Configure a training runtime
A training runtime is a pre-built environment that combines a deep learning framework, a container image, and a parallelism strategy. Runtimes are prerequisites for submitting training jobs.
Create and manage training runtimes that define the framework, image, and parallelism strategy for training jobs.
View runtimes
Open Workbench
Click Workbench in the left sidebar.
Navigate to Runtimes
Navigate to Train > Training > Runtimes.
Browse the list
The list shows each runtime with its name, framework, description, base image, and status.
Create a custom runtime
Open the creation form
On the Runtimes page, click Create Runtime.
Configure the runtime
Set the following fields:
| Field | Required | What it does |
|---|---|---|
| Runtime name | required | A short identifier, for example pytorch-distributed. |
| Framework | required | The ML framework: torch, deepspeed, mlx, huggingface, or custom. Determines which ML policy fields are shown. |
| Base image | required | The container image URI, for example docker.io/pytorch/pytorch:2.1.0-cuda11.8-cudnn8-runtime. |
| Number of nodes | optional | Number of nodes for multi-node training. Shown when a framework is selected. |
| Processes per node | optional | Enter auto or a positive integer. Available for torch, deepspeed, and huggingface frameworks. |
| NCCL config | optional | DeepSpeed only. KEY=VALUE pairs for NCCL configuration. |
| Compute pool | optional | Optional compute pool to constrain where the runtime runs. |
| Description | optional | Short description of this runtime. |
Create
Click Create runtime.
Success looks like this: the runtime environment is configured, and training jobs can use the specified dependencies and framework versions.
Edit a runtime
Open the detail page
Click the runtime name to open the detail page.
Edit and save
Click Edit and modify the image, policy, or other settings. Click Save.
Set a default runtime
On the Runtimes page, click the Set as default action on a runtime row. The default runtime is pre-selected when creating new training jobs.
Runtimes are workspace-scoped. Administrators can publish runtimes that are available to all users in the workspace.