fine-tune-3x1-l4.yaml

apiVersion: jobset.x-k8s.io/v1alpha2
kind: JobSet
metadata:
  name: pytorch
  
spec:
  replicatedJobs:
  - name: workers
    template:
      spec:
        parallelism: 3
        completions: 3
        backoffLimit: 0
        template:
          metadata:
            annotations:
               gke-gcsfuse/volumes: "true"
          spec:
           nodeSelector:
             cloud.google.com/gke-accelerator: nvidia-l4 
           serviceAccountName: csi-storage
           volumes:
           - name: gcs-fuse-csi-ephemeral
             csi:
               driver: gcsfuse.csi.storage.gke.io
               readOnly: true
               volumeAttributes:
                  bucketName: "mlops-repo"
                  mountOptions: "implicit-dirs"
                  gcsfuseLoggingSeverity: warning
           - name: hf-cache
             emptyDir: {}
           - name: model-cache
             emptyDir: {}
           - name: output
             emptyDir: {}
              
           containers:
            - name: gke-gcsfuse-sidecar
              image: gke.gcr.io/gcs-fuse-csi-driver-sidecar-mounter:v1.2.0-gke.0@sha256:31880114306b1fb5d9e365ae7d4771815ea04eb56f0464a514a810df9470f88f
            - name: pytorch
              image: us-east1-docker.pkg.dev/rick-vertex-ai/gke-llm/llama-factory:latest
              ports:
              - containerPort: 3389
              env:
              - name: HF_TOKEN
                valueFrom:
                  secretKeyRef:
                    name: huggingface
                    key: HF_TOKEN
              - name: HUGGING_FACE_HUB_TOKEN
                valueFrom:
                  secretKeyRef:
                     name: huggingface
                     key: HF_TOKEN
              - name: MASTER_ADDR
                value: "pytorch-workers-0-0.pytorch"
              - name: MASTER_PORT
                value: "3389"
              - name: RANK
                valueFrom:
                  fieldRef:
                    fieldPath: metadata.annotations['batch.kubernetes.io/job-completion-index']
              # Force python to not buffer output and write directly to stdout, so we can view training logs via `kubectl logs`.
              - name: PYTHONUNBUFFERED
                value: "0"
              resources:
                limits:
                  nvidia.com/gpu: 1
              volumeMounts:
              - name: gcs-fuse-csi-ephemeral
                mountPath: /data
                readOnly: false
              - name: hf-cache
                mountPath: /root/.cache/huggingface
              - name: model-cache
                mountPath: /root/.cache/modelscope
              - name: output
                mountPath: /app/output
              command:
              - bash
              - -xc
              - |
                FORCE_TORCHRUN=1 NNODES=3 RANK=$RANK MASTER_ADDR=$MASTER_ADDR MASTER_PORT=$MASTER_PORT llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml