pcm-coordinator/internal/storeLink/modelarts.go

/*

 Copyright (c) [2023] [pcm]
 [pcm-coordinator] is licensed under Mulan PSL v2.
 You can use this software according to the terms and conditions of the Mulan PSL v2.
 You may obtain a copy of Mulan PSL v2 at:
          http://license.coscl.org.cn/MulanPSL2
 THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
 EITHER EXPaRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
 MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
 See the Mulan PSL v2 for more details.

*/

package storeLink

import (
	"context"
	"github.com/pkg/errors"
	"gitlink.org.cn/JointCloud/pcm-coordinator/internal/scheduler/schedulers/option"
	"gitlink.org.cn/JointCloud/pcm-coordinator/internal/scheduler/service/collector"
	"gitlink.org.cn/JointCloud/pcm-coordinator/internal/scheduler/service/inference"
	"gitlink.org.cn/JointCloud/pcm-coordinator/pkg/constants"
	"gitlink.org.cn/JointCloud/pcm-coordinator/pkg/utils"
	"gitlink.org.cn/JointCloud/pcm-coordinator/pkg/utils/timeutils"
	"gitlink.org.cn/JointCloud/pcm-modelarts/client/imagesservice"
	"gitlink.org.cn/JointCloud/pcm-modelarts/client/modelartsservice"
	"gitlink.org.cn/JointCloud/pcm-modelarts/modelarts"
	modelartsclient "gitlink.org.cn/JointCloud/pcm-modelarts/modelarts"
	"mime/multipart"
	"strconv"
	"strings"
	"time"
)

const (
	Ascend = "Ascend"
	Npu    = "npu"
)

type ModelArtsLink struct {
	modelArtsRpc    modelartsservice.ModelArtsService
	modelArtsImgRpc imagesservice.ImagesService
	platform        string
	participantId   int64
	pageIndex       int32
	pageSize        int32
}

func NewModelArtsLink(modelArtsRpc modelartsservice.ModelArtsService, modelArtsImgRpc imagesservice.ImagesService, name string, id int64, nickname string) *ModelArtsLink {
	return &ModelArtsLink{modelArtsRpc: modelArtsRpc, modelArtsImgRpc: modelArtsImgRpc, platform: nickname, participantId: id, pageIndex: 0, pageSize: 50}
}

func (m *ModelArtsLink) UploadImage(ctx context.Context, path string) (interface{}, error) {
	//TODO modelArts上传镜像
	return nil, nil
}

func (m *ModelArtsLink) DeleteImage(ctx context.Context, imageId string) (interface{}, error) {
	// TODO modelArts删除镜像
	return nil, nil
}

func (m *ModelArtsLink) QueryImageList(ctx context.Context) (interface{}, error) {
	// modelArts获取镜像列表
	req := &modelarts.ListRepoReq{
		Offset:   "0",
		Limit:    strconv.Itoa(int(m.pageSize)),
		Platform: m.platform,
	}
	resp, err := m.modelArtsImgRpc.ListReposDetails(ctx, req)
	if err != nil {
		return nil, err
	}

	return resp, nil
}

func (m *ModelArtsLink) SubmitTask(ctx context.Context, imageId string, cmd string, envs []string, params []string, resourceId string, datasetsId string, algorithmId string, aiType string) (interface{}, error) {
	// modelArts提交任务
	environments := make(map[string]string)
	parameters := make([]*modelarts.ParametersTrainJob, 0)
	for _, env := range envs {
		s := strings.Split(env, COMMA)
		environments[s[0]] = s[1]
	}
	for _, param := range params {
		s := strings.Split(param, COMMA)
		parameters = append(parameters, &modelarts.ParametersTrainJob{
			Name:  s[0],
			Value: s[1],
		})
	}
	req := &modelarts.CreateTrainingJobReq{
		Kind: "job",
		Metadata: &modelarts.MetadataS{
			Name:        TASK_NAME_PREFIX + utils.RandomString(10),
			WorkspaceId: "0",
		},
		Algorithm: &modelarts.Algorithms{
			Id: algorithmId,
			Engine: &modelarts.EngineCreateTraining{
				ImageUrl: imageId,
			},
			Command:      cmd,
			Environments: environments,
			Parameters:   parameters,
		},
		Spec: &modelarts.SpecsC{
			Resource: &modelarts.ResourceCreateTraining{
				FlavorId:  resourceId,
				NodeCount: 1,
			},
		},
		Platform: m.platform,
	}
	resp, err := m.modelArtsRpc.CreateTrainingJob(ctx, req)
	if err != nil {
		return nil, err
	}

	return resp, nil
}

func (m *ModelArtsLink) QueryTask(ctx context.Context, taskId string) (interface{}, error) {
	// 获取任务
	req := &modelarts.DetailTrainingJobsReq{
		TrainingJobId: taskId,
		Platform:      m.platform,
	}
	resp, err := m.modelArtsRpc.GetTrainingJobs(ctx, req)
	if err != nil {
		return nil, err
	}

	return resp, nil
}

func (m *ModelArtsLink) DeleteTask(ctx context.Context, taskId string) (interface{}, error) {
	// 删除任务
	req := &modelarts.DeleteTrainingJobReq{
		TrainingJobId: taskId,
		Platform:      m.platform,
	}
	resp, err := m.modelArtsRpc.DeleteTrainingJob(ctx, req)
	if err != nil {
		return nil, err
	}

	return resp, nil
}

func (m *ModelArtsLink) QuerySpecs(ctx context.Context) (interface{}, error) {
	// octopus查询资源规格
	req := &modelarts.TrainingJobFlavorsReq{
		Platform: m.platform,
	}
	resp, err := m.modelArtsRpc.GetTrainingJobFlavors(ctx, req)
	if err != nil {
		return nil, err
	}

	return resp, nil
}

func (m *ModelArtsLink) GetResourceStats(ctx context.Context) (*collector.ResourceStats, error) {
	req := &modelarts.GetPoolsRuntimeMetricsReq{}
	resp, err := m.modelArtsRpc.GetPoolsRuntimeMetrics(ctx, req)
	if err != nil {
		return nil, err
	}
	if resp.ErrorMsg != "" {
		return nil, errors.New("failed to get algorithms")
	}
	resourceStats := &collector.ResourceStats{}
	CpuCoreTotalSum := int64(0)
	CpuCoreAvailSum := int64(0)
	MemTotalSum := float64(0)
	MemAvailSum := float64(0)
	var CpuCoreTotal int64
	var CpuCoreAvail int64
	var MemTotal float64
	var MemAvail float64
	for _, items := range resp.Items {
		//TODO The value of taskType is temporarily fixed to "pytorch"
		CpuCoreTotal, err = strconv.ParseInt(items.Table.Capacity.Value.Cpu, 10, 64)
		CpuCoreTotalSum += CpuCoreTotal
		CpuCoreAvail, err = strconv.ParseInt(items.Table.Allocated.Value.Cpu, 10, 64)
		CpuCoreAvailSum += CpuCoreAvail
		MemTotal, err = strconv.ParseFloat(items.Table.Capacity.Value.Memory, 64)
		MemTotalSum += MemTotal
		MemAvail, err = strconv.ParseFloat(items.Table.Allocated.Value.Memory, 64)
		MemAvailSum += MemAvail
	}
	resourceStats.CpuCoreTotal = CpuCoreTotalSum
	resourceStats.CpuCoreAvail = CpuCoreAvailSum
	resourceStats.MemTotal = MemTotalSum
	resourceStats.MemAvail = MemAvailSum
	req1 := &modelarts.GetResourceFlavorsReq{}
	resp1, err := m.modelArtsRpc.GetResourceFlavors(ctx, req1)

	num32, _ := strconv.Atoi(resp1.Items[0].Spec.Npu.Size)
	var cards []*collector.Card
	card := &collector.Card{
		Platform:   MODELARTS,
		Type:       CARD,
		Name:       Npu,
		CardNum:    int32(num32),
		TOpsAtFp16: float64(num32 * 320),
	}
	cards = append(cards, card)
	resourceStats.CardsAvail = cards

	return resourceStats, nil
}

func (m *ModelArtsLink) GetDatasetsSpecs(ctx context.Context) ([]*collector.DatasetsSpecs, error) {
	return nil, nil
}

func (m *ModelArtsLink) GetAlgorithms(ctx context.Context) ([]*collector.Algorithm, error) {
	var algorithms []*collector.Algorithm

	req := &modelarts.ListAlgorithmsReq{
		Platform: m.platform,
		Offset:   m.pageIndex,
		Limit:    m.pageSize,
	}
	resp, err := m.modelArtsRpc.ListAlgorithms(ctx, req)
	if err != nil {
		return nil, err
	}
	if resp.ErrorMsg != "" {
		return nil, errors.New("failed to get algorithms")
	}

	for _, a := range resp.Items {
		//TODO The value of taskType is temporarily fixed to "pytorch"
		algorithm := &collector.Algorithm{Name: a.Metadata.Name, Platform: MODELARTS, TaskType: "pytorch"}
		algorithms = append(algorithms, algorithm)
	}
	return algorithms, nil
}

func (m *ModelArtsLink) GetComputeCards(ctx context.Context) ([]string, error) {
	var cards []string
	cards = append(cards, Ascend)
	return cards, nil
}

func (m *ModelArtsLink) GetUserBalance(ctx context.Context) (float64, error) {
	return 0, nil
}

func (m *ModelArtsLink) DownloadAlgorithmCode(ctx context.Context, resourceType string, card string, taskType string, dataset string, algorithm string) (string, error) {
	return "", nil
}

func (m *ModelArtsLink) UploadAlgorithmCode(ctx context.Context, resourceType string, card string, taskType string, dataset string, algorithm string, code string) error {
	return nil
}

// Get AI Application List
func (m *ModelArtsLink) GetModelId(ctx context.Context, option *option.InferOption) error {
	req := &modelarts.ListModelReq{
		Platform:  m.platform,
		ModelName: option.ModelName,
		ModelType: "Image",
		Limit:     int64(m.pageIndex),
		Offset:    int64(m.pageSize),
	}
	ListModelResp, err := m.modelArtsRpc.ListModels(ctx, req)
	if err != nil {
		return err
	}
	if ListModelResp.Code != 200 {
		return errors.New("failed to get ModelId")
	}
	for _, ListModel := range ListModelResp.Models {
		if ListModel.ModelName == option.ModelName {
			option.ModelId = ListModel.ModelId
			return nil
		}
	}
	return nil
}

// 创建ai应用
func (m *ModelArtsLink) CreateModel(ctx context.Context, option *option.InferOption) error {
	req := &modelarts.CreateModelReq{
		Platform:       m.platform,
		ModelName:      option.ModelName,
		ModelType:      "PyTorch",
		ModelVersion:   "0.0.1",
		SourceLocation: "",
	}
	ModelResp, err := m.modelArtsRpc.CreateModel(ctx, req)
	if err != nil {
		return err
	}
	if ModelResp.Code != 200 {
		return errors.New("failed to get ModelId")
	}
	option.ModelId = ModelResp.ModelId

	return nil
}

func (m *ModelArtsLink) GetSpecifications(ctx context.Context, option *option.AiOption, ifoption *option.InferOption) error {
	req := &modelarts.ListSpecificationsReq{
		//Platform:  m.platform,
		IsPersonalCluster: false,
		InferType:         "real-time",
		Limit:             m.pageIndex,
		OffSet:            m.pageSize,
	}
	ListSpecificationsResp, err := m.modelArtsRpc.ListSpecifications(ctx, req)
	if err != nil {
		return err
	}
	for _, ListSpecifications := range ListSpecificationsResp.Specifications {
		if ListSpecifications.Specification == "modelarts.kat1.xlarge" {
			ifoption.Specification = ListSpecifications.Specification
			return nil
		}
	}
	return nil
}

func (m *ModelArtsLink) GetTrainingTaskLog(ctx context.Context, taskId string, instanceNum string) (string, error) {
	req := &modelartsservice.GetTrainingJobLogsPreviewReq{
		Platform:      m.platform,
		TaskId:        "worker-0",
		TrainingJobId: taskId,
	}
	resp, err := m.modelArtsRpc.GetTrainingJobLogsPreview(ctx, req)
	if err != nil {
		return "", err
	}

	if strings.Contains(resp.Content, "404 Not Found") {
		resp.Content = "waiting for logs..."
	}
	return resp.Content, nil
}

func (m *ModelArtsLink) GetTrainingTask(ctx context.Context, taskId string) (*collector.Task, error) {
	resp, err := m.QueryTask(ctx, taskId)
	if err != nil {
		return nil, err
	}
	jobresp, ok := (resp).(*modelartsservice.JobResponse)
	if jobresp.ErrorMsg != "" || !ok {
		if jobresp.ErrorMsg != "" {
			return nil, errors.New(jobresp.ErrorMsg)
		} else {
			return nil, errors.New("get training task failed, empty error returned")
		}
	}
	var task collector.Task
	task.Id = jobresp.Metadata.Id

	switch strings.ToLower(jobresp.Status.Phase) {
	case "completed":
		milliTimestamp := int64(jobresp.Status.StartTime)
		task.Start = timeutils.MillisecondsToUTCString(milliTimestamp, time.DateTime)
		duration := int64(jobresp.Status.Duration)
		task.End = timeutils.MillisecondsToAddDurationToUTCString(milliTimestamp, duration, time.DateTime)
		task.Status = constants.Completed
	case "failed":
		milliTimestamp := int64(jobresp.Status.StartTime)
		task.Start = timeutils.MillisecondsToUTCString(milliTimestamp, time.DateTime)
		duration := int64(jobresp.Status.Duration)
		task.End = timeutils.MillisecondsToAddDurationToUTCString(milliTimestamp, duration, time.DateTime)
		task.Status = constants.Failed
	case "running":
		milliTimestamp := int64(jobresp.Status.StartTime)
		task.Start = timeutils.MillisecondsToUTCString(milliTimestamp, time.DateTime)
		task.Status = constants.Running
	case "stopped":
		task.Status = constants.Stopped
	case "pending":
		task.Status = constants.Pending
	case "terminated":
		//TODO Failed
		task.Status = constants.Failed
	default:
		task.Status = "undefined"
	}

	return &task, nil
}

func (m *ModelArtsLink) Execute(ctx context.Context, option *option.AiOption) (interface{}, error) {
	err := m.GenerateSubmitParams(ctx, option)
	if err != nil {
		return nil, err
	}
	task, err := m.SubmitTask(ctx, option.ImageId, option.Cmd, option.Envs, option.Params, option.ResourceId, option.DatasetsId, option.AlgorithmId, option.TaskType)
	if err != nil {
		return nil, err
	}
	return task, nil
}

func (m *ModelArtsLink) GenerateSubmitParams(ctx context.Context, option *option.AiOption) error {
	err := m.generateResourceId(ctx, option, nil)
	if err != nil {
		return err
	}
	err = m.generateAlgorithmId(ctx, option)
	if err != nil {
		return err
	}
	err = m.generateImageId(option)
	if err != nil {
		return err
	}
	err = m.generateCmd(option)
	if err != nil {
		return err
	}
	err = m.generateEnv(option)
	if err != nil {
		return err
	}
	err = m.generateParams(option)
	if err != nil {
		return err
	}
	return nil
}

func (m *ModelArtsLink) generateResourceId(ctx context.Context, option *option.AiOption, ifoption *option.InferOption) error {
	option.ResourceId = "modelarts.kat1.xlarge"
	return nil
}

func (m *ModelArtsLink) generateImageId(option *option.AiOption) error {

	return nil
}

func (m *ModelArtsLink) generateCmd(option *option.AiOption) error {

	return nil
}

func (m *ModelArtsLink) generateEnv(option *option.AiOption) error {

	return nil
}

func (m *ModelArtsLink) generateParams(option *option.AiOption) error {

	return nil
}

func (m *ModelArtsLink) generateAlgorithmId(ctx context.Context, option *option.AiOption) error {
	req := &modelarts.ListAlgorithmsReq{
		Platform: m.platform,
		Offset:   m.pageIndex,
		Limit:    m.pageSize,
	}
	resp, err := m.modelArtsRpc.ListAlgorithms(ctx, req)
	if err != nil {
		return err
	}
	if resp.ErrorMsg != "" {
		return errors.New("failed to get algorithmId")
	}

	for _, algorithm := range resp.Items {
		engVersion := algorithm.JobConfig.Engine.EngineVersion
		if strings.Contains(engVersion, option.TaskType) {
			ns := strings.Split(algorithm.Metadata.Name, DASH)
			if ns[0] != option.TaskType {
				continue
			}
			if ns[1] != option.DatasetsName {
				continue
			}
			if ns[2] != option.AlgorithmName {
				continue
			}
			option.AlgorithmId = algorithm.Metadata.Id
			return nil
		}
	}

	if option.AlgorithmId == "" {
		return errors.New("Algorithm does not exist")
	}

	return errors.New("failed to get AlgorithmId")
}

func (m *ModelArtsLink) GetClusterInferUrl(ctx context.Context, option *option.InferOption) (*inference.ClusterInferUrl, error) {
	var imageUrls []*inference.InferUrl
	urlReq := &modelartsclient.ImageReasoningUrlReq{
		ServiceName: option.ModelName,
		Type:        option.ModelType,
		Card:        "npu",
	}
	urlResp, err := m.modelArtsRpc.ImageReasoningUrl(ctx, urlReq)
	if err != nil {
		return nil, err
	}
	imageUrl := &inference.InferUrl{
		Url:  urlResp.Url,
		Card: "npu",
	}
	imageUrls = append(imageUrls, imageUrl)

	clusterWithUrl := &inference.ClusterInferUrl{
		ClusterName: m.platform,
		ClusterType: TYPE_MODELARTS,
		InferUrls:   imageUrls,
	}
	return clusterWithUrl, nil
}

func (m *ModelArtsLink) GetInferDeployInstanceList(ctx context.Context) ([]*inference.DeployInstance, error) {
	var insList []*inference.DeployInstance
	req := &modelarts.ListServicesReq{
		Platform: m.platform,
		OffSet:   m.pageIndex,
		Limit:    m.pageSize,
	}
	//list, err := m.modelArtsRpc.ListServices(ctx, req)
	resp, err := m.modelArtsRpc.ListServices(ctx, req)
	if err != nil {
		return nil, err
	}
	if resp.ErrorMsg != "" {
		return nil, errors.New(resp.Msg)
	}

	for _, services := range resp.Services {
		ins := &inference.DeployInstance{}
		ins.InstanceName = services.ServiceName
		ins.InstanceId = services.ServiceId
		ins.Status = services.Status
		ins.InferCard = "NPU"
		ins.ClusterName = m.platform
		ins.CreatedTime = string(services.StartTime)
		ins.ClusterType = TYPE_MODELARTS
		insList = append(insList, ins)
	}

	return insList, nil
}

func (m *ModelArtsLink) StartInferDeployInstance(ctx context.Context, id string) bool {
	req := &modelartsclient.UpdateServiceReq{
		ServiceId: id,
		Status:    "running",
	}
	resp, err := m.modelArtsRpc.UpdateService(ctx, req)
	if err != nil || resp.Code != 0 {
		return false
	}
	if resp.Code == 0 {
		return true
	}
	return false
}

func (m *ModelArtsLink) StopInferDeployInstance(ctx context.Context, id string) bool {
	req := &modelartsclient.UpdateServiceReq{
		ServiceId: id,
		Status:    "stopped",
	}
	resp, err := m.modelArtsRpc.UpdateService(ctx, req)
	if err != nil || resp.Code != 0 {
		return false
	}
	if resp.Code == 0 {
		return true
	}
	return false
}

func (m *ModelArtsLink) GetInferDeployInstance(ctx context.Context, id string) (*inference.DeployInstance, error) {
	req := &modelarts.ShowServiceReq{
		ServiceId: id,
	}
	resp, err := m.modelArtsRpc.ShowService(ctx, req)
	if err != nil {
		return nil, err
	}
	if resp.ErrorMsg != "" {
		return nil, errors.New(resp.Msg)
	}
	ins := &inference.DeployInstance{}
	ins.InstanceName = resp.ServiceName
	ins.InstanceId = resp.ServiceId
	ins.Status = resp.Status
	ins.InferCard = "NPU"
	ins.ClusterName = m.platform
	ins.CreatedTime = string(resp.StartTime)
	ins.ClusterType = TYPE_MODELARTS
	return ins, nil
}

func (m *ModelArtsLink) GetInferResult(ctx context.Context, url string, file multipart.File, fileName string) (string, error) {
	return "", nil
}

func (m *ModelArtsLink) CreateInferDeployInstance(ctx context.Context, option *option.InferOption) (string, error) {

	err := m.CreateModel(ctx, option)
	if err != nil {
		return "", err
	}

	configParam := &modelarts.ServiceConfig{
		Specification: "modelarts.kat1.xlarge",
		Weight:        100,
		ModelId:       option.ModelId,
		InstanceCount: 1,
	}
	var configItems []*modelarts.ServiceConfig
	configItems = append(configItems, configParam)
	req := &modelarts.CreateServiceReq{
		Platform:    m.platform,
		Config:      configItems,
		InferType:   "real-time",
		ServiceName: option.TaskName,
	}
	resp, err := m.modelArtsRpc.CreateService(ctx, req)
	if err != nil {
		return "", err
	}

	return resp.ServiceId, nil
}

func (m *ModelArtsLink) CheckModelExistence(ctx context.Context, name string, mtype string) bool {
	ifoption := &option.InferOption{
		ModelName: name,
		ModelType: mtype,
	}
	err := m.GetModelId(ctx, ifoption)
	if err != nil {
		return false
	}

	return true
}