Orchestrating Training Workflows with Workflows

​

Create Workflow

​

1. Go to the creation page

1. Go to Training > Workflows.
2. Click Create and select Create with Form or Create with Code.
   • Form Creation: Suitable for dragging nodes, configuring dependencies, and entering node parameters one by one in the console.
   • Code Creation: Suitable for scenarios where you already have a DAG configuration, need to reuse templates, or want to maintain workflow definitions uniformly through JSON/YAML.

​

2. Enter Basic information

​

3. Configure DAG

• Create by Form
• Create with Code

1. In the DAG area, drag nodes from the node list on the left to the canvas. SF Job Node and SF Inference Node are currently supported.
2. Connect the nodes in workflow order, or click Auto Layout to arrange the canvas.
3. Click the node’s edit button to open Node Configuration.
4. Configure upstream dependencies on the Predecessor Node Dependencies tab.
5. Configure the basic node information on the Common Fields tab.

   Configuration Item	Description
   Node Name	The unique name of the node in the DAG configuration, corresponding to name when created in code.
   Display Name	The display name of the node on the canvas.
   Node Type	The current node type, such as sfjob.
   Service Mode	Controls whether the node runs in service mode.
   Retry Count	The number of retries after the node fails.
   Timeout (Seconds)	The maximum time the node is allowed to run. Any excess is handled as a timeout.
   Labels	Add labels to the node to distinguish its stage or purpose.
   Environment Variables	Add runtime environment variables for the current node.
   Input Parameters	Define the parameters that the current node needs to read.
   Output Parameters	Define the parameters produced by the current node, which can be referenced by downstream nodes.

6. On the Load Parameters tab, configure the job or inference workload parameters to be run by the node.
   • SF Job Node: Configurations for the image, resources, Volume, dataset, model, startup command, and more are generally consistent with those for Training. For details, see Create a Training Task.
   • SF Inference Node: Configurations for the resource pool, inference engine, model, cache, service, and monitoring are generally consistent with those for creating an LLM inference service. For details, see Create an LLM Inference Service.
7. Click Save.
8. Verify that the nodes and dependencies on the canvas meet expectations.

1. In Code Format, select JSON or YAML.
2. In the Dag editor, enter or paste the node orchestration configuration.
3. In the Steps editor, enter or paste the node execution configuration.
4. Ensure that the configuration can represent the expected nodes, dependencies, and execution conditions.

DAG Configuration StructureThe Dag configuration describes the node orchestration, dependencies, and execution conditions of a workflow.

Field	Description
tasks	List of node orchestration configurations in the workflow.
name	Node name, which must correspond to the node name in the Steps configuration.
template	The template or node type used by the node. For example, an SF job node can use sfjob.
depends	Dependencies on preceding nodes. You can reference upstream nodes and their statuses, for example, data-preprocess.Success.
when	Node execution condition. You can reference output parameters from upstream nodes, for example, {{tasks.data-preprocess.outputs.parameters.next-step}} == training-1.

DAG structure example:

tasks:
  - name: data-preprocess
    template: sfjob
  - name: train-model
    template: sfjob
    depends: data-preprocess.Success
    when: "{{tasks.data-preprocess.outputs.parameters.next-step}} == training"

Steps Configure StructureThe Steps configuration describes each node’s runtime parameters, display information, inputs and outputs, and workload configuration.

Field	Description
type	Node type. For example, an SF job node can use sfjob.
name	Node name. This must correspond to the node name in the Dag configuration.
outputs.parameters	Node output parameters. You can use valueFrom.path to read parameter values from files generated during node execution.
inputs	Node input parameters.
displayName	Display name of the node.
labels	Node labels.
retries	Number of retries after the node fails.
timeoutSeconds	Node timeout period, in seconds.
env	Node environment variables.
spec	Node workload configuration. The specific structure depends on the node type.

Steps structure example:

- type: sfjob
  name: data-preprocess
  displayName: Data preprocessing
  retries: 1
  timeoutSeconds: 3600
  outputs:
    parameters:
      - name: next-step
        valueFrom:
          path: /etc/output/data-preprocess/result.json
  env:
    - name: STAGE
      value: preprocess
  spec:
    # Configure node load parameters by node type.
    # For SF job nodes, see Create Training Tasks; for SF inference nodes, see Create an LLM Inference Service.

​

4. Submit Creation

​

View and manage workflows

• The current workflow status and execution progress.
• Which nodes are running, have failed, or are waiting for upstream nodes.
• Whether a node failure is caused by task configuration, resource scheduling, an image, mounting, a command, or upstream dependencies.
• Whether the workflow needs to be rerun or a specific training job needs to be investigated separately.

​

Related Documentation

• Create Training