | Updated November 2026, 4 hours ago
Everyone's talking about PD disaggregation. Greg Pereira, Senior Machine Learning Engineer at Red Hat, explains when it actually costs you: 10 uncached input tokens and 1,000 output tokens means opening a KV transfer between two GPUs is more work than just running it on one. That's why llm-d ships a prefix-based PD decider.

From vLLM Office Hours #55: Mooncake + vLLM/llm-d Deep Dive. Full session: youtu.be/RFyeBEy1AP8

Join us live every other Thursday: https://red.ht/office-hours

#vLLM #llm-d #KVCache #AIInference #RedHat #MLOps
When you should NOT use prefill/decode disaggregationOpenShift 5 agentic troubleshooting | DemoGitOps Guide to the Galaxy (ep 109) | Critical User JourneysSupport new AI models on day zero with vLLMDeploy local AI agents to productionWhat is sovereign infrastructure?Prefill and decode want opposite things from your routerWhy AI compliance is more than just a checkboxHow to sandbox AI agents with OpenShell | AgentOps UnlockedWho powers the Fortune 500?Scaling with OrchestratorsOpenShift simplifies cloud-native ecosystems
Red Hat |

When you should NOT use prefill/decode disaggregation

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER