[hlts2]

Summary

• Fix the GPU health check so that nodes with GPU time-slicing are not reported as GPUCountMismatch
• Compute the expected GPU count as nvidia.com/gpu.count × nvidia.com/gpu.replicas
• Treat a missing or invalid nvidia.com/gpu.replicas label as 1, so the behaviour for non-shared GPU nodes is unchanged
• HasGPU is unchanged and still decides GPU nodes from nvidia.com/gpu.count only

Problem

With GPU time-slicing, nvidia.com/gpu.count reports the number of physical GPUs, while allocatable nvidia.com/gpu is physical GPUs × nvidia.com/gpu.replicas (NVIDIA GPU Operator docs).

Labels:
  nvidia.com/gpu.count=1
  nvidia.com/gpu.replicas=4
  nvidia.com/gpu.sharing-strategy=time-slicing
Allocatable:
  nvidia.com/gpu: 4

The checker expected 1 but saw 4, so healthy time-sliced nodes failed the GPU check and became reboot targets after the 10-minute threshold.