[Feature]: Support for preflight/prolog checks

[lalitadithya]

Prerequisites

• I searched existing issues

Feature Summary

As a part of Kubernetes 1.35, support for gang scheduling was introduced -- https://kubernetes.io/blog/2025/12/29/kubernetes-v1-35-introducing-workload-aware-scheduling/. As a part of this support, the pod spec now has a reference to the gang that the pod is a part of:

apiVersion: v1
kind: Pod
metadata:
  name: worker-0
  namespace: some-ns
spec:
  workloadRef:
    name: training-job-workload
    podGroup: workers

Using workloadRef NVSentinel should be able to inject an init container that can run a variety of checks including but not limited to:

• DCGM diag
• NCCL loopback
• NCCL all-reduce across a gang
• Pluggable third party checks

Problem/Use Case

As an operator, I want to run preflight checks before a pod starts that allow me to run active health checks to give the user jobs a higher chance of success

Proposed Solution

TBD

Component

Health Monitor