kserve/docs/samples/v1beta1/torchserve/v2/bert/bert_gpu.yaml

apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  name: "torchserve-bert-v2"
spec:
  predictor:
    pytorch:
      protocolVersion: v2
      storageUri: gs://kfserving-examples/models/torchserve/bert/v2
      resources:
        limits:
          memory: 4Gi
          nvidia.com/gpu: "1"