[Roadmap] vLLM Production Stack 2026 Roadmap

Router

• (P0) XpYd disaggregated prefill support
• (P1) Routing to external providers like OpenAI or Anthropic (#883 #902 )
• (P1) Router-side request queuing support (#876 #905 )
• (P2) Implement predictive routing based on future workloads
• (P2) Implement priority routing #996
• (P2) Improve router performance
  • Compile current router with Nuitka
  • Prototype Rust/Go/Nginx-based router frontend

Advanced Deployment

• (P0) Advanced autoscaling based on modern inference metrics
• (P0) Implement KV-cache–aware and prefix-aware routing logic in Gateway Inference extension and CRD (#881 )
• (P0) Implement disaggregated prefill support in CRD (#841 )
• (P0) Integrate KEDA-based scaling in CRD (#903 )
• (P1) Make existing deployment methods (Helm/CRD/Gateway Inference) more Kubernetes-native
• (P2) Support automatic configuration of engine parameters

Support for Modern Workloads

• (P0) Support deployment with vLLM Omni
• (P1) Intelligent routing for agent workloads and multimodal models
• (P2) End-to-end deployment for multimodal models
• (P2) First-class support for RL workloads

Safety and Security

• (P0) Request migration when the vLLM instance fails
• (P1) Ship router extension config with pre-defined safety policies, rate limits, and logging enabled
• (P1) Add optional PII detection in the request path

CI/CD and Misc

• (P0) Release bot for automated releases (Helm chart, K8s controller packages, Docker images)
• (P2) GitHub Actions for building router Docker images for multiple architectures
• (P2) Tutorials covering more cloud platforms, models, and features

If any of the items you wanted are not on the roadmap, your suggestion and contribution are strongly welcomed! Please feel free to comment in this thread, open a feature request, or create an RFC.

Happy vLLMing!

This is a broad roadmap spanning Router, Advanced Deployment, Modern Workloads, Safety and Security, and CI/CD, with no files or tests named. Start by selecting one unchecked item and reading its referenced issues where available; done means a single item has a defined scope and is implemented and validated according to that item's requirements.