gpu.md

GPU workloads

nomploy schedules NVIDIA GPUs through Nomad's nomad-device-nvidia plugin. There are two steps: enable GPU support on the node, then request a GPU in your job.

1. Enable GPU support on the node

The NVIDIA driver (nvidia-smi) must already be installed on the machine — nomploy does not install the kernel driver.

Open the server's setup/actions and use GPU Configuration → Enable GPU Support. This:

• installs the NVIDIA Container Toolkit,
• points Docker at the nvidia runtime, and
• installs the nomad-device-nvidia plugin so Nomad can fingerprint and schedule the GPUs.

The card shows the detected GPU status and becomes Reconfigure GPU once set up.

2. Request a GPU in your workload

There's no separate GPU form field — you request GPUs in your service definition.

Docker Compose

Use the standard compose device reservation:

services:
  trainer:
    image: my/cuda-app:latest
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]

A device is treated as a GPU when its driver is nvidia or its capabilities include gpu. The count is summed across matching devices; count: all or an omitted count is treated as 1. nomploy turns that into a Nomad device "nvidia/gpu" { count = N } block in the generated job. (nvidia/gpu matches any NVIDIA GPU the plugin fingerprints.)

Native Nomad HCL

If you deploy a native HCL job, request the device directly in the task's resources:

resources {
  cpu    = 2000
  memory = 4096
  device "nvidia/gpu" {
    count = 1
  }
}

This gives you the full device-plugin surface (constraints on model, memory, affinity, etc.) if you need it.

Notes

• GPUs are only schedulable on nodes where you enabled GPU support in step 1.
• Combine with a constraint if you have a mixed fleet and want a job pinned to GPU nodes.