Cluster · 4 nodes × 4 GPUs
ResearchProductPlatformheld but stuck: job cannot start
Four nodes with four H100s each are shared by three teams. Research runs big multi-GPU training, Product runs customer-facing inference, Platform runs occasional tests. Start in plain Kubernetes and submit two 8-GPU training jobs, then switch to Kueue and do the same. The difference between "GPUs allocated" and "GPUs doing work" is the point.
Switching restarts the demo with the same starting jobs.