Skip to content

[Feature]: Paged KV Integration #124

@christ-tt

Description

@christ-tt

Is your feature request related to a problem?

  1. Performance Gain
  2. Remove all the mess with preparing hidden states / mask... in the executor. prepare_batch. With external paged kernel, we could simplify our executor workflow.

Describe the Solution you'd like

  1. Update KV Cache Manager to include Paged KV;
  2. Check how to leverage open-sourced paged KV kernel on MAC;

Alternatives Considered (Optional)

No response

Additional Context (Optional)

No response

Metadata

Metadata

Labels

enhancementNew feature or request

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions