Confirm the symptom in your cluster events/logs.
Look for messages similar to the following, which indicate he selector can’t be satisfied by the current node group configuration when L40S isn’t supported by the label.

• “Attempt Scale Up… pod is pending but its affinity/anti-affinity could not be satisfied…”

• “NodeSelector: { "spotinst.io/gpu-type": "nvidia-l40s" }”
  
  

Remove the unsupported L40S spotinst.io/gpu-type selector.
In your pod manifest, remove usage of:

• spotinst.io/gpu-type: nvidia-l40s

This prevents the scheduler/autoscaler from blocking on a label value that isn’t recognized/satisfied.

Request a GPU using standard Kubernetes GPU resources.
In your container spec, request/limit a GPU so Kubernetes schedules the pod onto a GPU-capable node: 

resources:
  limits:
    nvidia.com/gpu: 1

This is the standard mechanism for GPU workloads and works with GPU-enabled nodes.

Create a dedicated VNG for L40S-capable instances.
Configure a Virtual Node Group that uses:

• A GPU-capable machine image (for example, a GPU-ready image/AMI in your cloud)

• Only the instance types that provide L40S hardware (so the VNG scales with the right capacity)

This ensures Ocean scales the correct nodes when a GPU pod is pending. For related configuration options, see this documentation.

Add a custom node label to the L40S VNG.
Add a custom label on nodes created by the L40S VNG (choose a label key/value that works for your environment), such as:

• workload.gpu.family: l40s

• or spot.ocean.vng: l40s

This gives you a reliable, supported way to target the correct VNG.

Target the L40S VNG using nodeSelector or node affinity.
In your pod manifest, select nodes using the custom label you configured on the L40S VNG, for example:

nodeSelector:
  workload.gpu.family: l40s

Or use node affinity if you need more control. This approach avoids relying on spotinst.io/gpu-type for a GPU model that isn’t currently supported.

(Optional) Mix GPUs by using multiple VNGs.
If you run both A10 and L40S workloads, create separate VNGs (one for A10, one for L40S) and steer workloads using labels/affinity. This prevents scheduling collisions and keeps scaling predictable.