JobSpec-阿里云帮助中心

Parameter	Type	Description	Example
	object	The detailed configuration of a worker type, including hardware specifications, runtime image, and number of replicas.
AssignNodeSpec	AssignNodeSpec	The specified node scheduling configuration.
AutoScalingSpec	AutoScalingSpec	The auto scaling configuration.
ConsiderInSuccessPolicy	boolean	Specifies whether to consider this role when determining job success. This parameter takes effect only when the success policy is set to Partial.
Driver	string	The GPU driver version used by the job.	535.54.03
EcsSpec	string	The hardware specifications of the worker. For a detailed list of specifications, see PAI-DLC billing overview. Important Prices vary by specification.	ecs.c6.large
ElasticSpotSpecs	array	The list of candidate instance types for this role to participate in elastic spot bidding.
	ElasticSpotSpec	A single candidate instance type for elastic spot bidding.
ExtraPodSpec	ExtraPodSpec	The additional pod configuration.
HyperNodeSchedulingConfig	HyperNodeSchedulingConfig	The hypernode scheduling configuration for this role.
Image	string	The runtime image URL for this type of worker. You can call ListImages to obtain images provided by the PAI platform. You can also specify a public third-party image.	registry-vpc.cn-hangzhou.aliyuncs.com/cloud-dsw/tensorflow:1.12PAI-gpu-py36-cu101-ubuntu18.04
ImageConfig	ImageConfig	The private image configuration.
IsCheif	boolean	[Deprecated] This field is misspelled and deprecated.
IsChief	boolean	Specifies whether this role is the chief role. Only one chief role is allowed.
LocalMountSpecs	array	The list of local mount configurations.
	LocalMountSpec	The local mount configuration.
OversoldType	string	The usage type of idle resources.	AcceptQuotaOverSold
PodCount	integer	The number of replicas.	1
QuotaId	string	The ID of the resource quota used by the job.	quota12345
ResourceConfig	ResourceConfig	The resource configuration.
RestartPolicy	string	The restart policy. Valid values: Always, Never, OnFailure, and ExitCode.	ExitCode
ServiceSpec	ServiceSpec	The service configuration.
SpotSpec	SpotSpec	The spot instance configuration.
StartupDependencies	array	The dependencies required before this role starts.
	StartupDependency	The startup dependency condition for a single job role.
SystemDisk	SystemDisk	The system cloud disk configuration for the job replicas.
Type	string	The type is closely related to the job type. Different job types support different worker types. TFJob: Supports Chief, PS, Worker, Evaluator, and GraphLearn. PyTorchJob: Supports Worker and Master. XGBoostJob: Supports Worker and Master. OneFlowJob: Supports Worker and Master. ElasticBatch: Supports Worker and Master. RayJob: Supports Head, Worker, and Worker[-xxx]. The Master role in PyTorchJob, XGBoostJob, OneFlowJob, and ElasticBatch is optional. If not specified, the system automatically assigns the first worker node as the master.	Worker
UseSpotInstance	boolean	Specifies whether to use spot instances.	false
UserCommand	string	The role-level startup command.	python train.py

object

The detailed configuration of a worker type, including hardware specifications, runtime image, and number of replicas.

AssignNodeSpec

The specified node scheduling configuration.

AutoScalingSpec

The auto scaling configuration.

ConsiderInSuccessPolicy

boolean

Specifies whether to consider this role when determining job success. This parameter takes effect only when the success policy is set to Partial.

Driver

string

The GPU driver version used by the job.

535.54.03

EcsSpec

string

The hardware specifications of the worker. For a detailed list of specifications, see PAI-DLC billing overview.

Important Prices vary by specification.

ecs.c6.large

ElasticSpotSpecs

array

The list of candidate instance types for this role to participate in elastic spot bidding.

ElasticSpotSpec

A single candidate instance type for elastic spot bidding.

ExtraPodSpec

The additional pod configuration.

HyperNodeSchedulingConfig

The hypernode scheduling configuration for this role.

Image

string

The runtime image URL for this type of worker. You can call ListImages to obtain images provided by the PAI platform. You can also specify a public third-party image.

registry-vpc.cn-hangzhou.aliyuncs.com/cloud-dsw/tensorflow:1.12PAI-gpu-py36-cu101-ubuntu18.04

ImageConfig

The private image configuration.

IsCheif

boolean

[Deprecated] This field is misspelled and deprecated.

IsChief

boolean

Specifies whether this role is the chief role. Only one chief role is allowed.

LocalMountSpecs

array

The list of local mount configurations.

LocalMountSpec

The local mount configuration.

OversoldType

string

The usage type of idle resources.

AcceptQuotaOverSold

PodCount

integer

The number of replicas.

1

QuotaId

string

The ID of the resource quota used by the job.

quota12345

ResourceConfig

The resource configuration.

RestartPolicy

string

The restart policy. Valid values: Always, Never, OnFailure, and ExitCode.

ExitCode

ServiceSpec

The service configuration.

SpotSpec

The spot instance configuration.

StartupDependencies

array

The dependencies required before this role starts.

StartupDependency

The startup dependency condition for a single job role.

SystemDisk

The system cloud disk configuration for the job replicas.

Type

string

The type is closely related to the job type. Different job types support different worker types.

TFJob: Supports Chief, PS, Worker, Evaluator, and GraphLearn.
PyTorchJob: Supports Worker and Master.
XGBoostJob: Supports Worker and Master.
OneFlowJob: Supports Worker and Master.
ElasticBatch: Supports Worker and Master.
RayJob: Supports Head, Worker, and Worker[-xxx].

The Master role in PyTorchJob, XGBoostJob, OneFlowJob, and ElasticBatch is optional. If not specified, the system automatically assigns the first worker node as the master.

Worker

UseSpotInstance

boolean

Specifies whether to use spot instances.

false

UserCommand

string

The role-level startup command.

python train.py