ipex-llm/docker/llm/finetune/lora/kubernetes/templates/bigdl-lora-finetuning-tdx-job.yaml

{{- if eq .Values.TEEMode "tdx" }}
apiVersion: kubeflow.org/v2beta1
kind: MPIJob
metadata:
  name: bigdl-lora-finetuning-job
  namespace: bigdl-lora-finetuning
spec:
  slotsPerWorker: 1
  runPolicy:
    cleanPodPolicy: Running
  sshAuthMountPath: /home/mpiuser/.ssh
  mpiImplementation: Intel
  mpiReplicaSpecs:
    Launcher:
      replicas: 1
      template:
         spec:
           volumes:
           - name: nfs-storage
             persistentVolumeClaim:
               claimName: nfs-pvc
           - name: dev
             hostPath:
               path: /dev
           {{- if eq .Values.enableTLS true }}
           - name: ssl-keys
             secret:
               secretName: ssl-keys
           {{- end }}
           runtimeClassName: kata-qemu-tdx
           containers:
           - image: {{ .Values.imageName }}
             name: bigdl-ppml-finetuning-launcher
             securityContext:
              runAsUser: 0
              privileged: true
             command: ["/bin/sh", "-c"]
             args:
               - |
                  nohup python /ppml/bigdl_aa.py > /ppml/bigdl_aa.log 2>&1 &
                  sudo -E -u mpiuser bash /ppml/bigdl-lora-finetuing-entrypoint.sh                  
             env:
             - name: WORKER_ROLE
               value: "launcher"
             - name: WORLD_SIZE
               value: "{{ .Values.trainerNum }}"
             - name: MICRO_BATCH_SIZE
               value: "{{ .Values.microBatchSize }}"
             - name: MASTER_PORT
               value: "42679"
             - name: MASTER_ADDR
               value: "bigdl-lora-finetuning-job-worker-0.bigdl-lora-finetuning-job-worker"
             - name: DATA_SUB_PATH
               value: "{{ .Values.dataSubPath }}"
             - name: OMP_NUM_THREADS
               value: "{{ .Values.ompNumThreads }}"
             - name: LOCAL_POD_NAME
               valueFrom:
                 fieldRef:
                   fieldPath: metadata.name
             - name: HF_DATASETS_CACHE
               value: "/ppml/output/cache"
             - name: ATTESTATION_API_SERVICE_PORT
               value: "{{ .Values.attestionApiServicePort }}"
             - name: ENABLE_TLS
               value: "{{ .Values.enableTLS }}"
             volumeMounts:
             - name: nfs-storage
               subPath: {{ .Values.modelSubPath }}
               mountPath: /ppml/model
             - name: nfs-storage
               subPath: {{ .Values.dataSubPath }}
               mountPath: "/ppml/data/{{ .Values.dataSubPath }}"
             - name: dev
               mountPath: /dev
             {{- if eq .Values.enableTLS true }}
             - name: ssl-keys
               mountPath: /ppml/keys
             {{- end }}
    Worker:
      replicas: {{ .Values.trainerNum }}
      template:
        spec:
          runtimeClassName: kata-qemu-tdx
          containers:
          - image: {{ .Values.imageName }}
            name: bigdl-ppml-finetuning-worker
            securityContext:
              runAsUser: 0
              privileged: true
            command: ["/bin/sh", "-c"]
            args:
              - |
                  chown nobody /home/mpiuser/.ssh/id_rsa &
                  sudo -E -u mpiuser bash /ppml/bigdl-lora-finetuing-entrypoint.sh                  
            env:
            - name: WORKER_ROLE
              value: "trainer"
            - name: WORLD_SIZE
              value: "{{ .Values.trainerNum }}"
            - name: MICRO_BATCH_SIZE
              value: "{{ .Values.microBatchSize }}"
            - name: MASTER_PORT
              value: "42679"
            - name: MASTER_ADDR
              value: "bigdl-lora-finetuning-job-worker-0.bigdl-lora-finetuning-job-worker"
            - name: LOCAL_POD_NAME
              valueFrom:
                fieldRef:
                  fieldPath: metadata.name
            volumeMounts:
            - name: nfs-storage
              subPath: {{ .Values.modelSubPath }}
              mountPath: /ppml/model
            - name: nfs-storage
              subPath: {{ .Values.dataSubPath }}
              mountPath: "/ppml/data/{{ .Values.dataSubPath }}"
            - name: dev
              mountPath: /dev
            resources:
              requests:
                cpu: {{ .Values.cpuPerPod }}
              limits:
                cpu: {{ .Values.cpuPerPod }}
          volumes:
          - name: nfs-storage
            persistentVolumeClaim:
              claimName: nfs-pvc
          - name: dev
            hostPath:
              path: /dev
---
apiVersion: v1
kind: Service
metadata:
  name: bigdl-lora-finetuning-launcher-attestation-api-service
  namespace: bigdl-lora-finetuning
spec:
  selector:
    job-name: bigdl-lora-finetuning-job-launcher
    training.kubeflow.org/job-name: bigdl-lora-finetuning-job
    training.kubeflow.org/job-role: launcher
  ports:
    - name: launcher-attestation-api-service-port
      protocol: TCP
      port: {{ .Values.attestionApiServicePort }}
      targetPort: {{ .Values.attestionApiServicePort }}
  type: ClusterIP
---
{{- if eq .Values.enableTLS true }}
apiVersion: v1
kind: Secret
metadata:
  name: ssl-keys
  namespace: bigdl-lora-finetuning
type: Opaque
data:
  server.crt: {{ .Values.base64ServerCrt }}
  server.key: {{ .Values.base64ServerKey }}
{{- end }}

{{- end }}