# multiplication-eval-deepseek-r1-distill-llama-70b.yaml
#
# Kubernetes Job to evaluate how well a language model solves basic
# multiplication problems using vLLM. The Job downloads and runs a Python script
# from SCRIPT_URL, which:
# - Downloads the specified Hugging Face model (MODEL_ID)
# - Generates 1,000 random multiplication problems
# - Queries the model for answers
# - Computes and prints accuracy to stdout

apiVersion: batch/v1
kind: Job
metadata:
  name: multiplication-eval-deepseek-r1-distill-llama-70b
  namespace: mk8s-docs-examples
spec:
  template:
    spec:
      runtimeClassName: nvidia
      tolerations:
      - key: "nvidia.com/gpu"
        operator: "Equal"
        value: "true"
        effect: "NoSchedule"
      containers:
      - name: vllm-eval-runner
        image: vllm/vllm-openai:v0.9.1
        command: ["/bin/bash"]
        args:
          - -c
          - |
            curl -L "$SCRIPT_URL" -o eval_multiplication.py && \
            python3 eval_multiplication.py "$MODEL_ID" --stdout
        env:
        - name: MODEL_ID
          value: deepseek-ai/DeepSeek-R1-Distill-Llama-70B
        - name: SCRIPT_URL
          value: https://docs.lambda.com/assets/code/eval_multiplication.py
        - name: HF_HOME
          value: /root/.cache/huggingface
        volumeMounts:
          - name: hf-cache
            mountPath: /root/.cache/huggingface
          - name: dev-shm
            mountPath: /dev/shm
        resources:
          limits:
            memory: "32Gi"
            nvidia.com/gpu: "4"
      volumes:
      - name: hf-cache
        persistentVolumeClaim:
          claimName: huggingface-cache
      - name: dev-shm
        emptyDir:
          medium: Memory
          sizeLimit: 1Gi
      restartPolicy: Never
  backoffLimit: 2
  ttlSecondsAfterFinished: 300