[Roadmap] UCM Roadmap Q1 2026

[yuanzhg078]

This roadmap is a living guide to Unified Cache Manager (UCM)’s evolution. We’re shaping its Q1 2026 direction and welcome your input—your feedback will help us prioritize what matters most for the community and production use.

Primary goal for 2026 Q1:
Refactor the UCM Store and sparse attention architectures to enhance modularity and stability, and optimize performance, while broadening northbound and southbound compatibility with more inference engines and storage backends.

Core

• Store
  • Refactor UCM Store Architecture
  • PipelineStore Framework Upgrade and Optimization
  • Add Layerwise Connector
  • KVCompress
  • Adapt Ds3fs Store
• Sparse
  • Spare Attention Framework Upgrade(v2)
  • Spare Attention Offload Performance Optimization
  • GSA/ESA on device
• Adapt SGLang
  • Support Prefix Cache
• CacheBlend Framework Optimization
• Model Window Extrapolation-Rerope

CI/CD

• Workflow (build/unittest/install/e2e inference)
• Correctness test (continuously updated)

Test

• Unified LLM interface and simplified test configuration
• Expanded accuracy validation (offline test + UC Eval)
• Added pre-run environment verification

Others

• Observability: Metrics Optimization
• Add UCM Logger module

Documentation & community management

• Add Contributing Guide
• Add Code of Conduct

[IgniteGo]

Thank you for your excellent work!
I understand that UCM is currently integrated with vLLM through a runtime patching mechanism. May I ask if there are any plans to integrate this project directly into vLLM or SGLang in the future?
I would greatly appreciate it if you could spare some time to answer this.

[yuanzhg078]

Thank you for your excellent work! I understand that UCM is currently integrated with vLLM through a runtime patching mechanism. May I ask if there are any plans to integrate this project directly into vLLM or SGLang in the future? I would greatly appreciate it if you could spare some time to answer this.

We have already merged the UCM prefix cache functionality into vllm-ascend. You can find more details in the vllm-ascend documentation.
Moving forward, we are continuously developing features within the UCM community to ensure full compatibility with both vLLM and SGLang. We are currently in active communication and collaboration with these communities to integrate the project directly into their main branches as soon as possible.

[IgniteGo]

Thank you for your excellent work! I understand that UCM is currently integrated with vLLM through a runtime patching mechanism. May I ask if there are any plans to integrate this project directly into vLLM or SGLang in the future? I would greatly appreciate it if you could spare some time to answer this.

We have already merged the UCM prefix cache functionality into vllm-ascend. You can find more details in the vllm-ascend documentation. Moving forward, we are continuously developing features within the UCM community to ensure full compatibility with both vLLM and SGLang. We are currently in active communication and collaboration with these communities to integrate the project directly into their main branches as soon as possible.

Thank you for your prompt reply!
Looking forward to the deep integration of UCM with the main branches of vLLM and SGLang. Thanks to the great work!