mirror of
https://github.com/hanzo-ml/kuberay.git
synced 2026-10-30 19:48:47 +00:00
208 lines
8.3 KiB
Go
208 lines
8.3 KiB
Go
package create
|
|
|
|
import (
|
|
"context"
|
|
"fmt"
|
|
|
|
"github.com/spf13/cobra"
|
|
corev1 "k8s.io/api/core/v1"
|
|
"k8s.io/apimachinery/pkg/api/resource"
|
|
metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
|
|
"k8s.io/cli-runtime/pkg/genericclioptions"
|
|
"k8s.io/client-go/rest"
|
|
cmdutil "k8s.io/kubectl/pkg/cmd/util"
|
|
"k8s.io/kubectl/pkg/util/templates"
|
|
|
|
"github.com/ray-project/kuberay/kubectl-plugin/pkg/util"
|
|
"github.com/ray-project/kuberay/kubectl-plugin/pkg/util/client"
|
|
rayv1 "github.com/ray-project/kuberay/ray-operator/apis/ray/v1"
|
|
)
|
|
|
|
type CreateWorkerGroupOptions struct {
|
|
cmdFactory cmdutil.Factory
|
|
ioStreams *genericclioptions.IOStreams
|
|
namespace string
|
|
clusterName string
|
|
groupName string
|
|
rayStartParams map[string]string
|
|
workerNodeSelectors map[string]string
|
|
rayVersion string
|
|
image string
|
|
workerCPU string
|
|
workerGPU string
|
|
workerTPU string
|
|
workerMemory string
|
|
workerReplicas int32
|
|
numOfHosts int32
|
|
workerMinReplicas int32
|
|
workerMaxReplicas int32
|
|
}
|
|
|
|
var (
|
|
createWorkerGroupLong = templates.LongDesc(`
|
|
Adds a worker group to an existing Ray cluster.
|
|
`)
|
|
|
|
createWorkerGroupExample = templates.Examples(fmt.Sprintf(`
|
|
# Create a worker group in an existing Ray cluster with defaults
|
|
kubectl ray create workergroup example-group --ray-cluster sample-cluster
|
|
|
|
# Create a worker group in an existing Ray cluster
|
|
kubectl ray create workergroup example-group --ray-cluster sample-cluster --image %s --worker-cpu 2 --worker-memory 5Gi
|
|
|
|
# Create a worker group in an existing Ray cluster with TPU
|
|
kubectl ray create workergroup example-tpu-group --ray-cluster sample-cluster --worker-tpu 1 --worker-node-selectors %s=tpu-v5-lite-podslice,%s=1x1
|
|
|
|
# For more details on TPU-related node selectors like %s and %s, refer to:
|
|
#https://cloud.google.com/kubernetes-engine/docs/concepts/plan-tpus#availability
|
|
`, util.RayImage, util.NodeSelectorGKETPUAccelerator, util.NodeSelectorGKETPUTopology, util.NodeSelectorGKETPUAccelerator, util.NodeSelectorGKETPUTopology))
|
|
)
|
|
|
|
func NewCreateWorkerGroupOptions(cmdFactory cmdutil.Factory, streams genericclioptions.IOStreams) *CreateWorkerGroupOptions {
|
|
return &CreateWorkerGroupOptions{
|
|
cmdFactory: cmdFactory,
|
|
ioStreams: &streams,
|
|
}
|
|
}
|
|
|
|
func NewCreateWorkerGroupCommand(cmdFactory cmdutil.Factory, streams genericclioptions.IOStreams) *cobra.Command {
|
|
options := NewCreateWorkerGroupOptions(cmdFactory, streams)
|
|
// Silence warnings to avoid messages like 'unknown field "spec.headGroupSpec.template.metadata.creationTimestamp"'
|
|
// See https://github.com/kubernetes/kubernetes/issues/67610 for more details.
|
|
rest.SetDefaultWarningHandler(rest.NoWarnings{})
|
|
|
|
cmd := &cobra.Command{
|
|
Use: "workergroup (WORKERGROUP) [(-r|--ray-cluster) RAYCLUSTER]",
|
|
Short: "Create worker group in an existing Ray cluster",
|
|
Long: createWorkerGroupLong,
|
|
Example: createWorkerGroupExample,
|
|
SilenceUsage: true,
|
|
RunE: func(cmd *cobra.Command, args []string) error {
|
|
if err := options.Validate(); err != nil {
|
|
return err
|
|
}
|
|
|
|
if err := options.Complete(cmd, args); err != nil {
|
|
return err
|
|
}
|
|
return options.Run(cmd.Context(), cmdFactory)
|
|
},
|
|
}
|
|
|
|
cmd.Flags().StringVarP(&options.clusterName, "ray-cluster", "c", "", "Ray cluster to add a worker group to")
|
|
cobra.CheckErr(cmd.MarkFlagRequired("ray-cluster"))
|
|
cmd.Flags().StringVar(&options.rayVersion, "ray-version", util.RayVersion, "Ray version to use")
|
|
cmd.Flags().StringVar(&options.image, "image", fmt.Sprintf("rayproject/ray:%s", options.rayVersion), "container image to use")
|
|
cmd.Flags().Int32Var(&options.workerReplicas, "worker-replicas", 1, "desired replicas")
|
|
cmd.Flags().Int32Var(&options.numOfHosts, "num-of-hosts", 1, "number of hosts in the worker group per replica")
|
|
cmd.Flags().Int32Var(&options.workerMinReplicas, "worker-min-replicas", 1, "minimum number of replicas")
|
|
cmd.Flags().Int32Var(&options.workerMaxReplicas, "worker-max-replicas", 10, "maximum number of replicas")
|
|
cmd.Flags().StringVar(&options.workerCPU, "worker-cpu", "2", "number of CPUs in each replica")
|
|
cmd.Flags().StringVar(&options.workerGPU, "worker-gpu", "0", "number of GPUs in each replica")
|
|
cmd.Flags().StringVar(&options.workerTPU, "worker-tpu", "0", "number of TPUs in each replica")
|
|
cmd.Flags().StringVar(&options.workerMemory, "worker-memory", "4Gi", "amount of memory in each replica")
|
|
cmd.Flags().StringToStringVar(&options.rayStartParams, "worker-ray-start-params", make(map[string]string), "a map of arguments to the Ray workers' 'ray start' entrypoint, e.g. '--worker-ray-start-params metrics-export-port=8080,num-cpus=2'")
|
|
cmd.Flags().StringToStringVar(&options.workerNodeSelectors, "worker-node-selectors", make(map[string]string), "Node selectors to apply to all worker pods in this worker group (e.g. --worker-node-selectors cloud.google.com/gke-accelerator=nvidia-l4,cloud.google.com/gke-nodepool=my-node-pool)")
|
|
|
|
return cmd
|
|
}
|
|
|
|
func (options *CreateWorkerGroupOptions) Complete(cmd *cobra.Command, args []string) error {
|
|
namespace, _, err := options.cmdFactory.ToRawKubeConfigLoader().Namespace()
|
|
if err != nil {
|
|
return fmt.Errorf("failed to get namespace: %w", err)
|
|
}
|
|
options.namespace = namespace
|
|
|
|
if options.rayStartParams == nil {
|
|
options.rayStartParams = map[string]string{}
|
|
}
|
|
|
|
if len(args) != 1 {
|
|
return cmdutil.UsageErrorf(cmd, "%s", cmd.Use)
|
|
}
|
|
options.groupName = args[0]
|
|
|
|
if options.image == "" {
|
|
options.image = fmt.Sprintf("rayproject/ray:%s", options.rayVersion)
|
|
}
|
|
|
|
return nil
|
|
}
|
|
|
|
func (options *CreateWorkerGroupOptions) Validate() error {
|
|
if err := util.ValidateTPU(&options.workerTPU, &options.numOfHosts, options.workerNodeSelectors); err != nil {
|
|
return fmt.Errorf("%w", err)
|
|
}
|
|
return nil
|
|
}
|
|
|
|
func (options *CreateWorkerGroupOptions) Run(ctx context.Context, factory cmdutil.Factory) error {
|
|
k8sClient, err := client.NewClient(factory)
|
|
if err != nil {
|
|
return fmt.Errorf("failed to create client: %w", err)
|
|
}
|
|
|
|
rayCluster, err := k8sClient.RayClient().RayV1().RayClusters(options.namespace).Get(ctx, options.clusterName, metav1.GetOptions{})
|
|
if err != nil {
|
|
return fmt.Errorf("error getting Ray cluster: %w", err)
|
|
}
|
|
|
|
newRayCluster := rayCluster.DeepCopy()
|
|
|
|
newRayCluster.Spec.WorkerGroupSpecs = append(newRayCluster.Spec.WorkerGroupSpecs, createWorkerGroupSpec(options))
|
|
|
|
newRayCluster, err = k8sClient.RayClient().RayV1().RayClusters(options.namespace).Update(ctx, newRayCluster, metav1.UpdateOptions{FieldManager: util.FieldManager})
|
|
if err != nil {
|
|
return fmt.Errorf("error updating Ray cluster with new worker group: %w", err)
|
|
}
|
|
|
|
fmt.Printf("Updated Ray cluster %s/%s with new worker group\n", newRayCluster.Namespace, newRayCluster.Name)
|
|
return nil
|
|
}
|
|
|
|
func createWorkerGroupSpec(options *CreateWorkerGroupOptions) rayv1.WorkerGroupSpec {
|
|
podTemplate := corev1.PodTemplateSpec{
|
|
Spec: corev1.PodSpec{
|
|
Containers: []corev1.Container{
|
|
{
|
|
Name: "ray-worker",
|
|
Image: options.image,
|
|
Resources: corev1.ResourceRequirements{
|
|
Requests: corev1.ResourceList{
|
|
corev1.ResourceCPU: resource.MustParse(options.workerCPU),
|
|
corev1.ResourceMemory: resource.MustParse(options.workerMemory),
|
|
},
|
|
Limits: corev1.ResourceList{
|
|
corev1.ResourceCPU: resource.MustParse(options.workerCPU),
|
|
corev1.ResourceMemory: resource.MustParse(options.workerMemory),
|
|
},
|
|
},
|
|
},
|
|
},
|
|
NodeSelector: options.workerNodeSelectors,
|
|
},
|
|
}
|
|
|
|
gpuResource := resource.MustParse(options.workerGPU)
|
|
if !gpuResource.IsZero() {
|
|
podTemplate.Spec.Containers[0].Resources.Requests[corev1.ResourceName(util.ResourceNvidiaGPU)] = gpuResource
|
|
podTemplate.Spec.Containers[0].Resources.Limits[corev1.ResourceName(util.ResourceNvidiaGPU)] = gpuResource
|
|
}
|
|
|
|
tpuResource := resource.MustParse(options.workerTPU)
|
|
if !tpuResource.IsZero() {
|
|
podTemplate.Spec.Containers[0].Resources.Requests[corev1.ResourceName(util.ResourceGoogleTPU)] = tpuResource
|
|
podTemplate.Spec.Containers[0].Resources.Limits[corev1.ResourceName(util.ResourceGoogleTPU)] = tpuResource
|
|
}
|
|
return rayv1.WorkerGroupSpec{
|
|
GroupName: options.groupName,
|
|
Replicas: &options.workerReplicas,
|
|
NumOfHosts: options.numOfHosts,
|
|
MinReplicas: &options.workerMinReplicas,
|
|
MaxReplicas: &options.workerMaxReplicas,
|
|
RayStartParams: options.rayStartParams,
|
|
Template: podTemplate,
|
|
}
|
|
}
|