Checklist

• 1. If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/sgl-project/sglang/discussions/new/choose Otherwise, it will be closed.
• 2. Please use English, otherwise it will be closed.

Motivation

In the early design of RBG, we gave the definition freedom of WorkloadType to users, but at the same time, it also increased the cost of understanding and use of users. But currently only Statefulset and LWS workloadtypes will be used during the deployment of the inference system. To reduce the cost of understanding and workload selection of users. we can encapsulate the complexity of workloadType selection by using a higher-level workload - InstanceSet to serve as the default workloadType option, capable of providing the capabilities of both StatefulSet and LeaderWorkSet.

1. InstanceSet offers both StatefulSet-style and LWS-style workload management.
2. Pod management is implemented based on a self-maintained Workload, which has higher design and development freedom, such as inplace-update and specified instances deletion capabilities.
3. Reduce our reliance on LWS/StatefulSet while also minimizing dependency on specific Kubernetes versions.

Related resources

No response