Amazon EKS supports several GPU operating models. EKS Auto Mode manages more of the node lifecycle and accelerator stack, while managed node groups and self-managed Karpenter leave more choices with the platform team. Polyaxon connects after that boundary is defined and keeps workload policy in queues and presets.

See the Amazon EKS AI/ML compute for the upstream configuration and requirements.

Prerequisites

• A documented EKS compute model: Auto Mode, managed node groups, or self-managed Karpenter.
• GPU nodes using a compatible accelerated AMI and one reviewed NVIDIA DRA driver or device-plugin path.
• EFA-capable instances and an advertised EFA resource when the workload requires multi-node communication.
• The cluster connected to Polyaxon by installing Polyaxon Agent (compute cluster), with a queue, storage connection, and pod identity configured for the workload namespace.

Connect the cluster

Follow Connect Kubernetes to Polyaxon to install Polyaxon Agent (compute cluster), configure storage and queues, and submit the shared GPU visibility example. Use the same namespace and queue that will run your workloads.

The example only checks GPU allocation and driver visibility. Validate storage, identity, and multi-node networking separately using the provider requirements below.

Provider requirements

• Record whether AWS or your platform team owns the node lifecycle, AMI, GPU device layer, EFA integration, and interruption handling.
• Treat each instance family, AMI, GPU allocation mechanism, and network topology as a separately validated capacity class.
• Keep node selectors, tolerations, identity, and capacity type in platform-owned queues and presets rather than application components.
• Test interruption and node rotation behavior before placing long-running training or stateful inference on reclaimable capacity.
• Run NCCL tests across the intended EFA topology before accepting a distributed-training queue.

For distributed training, continue with NCCL and RCCL validation.

For help choosing a provider, see Choosing GPU infrastructure for Polyaxon.

Troubleshooting

The GPU resource is missing

Confirm the node uses the intended accelerated AMI and that exactly one supported NVIDIA DRA driver or device plugin is active for that capacity model.

EFA is unavailable

Verify the instance type, attached interfaces, security group, placement, RDMA libraries, and the EFA DRA driver or device plugin used by the pool.

The connected compute cluster cannot submit workloads

Verify the Polyaxon Agent deployment, cluster registration, namespace access, pod identity, queue routing, and control-plane connectivity.

References

• Amazon EKS AI/ML compute — Operational ownership differences between EKS Auto Mode and self-managed Karpenter.
• NVIDIA CUDA container — Published CUDA container tags used for the GPU visibility check.
• NVIDIA devices on Amazon EKS — Accelerated AMIs, DRA, device-plugin choices, GPU resources, and topology considerations.
• AWS Elastic Fabric Adapter — Supported instance, networking, libfabric, and device guidance.
• Polyaxon job runtime — Job containers, resources, connections, initialization, and lifecycle settings.
• Polyaxon resource scheduling — Kubernetes requests, limits, custom resources, queues, and placement controls.

---