Skip to content

[RFC]: Support Prefix Cache and P-D transfer use UCM(Unified Cache Manager) #4402

@ygwpz

Description

@ygwpz

Motivation.

Using KV Pool for prefix caching has definitely expand the capacity of KV cache we can store, but it still has some problems:

  1. Capacity is limited by the size of the pooled memory. Compute and storage are tied together, so we can’t scale capacity.
  2. It doesn’t have capability of persistence or reliability that’s actually needed for prefix caching to work in real-world production systems.

For prefix cache, the capacity and read/write speeds fit right into the sweet spot of traditional storage services. If we switch to a tiered cache setup—using DRAM on each node as cache, and a unified file system (3FS, or an enterprise-grade centralized storage) as the main storage—then that second issue I mentioned is just something the storage backend naturally handles. Check out the diagram below for how this works:
Image

Proposed Change.

Auto Prefix Cache

We want to push UCM (Unified Cache Manager) into vllm-ascend community to support Prefix Cache feature:

  1. add a new UCMConnector class, this class only delivery request to UCMConnectorImpl(reference lmcache to vLLM)
  2. add document about how to enable prefix cache by UCM User Guide
Image

P-D Transfer

More Important, UCM can support P-D transfer. We can use vLLM-Ascend + UCM + 3FS to build a P-D Disaggregate Inference System, User Guide:

Image

thanks for review our pull request: #4411

Feedback Period.

2025/11/30

CC List.

@wangxiyuan
@Yikun

Any Other Things.

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    RFCRequest For Comments

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions