mirror of
https://github.com/luxfi/lux-gpu-ci.git
synced 2026-10-30 16:20:42 +00:00
zeekay
dde006617a
EKS cluster with Karpenter v1 NodePool that scales g4dn.xlarge spot instances 0..1 on demand for CUDA workflows. ARC RunnerScaleSet registers as [self-hosted, linux, amd64, cuda, gpu, t4]. Components: - terraform/ — VPC, EKS 1.31, system node group, Karpenter IAM/SQS - helm/karpenter-nodepool.yaml — v1 NodePool + EC2NodeClass for GPU spot - helm/nvidia-device-plugin-values.yaml — DS with GPU taint toleration - helm/arc-values.yaml + helm/runner-scaleset.yaml — ARC v0.9.x - bootstrap.sh — one-shot deploy - Makefile — init/apply/destroy/verify/cost-estimate Cost: ~$138/mo idle (EKS control plane + tiny system node + NAT), ~$0.027/build for a 10-minute g4dn.xlarge spot run.
26 lines
586 B
YAML
26 lines
586 B
YAML
# NVIDIA k8s-device-plugin DaemonSet — exposes nvidia.com/gpu resource.
|
|
# Tolerates the GPU taint so it actually runs on GPU nodes.
|
|
# Targets only nodes that advertise the nvidia.com/gpu.present=true label.
|
|
|
|
nodeSelector:
|
|
nvidia.com/gpu.present: "true"
|
|
|
|
tolerations:
|
|
- key: nvidia.com/gpu
|
|
operator: Equal
|
|
value: "present"
|
|
effect: NoSchedule
|
|
- key: CriticalAddonsOnly
|
|
operator: Exists
|
|
|
|
# AL2023 ships the runtime; default config is fine.
|
|
runtimeClassName: ""
|
|
|
|
resources:
|
|
requests:
|
|
cpu: "10m"
|
|
memory: "32Mi"
|
|
limits:
|
|
cpu: "100m"
|
|
memory: "128Mi"
|