KEP-2170: Add TrainJob and TrainingRuntime APIs #2223
There are no files selected for viewing
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,39 @@ | ||
| /* | ||
| Copyright 2024 The Kubeflow Authors. | ||
|
|
||
| Licensed under the Apache License, Version 2.0 (the "License"); | ||
| you may not use this file except in compliance with the License. | ||
| You may obtain a copy of the License at | ||
|
|
||
| http://www.apache.org/licenses/LICENSE-2.0 | ||
|
|
||
| Unless required by applicable law or agreed to in writing, software | ||
| distributed under the License is distributed on an "AS IS" BASIS, | ||
| WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| See the License for the specific language governing permissions and | ||
| limitations under the License. | ||
| */ | ||
|
|
||
| // Package v2alpha1 contains API Schema definitions for the kubeflow.org v2alpha1 API group | ||
| // +kubebuilder:object:generate=true | ||
| // +groupName=kubeflow.org | ||
| package v2alpha1 | ||
|
|
||
| import ( | ||
| "k8s.io/apimachinery/pkg/runtime/schema" | ||
| "sigs.k8s.io/controller-runtime/pkg/scheme" | ||
| ) | ||
|
|
||
| var ( | ||
| // GroupVersion is group version used to register these objects. | ||
| GroupVersion = schema.GroupVersion{Group: "kubeflow.org", Version: "v2alpha1"} | ||
|
|
||
| // SchemeBuilder is used to add go types to the GroupVersionKind scheme. | ||
| SchemeBuilder = &scheme.Builder{GroupVersion: GroupVersion} | ||
|
|
||
| // SchemeGroupVersion is alias to GroupVersion for client-go libraries. | ||
| SchemeGroupVersion = GroupVersion | ||
|
|
||
| // AddToScheme adds the types in this group-version to the given scheme. | ||
| AddToScheme = SchemeBuilder.AddToScheme | ||
| ) |
| Original file line number | Diff line number | Diff line change | ||||
|---|---|---|---|---|---|---|
|
|
@@ -15,3 +15,181 @@ limitations under the License. | |||||
| */ | ||||||
|
|
||||||
| package v2alpha1 | ||||||
|
|
||||||
| import ( | ||||||
| autoscalingv2 "k8s.io/api/autoscaling/v2" | ||||||
| metav1 "k8s.io/apimachinery/pkg/apis/meta/v1" | ||||||
| jobsetv1alpha2 "sigs.k8s.io/jobset/api/jobset/v1alpha2" | ||||||
| ) | ||||||
|
|
||||||
| // +kubebuilder:object:root=true | ||||||
|
|
||||||
| // ClusterTrainingRuntime represents a training runtime which can be referenced as part of | ||||||
| // `trainingRuntimeRef` API in TrainJob. This resource is a cluster-scoped and can be referenced | ||||||
| // by TrainJob that created in *any* namespace. | ||||||
| type ClusterTrainingRuntime struct { | ||||||
| metav1.TypeMeta `json:",inline"` | ||||||
|
|
||||||
| // Standard object's metadata. | ||||||
| metav1.ObjectMeta `json:"metadata,omitempty"` | ||||||
|
|
||||||
| // Specification of the desired ClusterTrainingRuntime. | ||||||
| Spec TrainingRuntimeSpec `json:"spec,omitempty"` | ||||||
| } | ||||||
|
|
||||||
| // +kubebuilder:object:root=true | ||||||
|
|
||||||
| // ClusterTrainingRuntimeList is a collection of cluster training runtimes. | ||||||
| type ClusterTrainingRuntimeList struct { | ||||||
| metav1.TypeMeta `json:",inline"` | ||||||
|
|
||||||
| // Standard list metadata. | ||||||
| metav1.ListMeta `json:"metadata,omitempty"` | ||||||
|
|
||||||
| // List of ClusterTrainingRuntimes. | ||||||
| Items []ClusterTrainingRuntime `json:"items"` | ||||||
| } | ||||||
|
|
||||||
| // +kubebuilder:object:root=true | ||||||
|
|
||||||
| // TrainingRuntime represents a training runtime which can be referenced as part of | ||||||
| // `trainingRuntimeRef` API in TrainJob. This resource is a namespaced-scoped and can be referenced | ||||||
| // by TrainJob that created in the *same* namespace as the TrainingRuntime. | ||||||
| type TrainingRuntime struct { | ||||||
| metav1.TypeMeta `json:",inline"` | ||||||
|
|
||||||
| // Standard object's metadata. | ||||||
| metav1.ObjectMeta `json:"metadata,omitempty"` | ||||||
|
|
||||||
| // Specification of the desired TrainingRuntime. | ||||||
| Spec TrainingRuntimeSpec `json:"spec"` | ||||||
|
There was a problem hiding this comment.
Suggested change
Similar to other specs? |
||||||
| } | ||||||
|
|
||||||
| // +kubebuilder:object:root=true | ||||||
|
|
||||||
| // TrainingRuntimeList is a collection of training runtimes. | ||||||
| type TrainingRuntimeList struct { | ||||||
| metav1.TypeMeta `json:",inline"` | ||||||
|
|
||||||
| // Standard list metadata. | ||||||
| metav1.ListMeta `json:"metadata,omitempty"` | ||||||
|
|
||||||
| // List of TrainingRuntimes. | ||||||
| Items []TrainingRuntime `json:"items"` | ||||||
| } | ||||||
|
|
||||||
| // TrainingRuntimeSpec represents a specification of the desired training runtime. | ||||||
| type TrainingRuntimeSpec struct { | ||||||
| // Configuration for the runtime-specific parameters, such as Torch or MPI. | ||||||
| MLSpec *MLSpec `json:"mlSpec,omitempty"` | ||||||
|
|
||||||
| // Number of training nodes. | ||||||
| // Defaults to 1. | ||||||
| NumNodes *int32 `json:"numNodes,omitempty"` | ||||||
|
|
||||||
| // JobSet configuration which will be used by TrainJob. | ||||||
| JobSetSpec *jobsetv1alpha2.JobSetSpec `json:",inline"` | ||||||
|
There was a problem hiding this comment. I think, we discussed it before, here is the example: https://github.com/kubeflow/training-operator/tree/master/docs/proposals/2170-kubeflow-training-v2#pytorch-distributed-runtime. There was a problem hiding this comment. I'm asking because I don't know what that argument actually does. I usually only see if for really small objects. Never a Spec so not sure if that means literally we are putting the object "inline" or it skips protobuf or api generation? like I've seen it for type.. There was a problem hiding this comment. We want to give user functionality to set the whole JobSet spec under Training Runtimes. There was a problem hiding this comment. type TrainingRuntimeSpec struct {
MLPolicy *MLPolicy `json:"mlPolicy,omitempty"`
JobSetSpec jonsetv2alpha2.JobSetSpec `json:"spec"`
}
type MLPolicy struct {
// Number of training nodes.
// Defaults to 1.
NumNodes *int32 `json:"numNodes,omitempty"`
MLPolicySource `json:",inline"`
}
type MLPolicySource struct {
PyTorch ...
}Maybe, we want to dedicated field for the JobSetSpec so that we can identify the JobSetSpec. |
||||||
|
|
||||||
| // Configuration for the PodGroup to enable gang-scheduling via supported plugins. | ||||||
| PodGroupSpec *PodGroupSpec `json:"podGroupSpec,omitempty"` | ||||||
|
|
||||||
| } | ||||||
|
|
||||||
| // PodGroupSpec represents a PodGroup configuration to enable gang-scheduling. | ||||||
| type PodGroupSpec struct { | ||||||
| // Plugin for the gang-scheduling. | ||||||
|
There was a problem hiding this comment. Are we going forward with a default? There was a problem hiding this comment. By default, the gang-scheduling is disabled for TrainJob, since it requires plugin to be installed (coscheduling or volcano). |
||||||
| Plugin GangSchedulerPlugin `json:"plugin"` | ||||||
|
|
||||||
| // Time threshold to schedule PodGroup for gang-scheduling. | ||||||
| ScheduleTimeoutSeconds *string `json:"scheduleTimeoutSeconds,omitempty"` | ||||||
| } | ||||||
|
|
||||||
| // GangSchedulerPlugin represents one of the supported gang-scheduling plugins. | ||||||
| type GangSchedulerPlugin string | ||||||
|
|
||||||
| const ( | ||||||
| // Volcano plugin for gang-scheduling. | ||||||
| GangSchedulerPluginVolcano GangSchedulerPlugin = "volcano" | ||||||
|
|
||||||
| // Coscheduling plugin from the Kubernetes scheduler-plugins for gang-scheduling. | ||||||
| GangSchedulerPluginCoscheduling GangSchedulerPlugin = "coscheduling" | ||||||
| ) | ||||||
|
|
||||||
| // MLSpec represents the runtime-specific configuration for various technologies. | ||||||
| // One of the following specs can be set. | ||||||
| type MLSpec struct { | ||||||
| // Configuration for the PyTorch runtime. | ||||||
| TorchSpec *TorchSpec `json:"torchSpec,omitempty"` | ||||||
|
|
||||||
| // Configuration for the MPI Runtime. | ||||||
| MPISpec *MPISpec `json:"mpiSpec,omitempty"` | ||||||
| } | ||||||
|
There was a problem hiding this comment. One concern I have is that frameworks configurations/spec change quite often. Have we considered using configmap for this so that we don't have a lot of responsibilities to maintain the compatibility, etc.? There was a problem hiding this comment. Our goal is not to add all frameworks configurations here, but only parameters that require additional orchestration such as Elastic Policy, MPISpec. For example, in the future we can add SlurmSpec or FluxSpec here as we discussed with @vsoch here: #2171 (comment). For @terrytangyuan How do you think we can use ConfigMap for those parameters ? |
||||||
|
|
||||||
| // TorchSpec represents a PyTorch runtime configuration. | ||||||
| type TorchSpec struct { | ||||||
| // Number of processes per node. | ||||||
| // This value is inserted into the `--nproc-per-node` argument of the `torchrun` CLI. | ||||||
| // Supported values: `auto`, `cpu`, `gpu`, or int value. | ||||||
|
There was a problem hiding this comment. You could probably use KubeBuilder validations for the enums here. There was a problem hiding this comment. As we discussed offline, we will add validations in the separate PRs. |
||||||
| // TODO (andreyvelich): Add kubebuilder validation. | ||||||
| // Defaults to `auto`. | ||||||
| NumProcPerNode *string `json:"numProcPerNode,omitempty"` | ||||||
|
|
||||||
| // Whether to run single-node multi-worker training. | ||||||
| // This value is inserted into the `--standalone` argument of the `torchrun` CLI. | ||||||
| // Defaults to false. | ||||||
| Standalone *bool `json:"standalone,omitempty"` | ||||||
|
|
||||||
| // Elastic policy for the PyTorch training. | ||||||
| ElasticPolicy *TorchElasticPolicy `json:"elasticPolicy,omitempty"` | ||||||
| } | ||||||
|
|
||||||
| // TorchElasticPolicy represents a configuration for the PyTorch elastic training. | ||||||
| // If this policy is set, the `.spec.numNodes` parameter must be omitted, since min and max node | ||||||
| // is used to configure the `torchrun` CLI argument: `--nnodes=minNodes:maxNodes`. | ||||||
| // Only `c10d` backend is supported for the Rendezvous communication. | ||||||
| type TorchElasticPolicy struct { | ||||||
| // How many times the training job can be restarted. | ||||||
| // This value is inserted into the `--max-restarts` argument of the `torchrun` CLI and | ||||||
| // the `.spec.failurePolicy.maxRestarts` parameter of the training Job. | ||||||
| MaxRestarts *int32 `json:"maxRestarts,omitempty"` | ||||||
|
|
||||||
| // Lower limit for the number of nodes to which training job can scale down. | ||||||
| MinNodes *int32 `json:"minNodes,omitempty"` | ||||||
|
|
||||||
| // Upper limit for the number of nodes to which training job can scale up. | ||||||
| MaxNodes *int32 `json:"maxNodes,omitempty"` | ||||||
|
|
||||||
| // Specification which are used to calculate the desired number of nodes. See the individual | ||||||
| // metric source types for more information about how each type of metric must respond. | ||||||
| // The HPA will be created to perform auto-scaling. | ||||||
| Metrics []autoscalingv2.MetricSpec `json:"metrics,omitempty"` | ||||||
| } | ||||||
|
|
||||||
| // MPISpec represents a MPI runtime configuration. | ||||||
| type MPISpec struct { | ||||||
| // Number of processes per node. | ||||||
| // This value is equal to the number of slots for each node in the hostfile. | ||||||
| NumProcPerNode *int32 `json:"numProcPerNode,omitempty"` | ||||||
|
|
||||||
| // Implementation name for the MPI to create the appropriate hostfile. | ||||||
| MPIImplementation *MPIImplementation `json:"mpiImplementation"` | ||||||
|
There was a problem hiding this comment. Since MPIImplementation is a const, does this need to be a pointer? With a pointer, a Also with Maybe you meant to include the There was a problem hiding this comment. Yes, I think we should remove pointer from here. @tenzen-y What do you think ? There was a problem hiding this comment. // Implementation name for the MPI to create the appropriate hostfile.
// Defaults to OpenMPI
MPIImplementation *MPIImplementation `json:"mpiImplementation,omitempty"`In most cases of optional fields, we should use the |
||||||
|
|
||||||
| // Directory where SSH keys are mounted. | ||||||
| SSHAuthMountPath *string `json:"SSHAuthMountPath,omitempty"` | ||||||
|
|
||||||
| // Whether to run training process on the launcher Job. | ||||||
| // Defaults to false. | ||||||
| RunLauncherAsNode *bool `json:"runLauncherAsNode,omitempty"` | ||||||
| } | ||||||
|
|
||||||
| // MPIImplementation represents one of the supported MPI implementations. | ||||||
| type MPIImplementation string | ||||||
|
|
||||||
| const ( | ||||||
| MPIImplementationOpenMPI MPIImplementation = "OpenMPI" | ||||||
| MPIImplementationIntel MPIImplementation = "Intel" | ||||||
| MPIImplementationMPICH MPIImplementation = "MPICH" | ||||||
| ) | ||||||
|
|
||||||
| func init() { | ||||||
| SchemeBuilder.Register(&ClusterTrainingRuntime{}, &ClusterTrainingRuntimeList{}, &TrainingRuntime{}, &TrainingRuntimeList{}) | ||||||
| } | ||||||
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Should the spec ever be empty for ClusterTrainingRUntime?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Not really, but I noticed that for all Kubernetes APIs the
specis set withomitempty: https://github.com/kubernetes/api/blob/master/apps/v1/types.go#L820.@tenzen-y @kannon92 Any specific reason why we do this ?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
IIUC, in any case, the
specfield is defined as an optional field in the Kubernetes. So, the optional TrainingRuntime spec would be better.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
TIL.