Adding a worker group to an additional network

Name of Feature or Improvement

RDMA Networks

Description of Problem the Feature Should Solve

RDMA Networks for Efficient LLM Training

Describe the Solution You Would Like to See

Description of the proposed solution.
"workerGroupSpecs": [
{
"replicas": cluster.config.num_workers,
"minReplicas": cluster.config.num_workers,
"maxReplicas": cluster.config.num_workers,
"groupName": f"small-group-{cluster.config.name}",
"rayStartParams": {
"block": "true",
"num-gpus": str(worker_gpu_count),
"resources": worker_resources,
},
"template": V1PodTemplateSpec(
metadata=V1ObjectMeta(
annotations={
"k8s.v1.cni.cncf.io/networks": [,,...]
}
),
spec=get_pod_spec(
cluster,
[get_worker_container_spec(cluster)],
cluster.config.worker_tolerations,
)
),
}
],

Describe Alternatives You Have Considered

Description of any alternative solutions or features you have considered.

Additional Context

Add any other context, screenshots, console logs, etc. about the request here.

Start by locating the Python code that builds workerGroupSpecs and the Kubernetes V1PodTemplateSpec shown in the issue. Trace how worker pod metadata and networks are currently configured; the work is done when a worker group can accept one or more additional network names and the resulting pods attach to them.