aiforge/modules/modelarts/modelarts.go

package modelarts

import (
	"encoding/json"
	"fmt"
	"path"
	"strconv"

	"code.gitea.io/gitea/models"
	"code.gitea.io/gitea/modules/context"
	"code.gitea.io/gitea/modules/log"
	"code.gitea.io/gitea/modules/setting"
	"code.gitea.io/gitea/modules/storage"
)

const (
	//notebook
	storageTypeOBS   = "obs"
	autoStopDuration = 4 * 60 * 60

	DataSetMountPath = "/home/ma-user/work"
	NotebookEnv      = "Python3"
	NotebookType     = "Ascend"
	FlavorInfo       = "Ascend: 1*Ascend 910 CPU: 24 核 96GiB （modelarts.kat1.xlarge）"

	//train-job
	// ResourcePools  = "{\"resource_pool\":[{\"id\":\"pool1328035d\", \"value\":\"专属资源池\"}]}"
	// Engines        = "{\"engine\":[{\"id\":1, \"value\":\"Ascend-Powered-Engine\"}]}"
	// EngineVersions = "{\"version\":[{\"id\":118,\"value\":\"MindSpore-1.0.0-c75-python3.7-euleros2.8-aarch64\"}," +
	// 	"{\"id\":119,\"value\":\"MindSpore-1.1.1-c76-python3.7-euleros2.8-aarch64\"}," +
	// 	"{\"id\":120,\"value\":\"MindSpore-1.1.1-c76-tr5-python3.7-euleros2.8-aarch64\"}," +
	// 	"{\"id\":117,\"value\":\"TF-1.15-c75-python3.7-euleros2.8-aarch64\"}" +
	// 	"]}"
	// TrainJobFlavorInfo = "{\"flavor\":[{\"code\":\"modelarts.bm.910.arm.public.2\",\"value\":\"Ascend : 2 * Ascend 910 CPU：48 核 512GiB\"}," +
	// 	"{\"code\":\"modelarts.bm.910.arm.public.8\",\"value\":\"Ascend : 8 * Ascend 910 CPU：192 核 2048GiB\"}," +
	// 	"{\"code\":\"modelarts.bm.910.arm.public.4\",\"value\":\"Ascend : 4 * Ascend 910 CPU：96 核 1024GiB\"}," +
	// 	"{\"code\":\"modelarts.bm.910.arm.public.1\",\"value\":\"Ascend : 1 * Ascend 910 CPU：24 核 256GiB\"}" +
	// 	"]}"
	CodePath         = "/code/"
	OutputPath       = "/output/"
	LogPath          = "/log/"
	JobPath          = "/job/"
	OrderDesc        = "desc" //向下查询
	OrderAsc         = "asc"  //向上查询
	Lines            = 500
	TrainUrl         = "train_url"
	DataUrl          = "data_url"
	PerPage          = 10
	IsLatestVersion  = "1"
	NotLatestVersion = "0"
	// ComputeResource  = "NPU"
	NPUResource  = "NPU"
	GPUResource  = "CPU/GPU"
	AllResource  = "all"
	DebugType    = -1
	VersionCount = 1

	SortByCreateTime  = "create_time"
	ConfigTypeCustom  = "custom"
	TotalVersionCount = 1
)

var (
	poolInfos   *models.PoolInfos
	FlavorInfos *models.FlavorInfos
)

type GenerateTrainJobReq struct {
	JobName           string
	Uuid              string
	Description       string
	CodeObsPath       string
	BootFile          string
	BootFileUrl       string
	DataUrl           string
	TrainUrl          string
	FlavorCode        string
	LogUrl            string
	PoolID            string
	WorkServerNumber  int
	EngineID          int64
	Parameters        []models.Parameter
	CommitID          string
	IsLatestVersion   string
	Params            string
	BranchName        string
	PreVersionId      int64
	PreVersionName    string
	FlavorName        string
	VersionCount      int
	EngineName        string
	TotalVersionCount int
}

type GenerateTrainJobVersionReq struct {
	JobName           string
	Uuid              string
	Description       string
	CodeObsPath       string
	BootFile          string
	BootFileUrl       string
	DataUrl           string
	TrainUrl          string
	FlavorCode        string
	LogUrl            string
	PoolID            string
	WorkServerNumber  int
	EngineID          int64
	Parameters        []models.Parameter
	Params            string
	PreVersionId      int64
	CommitID          string
	BranchName        string
	FlavorName        string
	EngineName        string
	PreVersionName    string
	TotalVersionCount int
}

type VersionInfo struct {
	Version []struct {
		ID    int    `json:"id"`
		Value string `json:"value"`
	} `json:"version"`
}

type Flavor struct {
	Info []struct {
		Code  string `json:"code"`
		Value string `json:"value"`
	} `json:"flavor"`
}

type Engine struct {
	Info []struct {
		ID    int    `json:"id"`
		Value string `json:"value"`
	} `json:"engine"`
}

type ResourcePool struct {
	Info []struct {
		ID    string `json:"id"`
		Value string `json:"value"`
	} `json:"resource_pool"`
}

// type Parameter struct {
// 	Label string `json:"label"`
// 	Value string `json:"value"`
// }

// type Parameters struct {
// 	Parameter []Parameter `json:"parameter"`
// }

type Parameters struct {
	Parameter []struct {
		Label string `json:"label"`
		Value string `json:"value"`
	} `json:"parameter"`
}

func GenerateTask(ctx *context.Context, jobName, uuid, description, flavor string) error {
	var dataActualPath string
	if uuid != "" {
		dataActualPath = setting.Bucket + "/" + setting.BasePath + path.Join(uuid[0:1], uuid[1:2]) + "/" + uuid + "/"
	} else {
		userPath := setting.UserBasePath + ctx.User.Name + "/"
		isExist, err := storage.ObsHasObject(userPath)
		if err != nil {
			log.Error("ObsHasObject failed:%v", err.Error(), ctx.Data["MsgID"])
			return err
		}

		if !isExist {
			if err = storage.ObsCreateObject(userPath); err != nil {
				log.Error("ObsCreateObject failed:%v", err.Error(), ctx.Data["MsgID"])
				return err
			}
		}

		dataActualPath = setting.Bucket + "/" + userPath
	}

	if poolInfos == nil {
		json.Unmarshal([]byte(setting.PoolInfos), &poolInfos)
	}
	jobResult, err := CreateJob(models.CreateNotebookParams{
		JobName:     jobName,
		Description: description,
		ProfileID:   setting.ProfileID,
		Flavor:      flavor,
		Pool: models.Pool{
			ID:   poolInfos.PoolInfo[0].PoolId,
			Name: poolInfos.PoolInfo[0].PoolName,
			Type: poolInfos.PoolInfo[0].PoolType,
		},
		Spec: models.Spec{
			Storage: models.Storage{
				Type: storageTypeOBS,
				Location: models.Location{
					Path: dataActualPath,
				},
			},
			AutoStop: models.AutoStop{
				Enable:   true,
				Duration: autoStopDuration,
			},
		},
	})
	if err != nil {
		log.Error("CreateJob failed: %v", err.Error())
		return err
	}
	err = models.CreateCloudbrain(&models.Cloudbrain{

		Status:  string(models.JobWaiting),
		UserID:  ctx.User.ID,
		RepoID:  ctx.Repo.Repository.ID,
		JobID:   jobResult.ID,
		JobName: jobName,
		JobType: string(models.JobTypeDebug),
		Type:    models.TypeCloudBrainTwo,
		Uuid:    uuid,
	})

	if err != nil {
		return err
	}

	return nil
}

func GenerateTrainJob(ctx *context.Context, req *GenerateTrainJobReq) (err error) {
	jobResult, err := createTrainJob(models.CreateTrainJobParams{
		JobName:     req.JobName,
		Description: req.Description,
		Config: models.Config{
			WorkServerNum: req.WorkServerNumber,
			AppUrl:        req.CodeObsPath,
			BootFileUrl:   req.BootFileUrl,
			DataUrl:       req.DataUrl,
			EngineID:      req.EngineID,
			TrainUrl:      req.TrainUrl,
			LogUrl:        req.LogUrl,
			PoolID:        req.PoolID,
			CreateVersion: true,
			Flavor: models.Flavor{
				Code: req.FlavorCode,
			},
			Parameter: req.Parameters,
		},
	})
	if err != nil {
		log.Error("CreateJob failed: %v", err.Error())
		return err
	}

	attach, err := models.GetAttachmentByUUID(req.Uuid)
	if err != nil {
		log.Error("GetAttachmentByUUID(%s) failed:%v", strconv.FormatInt(jobResult.JobID, 10), err.Error())
		return err
	}

	err = models.CreateCloudbrain(&models.Cloudbrain{
		Status:            TransTrainJobStatus(jobResult.Status),
		UserID:            ctx.User.ID,
		RepoID:            ctx.Repo.Repository.ID,
		JobID:             strconv.FormatInt(jobResult.JobID, 10),
		JobName:           req.JobName,
		JobType:           string(models.JobTypeTrain),
		Type:              models.TypeCloudBrainTwo,
		VersionID:         jobResult.VersionID,
		VersionName:       jobResult.VersionName,
		Uuid:              req.Uuid,
		DatasetName:       attach.Name,
		CommitID:          req.CommitID,
		IsLatestVersion:   req.IsLatestVersion,
		ComputeResource:   NPUResource,
		EngineID:          req.EngineID,
		TrainUrl:          req.TrainUrl,
		BranchName:        req.BranchName,
		Parameters:        req.Params,
		BootFile:          req.BootFile,
		DataUrl:           req.DataUrl,
		LogUrl:            req.LogUrl,
		FlavorCode:        req.FlavorCode,
		Description:       req.Description,
		WorkServerNumber:  req.WorkServerNumber,
		FlavorName:        req.FlavorName,
		EngineName:        req.EngineName,
		VersionCount:      req.VersionCount,
		TotalVersionCount: req.TotalVersionCount,
	})

	if err != nil {
		log.Error("CreateCloudbrain(%s) failed:%v", req.JobName, err.Error())
		return err
	}

	return nil
}

func GenerateTrainJobVersion(ctx *context.Context, req *GenerateTrainJobReq, jobId string) (err error) {
	jobResult, err := createTrainJobVersion(models.CreateTrainJobVersionParams{
		Description: req.Description,
		Config: models.TrainJobVersionConfig{
			WorkServerNum: req.WorkServerNumber,
			AppUrl:        req.CodeObsPath,
			BootFileUrl:   req.BootFileUrl,
			DataUrl:       req.DataUrl,
			EngineID:      req.EngineID,
			TrainUrl:      req.TrainUrl,
			LogUrl:        req.LogUrl,
			PoolID:        req.PoolID,
			Flavor: models.Flavor{
				Code: req.FlavorCode,
			},
			Parameter:    req.Parameters,
			PreVersionId: req.PreVersionId,
		},
	}, jobId)
	if err != nil {
		log.Error("CreateJob failed: %v", err.Error())
		return err
	}

	attach, err := models.GetAttachmentByUUID(req.Uuid)
	if err != nil {
		log.Error("GetAttachmentByUUID(%s) failed:%v", strconv.FormatInt(jobResult.JobID, 10), err.Error())
		return err
	}

	repo := ctx.Repo.Repository
	VersionTaskList, VersionListCount, err := models.CloudbrainsVersionList(&models.CloudbrainsOptions{
		RepoID:  repo.ID,
		Type:    models.TypeCloudBrainTwo,
		JobType: string(models.JobTypeTrain),
		JobID:   strconv.FormatInt(jobResult.JobID, 10),
	})
	if err != nil {
		ctx.ServerError("Cloudbrain", err)
		return err
	}
	//将当前版本的isLatestVersion设置为"1"和任务数量更新,任务数量包括当前版本数VersionCount和历史创建的总版本数TotalVersionCount

	err = models.CreateCloudbrain(&models.Cloudbrain{
		Status:            TransTrainJobStatus(jobResult.Status),
		UserID:            ctx.User.ID,
		RepoID:            ctx.Repo.Repository.ID,
		JobID:             strconv.FormatInt(jobResult.JobID, 10),
		JobName:           req.JobName,
		JobType:           string(models.JobTypeTrain),
		Type:              models.TypeCloudBrainTwo,
		VersionID:         jobResult.VersionID,
		VersionName:       jobResult.VersionName,
		Uuid:              req.Uuid,
		DatasetName:       attach.Name,
		CommitID:          req.CommitID,
		IsLatestVersion:   req.IsLatestVersion,
		PreVersionName:    req.PreVersionName,
		ComputeResource:   NPUResource,
		EngineID:          req.EngineID,
		TrainUrl:          req.TrainUrl,
		BranchName:        req.BranchName,
		Parameters:        req.Params,
		BootFile:          req.BootFile,
		DataUrl:           req.DataUrl,
		LogUrl:            req.LogUrl,
		PreVersionId:      req.PreVersionId,
		FlavorCode:        req.FlavorCode,
		Description:       req.Description,
		WorkServerNumber:  req.WorkServerNumber,
		FlavorName:        req.FlavorName,
		EngineName:        req.EngineName,
		TotalVersionCount: VersionTaskList[0].TotalVersionCount + 1,
		VersionCount:      VersionListCount + 1,
	})
	if err != nil {
		log.Error("CreateCloudbrain(%s) failed:%v", req.JobName, err.Error())
		return err
	}

	//将训练任务的上一版本的isLatestVersion设置为"0"
	err = models.SetVersionCountAndLatestVersion(strconv.FormatInt(jobResult.JobID, 10), VersionTaskList[0].VersionName, VersionCount, NotLatestVersion, TotalVersionCount)
	if err != nil {
		ctx.ServerError("Update IsLatestVersion failed", err)
		return err
	}

	return err
}

func TransTrainJobStatus(status int) string {
	switch status {
	case 0:
		return "UNKNOWN"
	case 1:
		return "INIT"
	case 2:
		return "IMAGE_CREATING"
	case 3:
		return "IMAGE_FAILED"
	case 4:
		return "SUBMIT_TRYING"
	case 5:
		return "SUBMIT_FAILED"
	case 6:
		return "DELETE_FAILED"
	case 7:
		return "WAITING"
	case 8:
		return "RUNNING"
	case 9:
		return "KILLING"
	case 10:
		return "COMPLETED"
	case 11:
		return "FAILED"
	case 12:
		return "KILLED"
	case 13:
		return "CANCELED"
	case 14:
		return "LOST"
	case 15:
		return "SCALING"
	case 16:
		return "SUBMIT_MODEL_FAILED"
	case 17:
		return "DEPLOY_SERVICE_FAILED"
	case 18:
		return "CHECK_INIT"
	case 19:
		return "CHECK_RUNNING"
	case 20:
		return "CHECK_RUNNING_COMPLETED"
	case 21:
		return "CHECK_FAILED"

	default:
		return strconv.Itoa(status)
	}
}

func GetVersionOutputPathByTotalVersionCount(TotalVersionCount int) (VersionOutputPath string) {
	talVersionCountToString := fmt.Sprintf("%04d", TotalVersionCount)
	VersionOutputPath = "V" + talVersionCountToString
	return VersionOutputPath
}