Platform For AI:PAI DLC job — Distributed training

Parameter

Type

Default

Description

--name

string

—

Job name Required

--command

string

—

Command to run on each node Required

--image

string

—

Container image (applied to all roles that don't specify their own image)

--job-type

string

PyTorchJob

Framework type PyTorchJob TFJob MPIJob XGBoostJob ElasticBatchJob RayJob FlinkBatchJob DataJuicerJob

--workspace-id

string

config

Workspace ID

Pay-as-you-go (ECS instance type)

--worker-spec

string

—

Worker ECS instance type (e.g., ecs.gn6v-c8g1.2xlarge)

--worker-count

int

1

Number of worker replicas

Subscription (resource quota)

--resource-id

string

—

Resource quota or resource group ID

--cpu

string

—

Number of CPU cores (e.g., 8)

--memory

string

—

Memory size (e.g., 64Gi)

--gpu

string

—

Number of GPUs (e.g., 1)

--gpu-type

string

—

GPU model (e.g., A100)

--shared-memory

string

—

Shared memory size (e.g., 16Gi)

Multi-role topology (mutually exclusive with --worker-*)

--role

stringArray

—

Role definition in NAME:key=value,... format. Repeatable.

Valid NAME values: chief / evaluator / graphlearn / head / master / ps / worker. Keys: count / cpu / gpu / gpu-type / memory / shared-memory / spec / image / restart-policy / spot / spot-strategy / spot-price-limit / spot-discount-limit / oversold-type / chief

--role-json

stringArray

—

Role extension fields in NAME:{PascalCase JSON} format. Repeatable. Use for fields not covered by --role, such as SystemDisk and ExtraPodSpec.

Data and code

--data-source

stringArray

—

Data source configuration. Repeatable. Shorthand: id[:mount_path[:version]]; key-value: uri=oss://bucket/path/, mount-path=/mnt/data, enable-cache=true

--code-source

string

—

Code source ID

--code-branch

string

—

Code branch (requires --code-source)

--code-commit

string

—

Code commit hash (requires --code-source)

--code-mount-path

string

—

Mount path for code

--thirdparty-lib

stringArray

—

Pre-installed Python packages (e.g., numpy==1.26.4). Repeatable.

--thirdparty-lib-dir

string

—

Directory containing requirements.txt

Scheduling and run control

--priority

int32

0

Scheduling priority, 1–9 (0 uses the server-side default)

--scheduling-strategy

string

—

Scheduling strategy IntelligentScheduling CentralizedScheduling DistributedScheduling

--success-policy

string

—

Distributed success policy ChiefWorker AllWorkers

--max-running-time-minutes

int

0

Maximum running time in minutes (0 = unlimited)

--allow-nodes

string

—

Allowed scheduling nodes (comma-separated node names)

--deny-nodes

string

—

Excluded nodes (comma-separated node names)

--env

stringArray

—

Environment variables in KEY=VALUE format. Repeatable.

--accessibility

string

—

Visibility PRIVATE PUBLIC

--description

string

—

Job description

--image-username

string

—

Username for private image registry

--image-password

string

—

Password for private image registry

--credential-config

string

—

Credential configuration JSON

--elastic-spec

string

—

Elastic scaling configuration JSON

--settings

string

—

Advanced job settings JSON

--user-vpc

string

—

VPC network configuration JSON

Wait and debug

--wait

bool

false

Block after submit until the job reaches a terminal state

--wait-for

string

Succeeded

Target status for --wait

--wait-interval

int

10

Polling interval in seconds

--wait-timeout

int

3600

Maximum wait time in seconds

--dry-run

bool

false

Print the request body without submitting

--generate-skeleton

bool

false

Print a request template and exit. Defaults to JSON; add -o yaml for a YAML template with per-field comments that map to CLI flags. Both formats are accepted by --body-json.

--body-json

string

—

Full CreateJob request body (PascalCase field names). Accepts inline JSON, @job.json / @job.yaml to read from file, and @- to read from stdin. Scalar flags (--name, --command, etc.) override matching fields; topology and list flags (--role, --worker-*, --env, --data-source, etc.) are mutually exclusive with this flag.

--from-file

string

—

Read the job from a YAML or JSON file. Field names correspond to CLI flags without leading dashes (name, command, worker-count…). CLI flags override file values, which is useful for saving a job template and tweaking one field per run. Use - for stdin. Mutually exclusive with --body-json.

--batch

string

—

Batch-submit all job files (.json/.yaml/*.yml, same format as --from-file, one job per file) in a directory. All files are validated before submission. Requires --yes, or preview first with --dry-run. Already-created jobs aren't rolled back on failure (their IDs are printed to stdout).

--override / --override-string / --override-unset

stringArray

—

Modify request body fields before submission. See --override request body editing

Parameter

Type

Default

Description

--workspace-id

string

config

Workspace ID

--all-workspaces

bool

false

Query across all accessible workspaces

--status

string

—

Filter by status (Running / Succeeded / Failed, etc.)

--display-name

string

—

Filter by job name

--job-type

string

—

Filter by framework type (PyTorchJob / TFJob, etc.)

--show-own

bool

false

Show only jobs submitted by the current user

--since

string

7d

Time window (e.g., 7d, 48h, 365d). Jobs outside this window aren't returned

--sort-by

string

—

Sort field (e.g., GmtCreateTime)

--order

string

—

Sort direction asc desc

--page

int

1

Page number

--page-size

int

20

Items per page

--columns

string

—

Custom output columns. Default: job_id, display_name, job_type, status, gmt_create_time, duration

Parameter

Type

Default

Description

--brief

bool

false

Compact output — omit Pods / Specs / History details

Parameter

Type

Default

Description

-y, --yes

bool

false

Skip the confirmation prompt

--dry-run

bool

false

Preview the operation without executing it

Parameter

Type

Default

Description

--pod

string

Auto-select primary pod

Pod ID. Omit to read from the job's primary pod

--pod-uid

string

—

Pod UID, used to distinguish same-named pods after a restart

--all-pods

bool

false

Read all pods at once, prefixing each log line with its source pod. --tail applies per-pod — use a small value for multi-node jobs. Mutually exclusive with --pod and --follow

--follow

bool

false

Stream logs continuously (poll until interrupted)

--since

string

7 days ago

Show only logs after the specified time (e.g., 30m, 2h, 3d). For jobs that ended more than 7 days ago, widen this value explicitly to avoid InvalidTimeRange errors

--tail

int

2000

Show only the last N lines (counted from newest)

Parameter

Type

Default

Description

--pod

string

—

Read events for this pod (uses a different API). Omit to read the job's own events

--all-pods

bool

false

Read events for all pods at once, prefixing each line with its source pod. --max applies per-pod — use a small value for multi-node jobs

--pod-uid

string

—

Pod UID, used to distinguish same-named pods after a restart (requires --pod)

--max

int

2000

Maximum number of events to return (API MaxEventsNum)

--since

string

7d

Show only events after the specified time (e.g., 30m, 2h, 3d). Widen this for long-running jobs

Parameter

Type

Default

Description

--pod

string

Auto-select primary pod

Target Pod ID. Omit to auto-select using the same logic as job logs. The selection is printed to stderr; stdout carries the command's own output

--pod-uid

string

—

Pod UID, used to reject stale targets after a pod restart

--timeout

int

120

Execution timeout in seconds

Parameter

Type

Default

Description

--name

string

Original name

New job name, useful for associating it with the original

--command / --priority / --job-type and other scalar parameters

—

Original values

Replace the corresponding fields before resubmitting

--export

bool

false

Print the rebuilt request body without submitting. Edit it, then submit with job submit --body-json @file

--dry-run

bool

false

Preview the rebuilt request body without sending

Parameter

Type

Default

Description

--for

string

Succeeded

Target status Creating Queuing Dequeued Running Succeeded Failed Stopped Stopping Restarting WaitingConfig Deleted

--interval

int

10

Polling interval in seconds

--timeout

int

3600

Maximum wait time in seconds