Download deployment/k8s/gpu-node-pool.yaml from ACA050/aegislm: direct link, hf CLI and curl.
- Browser
- Download file 5.72 kB
-
https://huggingface.co/spaces/ACA050/aegislm/resolve/main/deployment/k8s/gpu-node-pool.yaml
- Command line
-
hf download hf://spaces/ACA050/aegislm/deployment/k8s/gpu-node-pool.yaml
-
curl -L -o gpu-node-pool.yaml https://huggingface.co/spaces/ACA050/aegislm/resolve/main/deployment/k8s/gpu-node-pool.yaml
5.72 kB
| # AegisLM GPU Node Pool Configuration | |
| # Kubernetes node pool configuration for GPU workloads | |
| # Multi-Agent Adversarial LLM Evaluation Framework | |
| # GPU Node Pool Configuration | |
| # This defines the node selector and taints for GPU workloads | |
| # Apply with: kubectl apply -f gpu-node-pool.yaml | |
| # Node Pool Configuration for GPU workloads | |
| apiVersion: v1 | |
| kind: ConfigMap | |
| metadata: | |
| name: aegislm-gpu-config | |
| namespace: aegislm | |
| labels: | |
| app: aegislm | |
| component: gpu-pool | |
| data: | |
| # Node selector labels | |
| node-selector.json: | | |
| { | |
| "workload": "benchmark", | |
| "accelerator": "nvidia" | |
| } | |
| # GPU tolerations | |
| gpu-tolerations.json: | | |
| [ | |
| { | |
| "key": "nvidia.com/gpu", | |
| "operator": "Exists", | |
| "effect": "NoSchedule" | |
| } | |
| ] | |
| # Resource limits per pod | |
| resource-limits.json: | | |
| { | |
| "nvidia.com/gpu": 1, | |
| "cpu": "4", | |
| "memory": "16Gi" | |
| } | |
| # Cluster Autoscaler Configuration | |
| # Add this to your cluster autoscaler deployment | |
| apiVersion: v1 | |
| kind: ConfigMap | |
| metadata: | |
| name: cluster-autoscaler-status | |
| namespace: kube-system | |
| labels: | |
| app: cluster-autoscaler | |
| data: | |
| # GPU node pool scaling configuration | |
| # Note: This is informational - actual autoscaler config depends on your cloud provider | |
| gpu-scaling-rules.yaml: | | |
| # GPU Node Pool Auto-Scaling Rules | |
| # | |
| # Scale UP condition: | |
| # pending_gpu_jobs > available_gpus | |
| # | |
| # Scale DOWN condition: | |
| # gpu_utilization < threshold for N minutes | |
| # | |
| # Example configuration for GKE: | |
| # gcloud container clusters update CLUSTER_NAME \ | |
| # --enable-autoscaling \ | |
| # --min-nodes=0 \ | |
| # --max-nodes=10 \ | |
| # --node-pool=gpu-pool | |
| # Scaling formula: | |
| # T_scale_up = max(1, pending_jobs - available_gpus) | |
| # T_scale_down = max(0, idle_minutes - cool_down_period) | |
| # Parameters: | |
| # - min_nodes: 0 (scale to zero when idle) | |
| # - max_nodes: 10 (max GPU nodes) | |
| # - scale_up_delay: 1 minute | |
| # - scale_down_delay: 10 minutes (cool down period) | |
| # - gpu_utilization_threshold: 30% (scale down if below) | |
| # Spot/Preemptive instance configuration (future): | |
| # - Use spot instances for cost optimization | |
| # - Handle spot interruptions gracefully | |
| # GPU Node Labels (for manual node labeling) | |
| apiVersion: v1 | |
| kind: ConfigMap | |
| metadata: | |
| name: node-labels | |
| namespace: aegislm | |
| labels: | |
| app: aegislm | |
| data: | |
| # Label GPU nodes with these commands: | |
| # kubectl label nodes <node-name> workload=benchmark | |
| # kubectl label nodes <node-name> accelerator=nvidia | |
| # kubectl label nodes <node-name> gpu-type=<type> # e.g., v100, a100, a10 | |
| label-commands.sh: | | |
| #!/bin/bash | |
| # Label GPU nodes for AegisLM workloads | |
| # For each GPU node: | |
| # kubectl label nodes $NODE_NAME \ | |
| # workload=benchmark \ | |
| # accelerator=nvidia \ | |
| # gpu-type=a10 \ | |
| # gpu-count=1 | |
| # GPU Resource Quota (optional - for multi-tenant) | |
| apiVersion: v1 | |
| kind: ResourceQuota | |
| metadata: | |
| name: gpu-quota | |
| namespace: aegislm | |
| labels: | |
| app: aegislm | |
| spec: | |
| hard: | |
| # Limit total GPUs across all pods | |
| nvidia.com/gpu: "10" | |
| # Limit CPU and memory | |
| requests.cpu: "40" | |
| requests.memory: "64Gi" | |
| limits.cpu: "80" | |
| limits.memory: "128Gi" | |
| scopeSelector: | |
| matchExpressions: | |
| - operator: In | |
| scopeName: PriorityClass | |
| values: ["high-priority", "normal-priority"] | |
| # LimitRange for GPU Pods | |
| apiVersion: v1 | |
| kind: LimitRange | |
| metadata: | |
| name: gpu-limits | |
| namespace: aegislm | |
| labels: | |
| app: aegislm | |
| spec: | |
| limits: | |
| # Default limits for GPU pods | |
| - type: Container | |
| default: | |
| nvidia.com/gpu: "1" | |
| cpu: "4" | |
| memory: "16Gi" | |
| defaultRequest: | |
| nvidia.com/gpu: "1" | |
| cpu: "2" | |
| memory: "8Gi" | |
| max: | |
| nvidia.com/gpu: "2" | |
| cpu: "8" | |
| memory: "32Gi" | |
| min: | |
| nvidia.com/gpu: "0" | |
| cpu: "100m" | |
| memory: "128Mi" | |
| # Pod-level limits | |
| - type: Pod | |
| max: | |
| nvidia.com/gpu: "2" | |
| cpu: "16" | |
| memory: "64Gi" | |
| # PriorityClass for GPU workloads | |
| apiVersion: scheduling.k8s.io/v1 | |
| kind: PriorityClass | |
| metadata: | |
| name: high-priority | |
| labels: | |
| app: aegislm | |
| value: 1000000 | |
| globalDefault: false | |
| description: "High priority for GPU-accelerated benchmark jobs" | |
| # PriorityClass for normal workloads | |
| apiVersion: scheduling.k8s.io/v1 | |
| kind: PriorityClass | |
| metadata: | |
| name: normal-priority | |
| labels: | |
| app: aegislm | |
| value: 0 | |
| globalDefault: true | |
| description: "Normal priority for standard workloads" | |
| # Horizontal Pod Autoscaler for GPU node pool (if using metrics-server) | |
| apiVersion: autoscaling/v2 | |
| kind: HorizontalPodAutoscaler | |
| metadata: | |
| name: aegislm-gpu-hpa | |
| namespace: aegislm | |
| labels: | |
| app: aegislm | |
| spec: | |
| scaleTargetRef: | |
| apiVersion: apps/v1 | |
| kind: Deployment | |
| name: aegislm-worker | |
| minReplicas: 0 | |
| maxReplicas: 10 | |
| metrics: | |
| - type: Resource | |
| resource: | |
| name: nvidia.com/gpu | |
| target: | |
| type: Utilization | |
| averageUtilization: 70 | |
| behavior: | |
| scaleDown: | |
| stabilizationWindowSeconds: 300 | |
| policies: | |
| - type: Percent | |
| value: 50 | |
| periodSeconds: 60 | |
| scaleUp: | |
| stabilizationWindowSeconds: 60 | |
| policies: | |
| - type: Percent | |
| value: 100 | |
| periodSeconds: 15 | |
| - type: Pods | |
| value: 2 | |
| periodSeconds: 15 | |
| selectPolicy: Max | |