Create BYOC Node Group

Background

Bring Your Own Compute (BYOC) allows you to bring your own compute resources to DGX Cloud Lepton. This is useful if you have existing resources you want to use with DGX Cloud Lepton.

Node groups let you group multiple nodes for a given workload. There are two types of node groups:

• BYOC Node Group: Uses your own compute resources.
• Lepton Managed Node Group: All resources are managed by DGX Cloud Lepton.

This document focuses on how to create and manage BYOC node groups.

Create a BYOC Node Group

Navigate to the Node Groups page, click Create Node Group, and select BYOC.

• Node Group Name: The name of the node group.
• Used By: List of workspaces that can allocate resources from this node group. The current workspace is pre-selected and cannot be removed.
• Visible To: List of workspaces that can view the node group.
• Accelerator Type: Defaults to the automatically detected accelerator type. You can also manually select the accelerator type from the preset options. The accelerator type name should identically match the GPU name shown when nvidia-smi is run on the GPU node.
• InfiniBand: Whether to enable InfiniBand. Defaults to disabled.
• Advanced Configuration:
  • Job Default Envs: Environment variables to be set for all jobs in the node group. You can add multiple environment variables, including variable env and secret env.
  • Private Net Interface: A list of network interfaces that are used for host LAN network traffic.
  • Host Network: Whether to enable host network for the node group. Defaults to disabled. Enable host network to allow sharing the node's network namespace, which is useful for workloads requiring direct access to the host's network interfaces.
  • Privileged Mode: Whether to enable privileged mode for the node group. Defaults to disabled. Privileged mode allows the node group to run privileged pods. This is useful for workloads that require access to the host's resources, such as the host's filesystem.
  • Expected GPUd Check: Configurations for the expected GPUd check. You can set expected values so that GPUd can help check if the node group is healthy, including InfiniBand Count, Minimum InfiniBand Bandwidth, and GPU Count.