Describe the bug

I have two preferred nodes with instance type Standard_NC80adis_H100_v5, which has 2 H100 each (total 4 GPUs). However, when I deployed the following workspace, the two pods generated by the managed StatefulSet are requesting 4 GPUs each:

apiVersion: kaito.sh/v1beta1
kind: Workspace
metadata:
  name: workspace-llama-3-3-70b-instruct
resource:
  count: 2
  instanceType: "Standard_NC80adis_H100_v5"
  labelSelector:
    matchLabels:
      node.kubernetes.io/instance-type: Standard_NC80adis_H100_v5
inference:
  preset:
    name: llama-3.3-70b-instruct
    presetOptions:
      modelAccessSecret: hf-token
  config: "llama-inference-params"
---
apiVersion: v1
kind: ConfigMap
metadata:
  name: "llama-inference-params"
data:
  inference_config.yaml: |
    vllm:
      cpu-offload-gb: 0
      gpu-memory-utilization: 0.95
      swap-space: 4
      max-model-len: 16384

Expected behavior

Each pod should request 2 GPUs, based on how many available GPUs there are in the preferred nodes

Logs

I was debugging it and noticed the field selector here is not working:

causing this function to fall back to the default number of GPUs required for the model

Environment

• Kubernetes version (use kubectl version):
• OS (e.g: cat /etc/os-release):
• Install tools:
• Others:

Additional context