vllm-project / vllm-project/aibrix
[RFC] Enhancing KV Cache Orchestration for AIbrix Components
- Dominant language
- Go
- Stars
- 5.1k
- Forks
- 694
- Avg merge
- 1d 20h
- Merged PRs (30d)
- 98
Description
### 🚀 Feature Description and Motivation
Currently, we are leveraging the Vineyard Operator to orchestrate workloads. While it provides a foundation, we've extended the upstream operator with advanced scheduling features to better align with our needs. However, several challenges remain:
- Limited Compatibility with Use Cases: Many existing operator features do not integrate well with our specific use cases, requiring significant custom modifications.
- Insufficient Failure Handling: The original operator lacks a robust design for handling failures effectively, leading to potential operational inefficiencies. like deleting etcd doesn't trigger the reconcilation
- Kubernetes Version Compatibility Issues: The operator struggles with compatibility across different Kubernetes versions, posing challenges for upgrades and maintenance.
Given these limitations, we propose evaluating the feasibility of using AIbrix components to orchestrate the KV cache pool. This shift would enable us to:
- Leverage AIbrix’s flexibility to build a robust and scalable KV orchestration system.
- Introduce advanced failure handling and improved version compatibility.
- Ensure the design is pluggable and adaptable for future integrations.
Additionally, since we do not yet have a finalized model API design, the KV orchestration design must account for potential migration paths to a model API in the future. This will ensure that the system remains extensible and capable of supporting evolving requirements.
### Use Case
_No response_
### Proposed Solution
- Assessment: Evaluate AIbrix components for KV orchestration and identify potential integration points.
- Pluggable Design: Propose a modular and pluggable architecture for KV orchestration.
- Migration Path: Define a strategy for transitioning to a model API once the design is finalized.
Feedback and contributions are welcome to ensure the proposed approach addresses both current needs and future scalability.
Contributor guide
Assessment
This issue has not been assessed yet.