You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

modelarts.go 36 kB

4 years ago
3 years ago
3 years ago
4 years ago
4 years ago
3 years ago
4 years ago
3 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
3 years ago
4 years ago
3 years ago
3 years ago
3 years ago
4 years ago
3 years ago
4 years ago
4 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
4 years ago
4 years ago
4 years ago
3 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
3 years ago
3 years ago
4 years ago
4 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
4 years ago
4 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
4 years ago
3 years ago
4 years ago
4 years ago
4 years ago
3 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
3 years ago
3 years ago
3 years ago
3 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
4 years ago
3 years ago
4 years ago
4 years ago
4 years ago
3 years ago
4 years ago
3 years ago
4 years ago
3 years ago
4 years ago
3 years ago
4 years ago
3 years ago
3 years ago
3 years ago
4 years ago
4 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
4 years ago
3 years ago
4 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263
  1. package modelarts
  2. import (
  3. "code.gitea.io/gitea/modules/modelarts_cd"
  4. "encoding/json"
  5. "errors"
  6. "fmt"
  7. "path"
  8. "strconv"
  9. "strings"
  10. "code.gitea.io/gitea/models"
  11. "code.gitea.io/gitea/modules/context"
  12. "code.gitea.io/gitea/modules/log"
  13. "code.gitea.io/gitea/modules/notification"
  14. "code.gitea.io/gitea/modules/setting"
  15. "code.gitea.io/gitea/modules/storage"
  16. "code.gitea.io/gitea/modules/timeutil"
  17. )
  18. const (
  19. //notebook
  20. storageTypeOBS = "obs"
  21. autoStopDuration = 4 * 60 * 60
  22. autoStopDurationMs = 4 * 60 * 60 * 1000
  23. MORDELART_USER_IMAGE_ENGINE_ID = -1
  24. DataSetMountPath = "/home/ma-user/work"
  25. NotebookEnv = "Python3"
  26. NotebookType = "Ascend"
  27. FlavorInfo = "Ascend: 1*Ascend 910 CPU: 24 核 96GiB (modelarts.kat1.xlarge)"
  28. //train-job
  29. // ResourcePools = "{\"resource_pool\":[{\"id\":\"pool1328035d\", \"value\":\"专属资源池\"}]}"
  30. // Engines = "{\"engine\":[{\"id\":1, \"value\":\"Ascend-Powered-Engine\"}]}"
  31. // EngineVersions = "{\"version\":[{\"id\":118,\"value\":\"MindSpore-1.0.0-c75-python3.7-euleros2.8-aarch64\"}," +
  32. // "{\"id\":119,\"value\":\"MindSpore-1.1.1-c76-python3.7-euleros2.8-aarch64\"}," +
  33. // "{\"id\":120,\"value\":\"MindSpore-1.1.1-c76-tr5-python3.7-euleros2.8-aarch64\"}," +
  34. // "{\"id\":117,\"value\":\"TF-1.15-c75-python3.7-euleros2.8-aarch64\"}" +
  35. // "]}"
  36. // TrainJobFlavorInfo = "{\"flavor\":[{\"code\":\"modelarts.bm.910.arm.public.2\",\"value\":\"Ascend : 2 * Ascend 910 CPU:48 核 512GiB\"}," +
  37. // "{\"code\":\"modelarts.bm.910.arm.public.8\",\"value\":\"Ascend : 8 * Ascend 910 CPU:192 核 2048GiB\"}," +
  38. // "{\"code\":\"modelarts.bm.910.arm.public.4\",\"value\":\"Ascend : 4 * Ascend 910 CPU:96 核 1024GiB\"}," +
  39. // "{\"code\":\"modelarts.bm.910.arm.public.1\",\"value\":\"Ascend : 1 * Ascend 910 CPU:24 核 256GiB\"}" +
  40. // "]}"
  41. CodePath = "/code/"
  42. OutputPath = "/output/"
  43. ResultPath = "/result/"
  44. LogPath = "/log/"
  45. JobPath = "/job/"
  46. OrderDesc = "desc" //向下查询
  47. OrderAsc = "asc" //向上查询
  48. Lines = 500
  49. TrainUrl = "train_url"
  50. DataUrl = "data_url"
  51. MultiDataUrl = "multi_data_url"
  52. ResultUrl = "result_url"
  53. CkptUrl = "ckpt_url"
  54. DeviceTarget = "device_target"
  55. Ascend = "Ascend"
  56. PerPage = 10
  57. IsLatestVersion = "1"
  58. NotLatestVersion = "0"
  59. VersionCountOne = 1
  60. SortByCreateTime = "create_time"
  61. ConfigTypeCustom = "custom"
  62. TotalVersionCount = 1
  63. )
  64. var (
  65. poolInfos *models.PoolInfos
  66. TrainFlavorInfos *Flavor
  67. SpecialPools *models.SpecialPools
  68. MultiNodeConfig *MultiNodes
  69. )
  70. type GenerateTrainJobReq struct {
  71. JobName string
  72. DisplayJobName string
  73. Uuid string
  74. Description string
  75. CodeObsPath string
  76. BootFile string
  77. BootFileUrl string
  78. DataUrl string
  79. TrainUrl string
  80. LogUrl string
  81. PoolID string
  82. WorkServerNumber int
  83. EngineID int64
  84. Parameters []models.Parameter
  85. CommitID string
  86. IsLatestVersion string
  87. Params string
  88. BranchName string
  89. PreVersionId int64
  90. PreVersionName string
  91. FlavorCode string
  92. FlavorName string
  93. VersionCount int
  94. EngineName string
  95. TotalVersionCount int
  96. UserImageUrl string
  97. UserCommand string
  98. DatasetName string
  99. Spec *models.Specification
  100. ModelName string
  101. LabelName string
  102. CkptName string
  103. ModelVersion string
  104. PreTrainModelUrl string
  105. }
  106. type GenerateInferenceJobReq struct {
  107. JobName string
  108. DisplayJobName string
  109. Uuid string
  110. Description string
  111. CodeObsPath string
  112. BootFile string
  113. BootFileUrl string
  114. DataUrl string
  115. TrainUrl string
  116. LogUrl string
  117. PoolID string
  118. WorkServerNumber int
  119. EngineID int64
  120. Parameters []models.Parameter
  121. CommitID string
  122. Params string
  123. BranchName string
  124. FlavorName string
  125. EngineName string
  126. LabelName string
  127. IsLatestVersion string
  128. VersionCount int
  129. TotalVersionCount int
  130. ModelName string
  131. ModelVersion string
  132. CkptName string
  133. ResultUrl string
  134. Spec *models.Specification
  135. DatasetName string
  136. }
  137. type VersionInfo struct {
  138. Version []struct {
  139. ID int `json:"id"`
  140. Value string `json:"value"`
  141. Url string `json:"url"`
  142. } `json:"version"`
  143. }
  144. type Flavor struct {
  145. Info []struct {
  146. Code string `json:"code"`
  147. Value string `json:"value"`
  148. UnitPrice int64 `json:"unitPrice"`
  149. } `json:"flavor"`
  150. }
  151. type Engine struct {
  152. Info []struct {
  153. ID int `json:"id"`
  154. Value string `json:"value"`
  155. } `json:"engine"`
  156. }
  157. type ResourcePool struct {
  158. Info []struct {
  159. ID string `json:"id"`
  160. Value string `json:"value"`
  161. } `json:"resource_pool"`
  162. }
  163. type MultiNodes struct{
  164. Info []OrgMultiNode `json:"multinode"`
  165. }
  166. type OrgMultiNode struct{
  167. Org string `json:"org"`
  168. Node []int `json:"node"`
  169. }
  170. // type Parameter struct {
  171. // Label string `json:"label"`
  172. // Value string `json:"value"`
  173. // }
  174. // type Parameters struct {
  175. // Parameter []Parameter `json:"parameter"`
  176. // }
  177. type Parameters struct {
  178. Parameter []struct {
  179. Label string `json:"label"`
  180. Value string `json:"value"`
  181. } `json:"parameter"`
  182. }
  183. func GenerateTask(ctx *context.Context, jobName, uuid, description, flavor string) error {
  184. var dataActualPath string
  185. if uuid != "" {
  186. dataActualPath = setting.Bucket + "/" + setting.BasePath + path.Join(uuid[0:1], uuid[1:2]) + "/" + uuid + "/"
  187. } else {
  188. userPath := setting.UserBasePath + ctx.User.Name + "/"
  189. isExist, err := storage.ObsHasObject(userPath)
  190. if err != nil {
  191. log.Error("ObsHasObject failed:%v", err.Error(), ctx.Data["MsgID"])
  192. return err
  193. }
  194. if !isExist {
  195. if err = storage.ObsCreateObject(userPath); err != nil {
  196. log.Error("ObsCreateObject failed:%v", err.Error(), ctx.Data["MsgID"])
  197. return err
  198. }
  199. }
  200. dataActualPath = setting.Bucket + "/" + userPath
  201. }
  202. if poolInfos == nil {
  203. json.Unmarshal([]byte(setting.PoolInfos), &poolInfos)
  204. }
  205. createTime := timeutil.TimeStampNow()
  206. jobResult, err := CreateJob(models.CreateNotebookParams{
  207. JobName: jobName,
  208. Description: description,
  209. ProfileID: setting.ProfileID,
  210. Flavor: flavor,
  211. Pool: models.Pool{
  212. ID: poolInfos.PoolInfo[0].PoolId,
  213. Name: poolInfos.PoolInfo[0].PoolName,
  214. Type: poolInfos.PoolInfo[0].PoolType,
  215. },
  216. Spec: models.Spec{
  217. Storage: models.Storage{
  218. Type: storageTypeOBS,
  219. Location: models.Location{
  220. Path: dataActualPath,
  221. },
  222. },
  223. AutoStop: models.AutoStop{
  224. Enable: true,
  225. Duration: autoStopDuration,
  226. },
  227. },
  228. })
  229. if err != nil {
  230. log.Error("CreateJob failed: %v", err.Error())
  231. return err
  232. }
  233. err = models.CreateCloudbrain(&models.Cloudbrain{
  234. Status: string(models.JobWaiting),
  235. UserID: ctx.User.ID,
  236. RepoID: ctx.Repo.Repository.ID,
  237. JobID: jobResult.ID,
  238. JobName: jobName,
  239. JobType: string(models.JobTypeDebug),
  240. Type: models.TypeCloudBrainTwo,
  241. Uuid: uuid,
  242. ComputeResource: models.NPUResource,
  243. CreatedUnix: createTime,
  244. UpdatedUnix: createTime,
  245. })
  246. if err != nil {
  247. return err
  248. }
  249. notification.NotifyOtherTask(ctx.User, ctx.Repo.Repository, jobResult.ID, jobName, models.ActionCreateDebugNPUTask)
  250. return nil
  251. }
  252. func GenerateNotebook2(ctx *context.Context, displayJobName, jobName, uuid, description, imageId string, spec *models.Specification) error {
  253. if poolInfos == nil {
  254. json.Unmarshal([]byte(setting.PoolInfos), &poolInfos)
  255. }
  256. imageName, err := GetNotebookImageName(imageId)
  257. if err != nil {
  258. log.Error("GetNotebookImageName failed: %v", err.Error())
  259. return err
  260. }
  261. createTime := timeutil.TimeStampNow()
  262. jobResult, err := createNotebook2(models.CreateNotebook2Params{
  263. JobName: jobName,
  264. Description: description,
  265. Flavor: spec.SourceSpecId,
  266. Duration: autoStopDurationMs,
  267. ImageID: imageId,
  268. PoolID: poolInfos.PoolInfo[0].PoolId,
  269. Feature: models.NotebookFeature,
  270. Volume: models.VolumeReq{
  271. Capacity: setting.Capacity,
  272. Category: models.EVSCategory,
  273. Ownership: models.ManagedOwnership,
  274. },
  275. WorkspaceID: "0",
  276. })
  277. if err != nil {
  278. log.Error("createNotebook2 failed: %v", err.Error())
  279. if strings.HasPrefix(err.Error(), UnknownErrorPrefix) {
  280. log.Info("(%s)unknown error, set temp status", displayJobName)
  281. errTemp := models.InsertCloudbrainTemp(&models.CloudbrainTemp{
  282. JobID: models.TempJobId,
  283. VersionID: models.TempVersionId,
  284. Status: models.TempJobStatus,
  285. Type: models.TypeCloudBrainTwo,
  286. JobName: jobName,
  287. JobType: string(models.JobTypeDebug),
  288. })
  289. if errTemp != nil {
  290. log.Error("InsertCloudbrainTemp failed: %v", errTemp.Error())
  291. return errTemp
  292. }
  293. }
  294. return err
  295. }
  296. task := &models.Cloudbrain{
  297. Status: jobResult.Status,
  298. UserID: ctx.User.ID,
  299. RepoID: ctx.Repo.Repository.ID,
  300. JobID: jobResult.ID,
  301. JobName: jobName,
  302. FlavorCode: spec.SourceSpecId,
  303. DisplayJobName: displayJobName,
  304. JobType: string(models.JobTypeDebug),
  305. Type: models.TypeCloudBrainTwo,
  306. Uuid: uuid,
  307. ComputeResource: models.NPUResource,
  308. Image: imageName,
  309. Description: description,
  310. CreatedUnix: createTime,
  311. UpdatedUnix: createTime,
  312. Spec: spec,
  313. }
  314. err = models.CreateCloudbrain(task)
  315. if err != nil {
  316. return err
  317. }
  318. stringId := strconv.FormatInt(task.ID, 10)
  319. notification.NotifyOtherTask(ctx.User, ctx.Repo.Repository, stringId, displayJobName, models.ActionCreateDebugNPUTask)
  320. return nil
  321. }
  322. func GenerateTrainJob(ctx *context.Context, req *GenerateTrainJobReq) (err error) {
  323. createTime := timeutil.TimeStampNow()
  324. var jobResult *models.CreateTrainJobResult
  325. var createErr error
  326. if req.EngineID < 0 {
  327. jobResult, createErr = createTrainJobUserImage(models.CreateUserImageTrainJobParams{
  328. JobName: req.JobName,
  329. Description: req.Description,
  330. Config: models.UserImageConfig{
  331. WorkServerNum: req.WorkServerNumber,
  332. AppUrl: req.CodeObsPath,
  333. BootFileUrl: req.BootFileUrl,
  334. DataUrl: req.DataUrl,
  335. TrainUrl: req.TrainUrl,
  336. LogUrl: req.LogUrl,
  337. PoolID: req.PoolID,
  338. CreateVersion: true,
  339. Flavor: models.Flavor{
  340. Code: req.Spec.SourceSpecId,
  341. },
  342. Parameter: req.Parameters,
  343. UserImageUrl: req.UserImageUrl,
  344. UserCommand: req.UserCommand,
  345. },
  346. })
  347. } else {
  348. jobResult, createErr = createTrainJob(models.CreateTrainJobParams{
  349. JobName: req.JobName,
  350. Description: req.Description,
  351. Config: models.Config{
  352. WorkServerNum: req.WorkServerNumber,
  353. AppUrl: req.CodeObsPath,
  354. BootFileUrl: req.BootFileUrl,
  355. DataUrl: req.DataUrl,
  356. EngineID: req.EngineID,
  357. TrainUrl: req.TrainUrl,
  358. LogUrl: req.LogUrl,
  359. PoolID: req.PoolID,
  360. CreateVersion: true,
  361. Flavor: models.Flavor{
  362. Code: req.Spec.SourceSpecId,
  363. },
  364. Parameter: req.Parameters,
  365. },
  366. })
  367. }
  368. if createErr != nil {
  369. log.Error("createTrainJob failed: %v", createErr.Error())
  370. if strings.HasPrefix(createErr.Error(), UnknownErrorPrefix) {
  371. log.Info("(%s)unknown error, set temp status", req.DisplayJobName)
  372. errTemp := models.InsertCloudbrainTemp(&models.CloudbrainTemp{
  373. JobID: models.TempJobId,
  374. VersionID: models.TempVersionId,
  375. Status: models.TempJobStatus,
  376. Type: models.TypeCloudBrainTwo,
  377. JobName: req.JobName,
  378. JobType: string(models.JobTypeTrain),
  379. })
  380. if errTemp != nil {
  381. log.Error("InsertCloudbrainTemp failed: %v", errTemp.Error())
  382. return errTemp
  383. }
  384. }
  385. return createErr
  386. }
  387. jobId := strconv.FormatInt(jobResult.JobID, 10)
  388. createErr = models.CreateCloudbrain(&models.Cloudbrain{
  389. Status: TransTrainJobStatus(jobResult.Status),
  390. UserID: ctx.User.ID,
  391. RepoID: ctx.Repo.Repository.ID,
  392. JobID: jobId,
  393. JobName: req.JobName,
  394. DisplayJobName: req.DisplayJobName,
  395. JobType: string(models.JobTypeTrain),
  396. Type: models.TypeCloudBrainTwo,
  397. VersionID: jobResult.VersionID,
  398. VersionName: jobResult.VersionName,
  399. Uuid: req.Uuid,
  400. DatasetName: req.DatasetName,
  401. CommitID: req.CommitID,
  402. IsLatestVersion: req.IsLatestVersion,
  403. ComputeResource: models.NPUResource,
  404. EngineID: req.EngineID,
  405. TrainUrl: req.TrainUrl,
  406. BranchName: req.BranchName,
  407. Parameters: req.Params,
  408. BootFile: req.BootFile,
  409. DataUrl: req.DataUrl,
  410. LogUrl: req.LogUrl,
  411. FlavorCode: req.Spec.SourceSpecId,
  412. Description: req.Description,
  413. WorkServerNumber: req.WorkServerNumber,
  414. FlavorName: req.FlavorName,
  415. EngineName: req.EngineName,
  416. VersionCount: req.VersionCount,
  417. TotalVersionCount: req.TotalVersionCount,
  418. CreatedUnix: createTime,
  419. UpdatedUnix: createTime,
  420. Spec: req.Spec,
  421. ModelName: req.ModelName,
  422. ModelVersion: req.ModelVersion,
  423. LabelName: req.LabelName,
  424. PreTrainModelUrl: req.PreTrainModelUrl,
  425. CkptName: req.CkptName,
  426. })
  427. if createErr != nil {
  428. log.Error("CreateCloudbrain(%s) failed:%v", req.DisplayJobName, createErr.Error())
  429. return createErr
  430. }
  431. notification.NotifyOtherTask(ctx.User, ctx.Repo.Repository, jobId, req.DisplayJobName, models.ActionCreateTrainTask)
  432. return nil
  433. }
  434. func GenerateModelConvertTrainJob(req *GenerateTrainJobReq) (*models.CreateTrainJobResult, error) {
  435. return createTrainJobUserImage(models.CreateUserImageTrainJobParams{
  436. JobName: req.JobName,
  437. Description: req.Description,
  438. Config: models.UserImageConfig{
  439. WorkServerNum: req.WorkServerNumber,
  440. AppUrl: req.CodeObsPath,
  441. BootFileUrl: req.BootFileUrl,
  442. DataUrl: req.DataUrl,
  443. TrainUrl: req.TrainUrl,
  444. LogUrl: req.LogUrl,
  445. PoolID: req.PoolID,
  446. CreateVersion: true,
  447. Flavor: models.Flavor{
  448. Code: req.FlavorCode,
  449. },
  450. Parameter: req.Parameters,
  451. UserImageUrl: req.UserImageUrl,
  452. UserCommand: req.UserCommand,
  453. },
  454. })
  455. }
  456. func GenerateTrainJobVersion(ctx *context.Context, req *GenerateTrainJobReq, jobId string) (err error) {
  457. createTime := timeutil.TimeStampNow()
  458. var jobResult *models.CreateTrainJobResult
  459. var createErr error
  460. if req.EngineID < 0 {
  461. jobResult, createErr = createTrainJobVersionUserImage(models.CreateTrainJobVersionUserImageParams{
  462. Description: req.Description,
  463. Config: models.TrainJobVersionUserImageConfig{
  464. WorkServerNum: req.WorkServerNumber,
  465. AppUrl: req.CodeObsPath,
  466. BootFileUrl: req.BootFileUrl,
  467. DataUrl: req.DataUrl,
  468. TrainUrl: req.TrainUrl,
  469. LogUrl: req.LogUrl,
  470. PoolID: req.PoolID,
  471. Flavor: models.Flavor{
  472. Code: req.Spec.SourceSpecId,
  473. },
  474. Parameter: req.Parameters,
  475. PreVersionId: req.PreVersionId,
  476. UserImageUrl: req.UserImageUrl,
  477. UserCommand: req.UserCommand,
  478. },
  479. }, jobId)
  480. } else {
  481. jobResult, createErr = createTrainJobVersion(models.CreateTrainJobVersionParams{
  482. Description: req.Description,
  483. Config: models.TrainJobVersionConfig{
  484. WorkServerNum: req.WorkServerNumber,
  485. AppUrl: req.CodeObsPath,
  486. BootFileUrl: req.BootFileUrl,
  487. DataUrl: req.DataUrl,
  488. EngineID: req.EngineID,
  489. TrainUrl: req.TrainUrl,
  490. LogUrl: req.LogUrl,
  491. PoolID: req.PoolID,
  492. Flavor: models.Flavor{
  493. Code: req.Spec.SourceSpecId,
  494. },
  495. Parameter: req.Parameters,
  496. PreVersionId: req.PreVersionId,
  497. },
  498. }, jobId)
  499. }
  500. if createErr != nil {
  501. log.Error("createTrainJobVersion failed: %v", createErr.Error())
  502. if strings.HasPrefix(createErr.Error(), UnknownErrorPrefix) {
  503. log.Info("(%s)unknown error, set temp status", req.DisplayJobName)
  504. errTemp := models.InsertCloudbrainTemp(&models.CloudbrainTemp{
  505. JobID: jobId,
  506. VersionID: models.TempVersionId,
  507. Status: models.TempJobStatus,
  508. Type: models.TypeCloudBrainTwo,
  509. JobName: req.JobName,
  510. JobType: string(models.JobTypeTrain),
  511. })
  512. if errTemp != nil {
  513. log.Error("InsertCloudbrainTemp failed: %v", errTemp.Error())
  514. return errTemp
  515. }
  516. }
  517. return createErr
  518. }
  519. var jobTypes []string
  520. jobTypes = append(jobTypes, string(models.JobTypeTrain))
  521. repo := ctx.Repo.Repository
  522. VersionTaskList, VersionListCount, createErr := models.CloudbrainsVersionList(&models.CloudbrainsOptions{
  523. RepoID: repo.ID,
  524. Type: models.TypeCloudBrainTwo,
  525. JobTypes: jobTypes,
  526. JobID: strconv.FormatInt(jobResult.JobID, 10),
  527. })
  528. if createErr != nil {
  529. ctx.ServerError("Cloudbrain", createErr)
  530. return createErr
  531. }
  532. //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
  533. createErr = models.CreateCloudbrain(&models.Cloudbrain{
  534. Status: TransTrainJobStatus(jobResult.Status),
  535. UserID: ctx.User.ID,
  536. RepoID: ctx.Repo.Repository.ID,
  537. JobID: strconv.FormatInt(jobResult.JobID, 10),
  538. JobName: req.JobName,
  539. DisplayJobName: req.DisplayJobName,
  540. JobType: string(models.JobTypeTrain),
  541. Type: models.TypeCloudBrainTwo,
  542. VersionID: jobResult.VersionID,
  543. VersionName: jobResult.VersionName,
  544. Uuid: req.Uuid,
  545. DatasetName: req.DatasetName,
  546. CommitID: req.CommitID,
  547. IsLatestVersion: req.IsLatestVersion,
  548. PreVersionName: req.PreVersionName,
  549. ComputeResource: models.NPUResource,
  550. EngineID: req.EngineID,
  551. TrainUrl: req.TrainUrl,
  552. BranchName: req.BranchName,
  553. Parameters: req.Params,
  554. BootFile: req.BootFile,
  555. DataUrl: req.DataUrl,
  556. LogUrl: req.LogUrl,
  557. PreVersionId: req.PreVersionId,
  558. FlavorCode: req.Spec.SourceSpecId,
  559. Description: req.Description,
  560. WorkServerNumber: req.WorkServerNumber,
  561. FlavorName: req.FlavorName,
  562. EngineName: req.EngineName,
  563. TotalVersionCount: VersionTaskList[0].TotalVersionCount + 1,
  564. VersionCount: VersionListCount + 1,
  565. CreatedUnix: createTime,
  566. UpdatedUnix: createTime,
  567. Spec: req.Spec,
  568. ModelName: req.ModelName,
  569. ModelVersion: req.ModelVersion,
  570. LabelName: req.LabelName,
  571. PreTrainModelUrl: req.PreTrainModelUrl,
  572. CkptName: req.CkptName,
  573. })
  574. if createErr != nil {
  575. log.Error("CreateCloudbrain(%s) failed:%v", req.JobName, createErr.Error())
  576. return createErr
  577. }
  578. //将训练任务的上一版本的isLatestVersion设置为"0"
  579. createErr = models.SetVersionCountAndLatestVersion(strconv.FormatInt(jobResult.JobID, 10), VersionTaskList[0].VersionName, VersionCountOne, NotLatestVersion, TotalVersionCount)
  580. if createErr != nil {
  581. ctx.ServerError("Update IsLatestVersion failed", createErr)
  582. return createErr
  583. }
  584. return createErr
  585. }
  586. func TransTrainJobStatus(status int) string {
  587. switch status {
  588. case 0:
  589. return "UNKNOWN"
  590. case 1:
  591. return "INIT"
  592. case 2:
  593. return "IMAGE_CREATING"
  594. case 3:
  595. return "IMAGE_FAILED"
  596. case 4:
  597. return "SUBMIT_TRYING"
  598. case 5:
  599. return "SUBMIT_FAILED"
  600. case 6:
  601. return "DELETE_FAILED"
  602. case 7:
  603. return "WAITING"
  604. case 8:
  605. return "RUNNING"
  606. case 9:
  607. return "KILLING"
  608. case 10:
  609. return "COMPLETED"
  610. case 11:
  611. return "FAILED"
  612. case 12:
  613. return "KILLED"
  614. case 13:
  615. return "CANCELED"
  616. case 14:
  617. return "LOST"
  618. case 15:
  619. return "SCALING"
  620. case 16:
  621. return "SUBMIT_MODEL_FAILED"
  622. case 17:
  623. return "DEPLOY_SERVICE_FAILED"
  624. case 18:
  625. return "CHECK_INIT"
  626. case 19:
  627. return "CHECK_RUNNING"
  628. case 20:
  629. return "CHECK_RUNNING_COMPLETED"
  630. case 21:
  631. return "CHECK_FAILED"
  632. default:
  633. return strconv.Itoa(status)
  634. }
  635. }
  636. func GetOutputPathByCount(TotalVersionCount int) (VersionOutputPath string) {
  637. talVersionCountToString := fmt.Sprintf("%04d", TotalVersionCount)
  638. VersionOutputPath = "V" + talVersionCountToString
  639. return VersionOutputPath
  640. }
  641. func GenerateInferenceJob(ctx *context.Context, req *GenerateInferenceJobReq) (err error) {
  642. createTime := timeutil.TimeStampNow()
  643. jobResult, err := createInferenceJob(models.CreateInferenceJobParams{
  644. JobName: req.JobName,
  645. Description: req.Description,
  646. InfConfig: models.InfConfig{
  647. WorkServerNum: req.WorkServerNumber,
  648. AppUrl: req.CodeObsPath,
  649. BootFileUrl: req.BootFileUrl,
  650. DataUrl: req.DataUrl,
  651. EngineID: req.EngineID,
  652. // TrainUrl: req.TrainUrl,
  653. LogUrl: req.LogUrl,
  654. PoolID: req.PoolID,
  655. CreateVersion: true,
  656. Flavor: models.Flavor{
  657. Code: req.Spec.SourceSpecId,
  658. },
  659. Parameter: req.Parameters,
  660. },
  661. })
  662. if err != nil {
  663. log.Error("createInferenceJob failed: %v", err.Error())
  664. if strings.HasPrefix(err.Error(), UnknownErrorPrefix) {
  665. log.Info("(%s)unknown error, set temp status", req.DisplayJobName)
  666. err = models.InsertCloudbrainTemp(&models.CloudbrainTemp{
  667. JobID: models.TempJobId,
  668. VersionID: models.TempVersionId,
  669. Status: models.TempJobStatus,
  670. Type: models.TypeCloudBrainTwo,
  671. JobName: req.JobName,
  672. JobType: string(models.JobTypeInference),
  673. })
  674. if err != nil {
  675. log.Error("InsertCloudbrainTemp failed: %v", err.Error())
  676. return err
  677. }
  678. }
  679. return err
  680. }
  681. // attach, err := models.GetAttachmentByUUID(req.Uuid)
  682. // if err != nil {
  683. // log.Error("GetAttachmentByUUID(%s) failed:%v", strconv.FormatInt(jobResult.JobID, 10), err.Error())
  684. // return err
  685. // }
  686. jobID := strconv.FormatInt(jobResult.JobID, 10)
  687. err = models.CreateCloudbrain(&models.Cloudbrain{
  688. Status: TransTrainJobStatus(jobResult.Status),
  689. UserID: ctx.User.ID,
  690. RepoID: ctx.Repo.Repository.ID,
  691. JobID: jobID,
  692. JobName: req.JobName,
  693. DisplayJobName: req.DisplayJobName,
  694. JobType: string(models.JobTypeInference),
  695. Type: models.TypeCloudBrainTwo,
  696. VersionID: jobResult.VersionID,
  697. VersionName: jobResult.VersionName,
  698. Uuid: req.Uuid,
  699. DatasetName: req.DatasetName,
  700. CommitID: req.CommitID,
  701. EngineID: req.EngineID,
  702. TrainUrl: req.TrainUrl,
  703. BranchName: req.BranchName,
  704. Parameters: req.Params,
  705. BootFile: req.BootFile,
  706. DataUrl: req.DataUrl,
  707. LogUrl: req.LogUrl,
  708. FlavorCode: req.Spec.SourceSpecId,
  709. Description: req.Description,
  710. WorkServerNumber: req.WorkServerNumber,
  711. FlavorName: req.FlavorName,
  712. EngineName: req.EngineName,
  713. LabelName: req.LabelName,
  714. IsLatestVersion: req.IsLatestVersion,
  715. ComputeResource: models.NPUResource,
  716. VersionCount: req.VersionCount,
  717. TotalVersionCount: req.TotalVersionCount,
  718. ModelName: req.ModelName,
  719. ModelVersion: req.ModelVersion,
  720. CkptName: req.CkptName,
  721. ResultUrl: req.ResultUrl,
  722. CreatedUnix: createTime,
  723. UpdatedUnix: createTime,
  724. Spec: req.Spec,
  725. })
  726. if err != nil {
  727. log.Error("CreateCloudbrain(%s) failed:%v", req.JobName, err.Error())
  728. return err
  729. }
  730. notification.NotifyOtherTask(ctx.User, ctx.Repo.Repository, jobID, req.DisplayJobName, models.ActionCreateInferenceTask)
  731. return nil
  732. }
  733. func GetNotebookImageName(imageId string) (string, error) {
  734. var validImage = false
  735. var imageName = ""
  736. for _, imageInfo := range setting.StImageInfos.ImageInfo {
  737. if imageInfo.Id == imageId {
  738. validImage = true
  739. imageName = imageInfo.Value
  740. }
  741. }
  742. if !validImage {
  743. log.Error("the image id(%s) is invalid", imageId)
  744. return imageName, errors.New("the image id is invalid")
  745. }
  746. return imageName, nil
  747. }
  748. func InitSpecialPool() {
  749. if SpecialPools == nil && setting.ModelArtsSpecialPools != "" {
  750. json.Unmarshal([]byte(setting.ModelArtsSpecialPools), &SpecialPools)
  751. }
  752. }
  753. func InitMultiNode(){
  754. if MultiNodeConfig ==nil && setting.ModelArtsMultiNode!=""{
  755. json.Unmarshal([]byte(setting.ModelArtsMultiNode), &MultiNodeConfig)
  756. }
  757. }
  758. func HandleTrainJobInfo(task *models.Cloudbrain) error {
  759. result, err := GetTrainJob(task.JobID, strconv.FormatInt(task.VersionID, 10))
  760. if err != nil {
  761. log.Error("GetTrainJob(%s) failed:%v", task.DisplayJobName, err)
  762. return err
  763. }
  764. if result != nil {
  765. oldStatus := task.Status
  766. task.Status = TransTrainJobStatus(result.IntStatus)
  767. task.Duration = result.Duration / 1000
  768. task.TrainJobDuration = result.TrainJobDuration
  769. if task.StartTime == 0 && result.StartTime > 0 {
  770. task.StartTime = timeutil.TimeStamp(result.StartTime / 1000)
  771. }
  772. task.TrainJobDuration = models.ConvertDurationToStr(task.Duration)
  773. if task.EndTime == 0 && models.IsTrainJobTerminal(task.Status) && task.StartTime > 0 {
  774. task.EndTime = task.StartTime.Add(task.Duration)
  775. }
  776. task.CorrectCreateUnix()
  777. if oldStatus != task.Status {
  778. notification.NotifyChangeCloudbrainStatus(task, oldStatus)
  779. }
  780. err = models.UpdateJob(task)
  781. if err != nil {
  782. log.Error("UpdateJob(%s) failed:%v", task.JobName, err)
  783. return err
  784. }
  785. }
  786. return nil
  787. }
  788. func HandleNotebookInfo(task *models.Cloudbrain) error {
  789. var result *models.GetNotebook2Result
  790. var err error
  791. if task.Type == models.TypeCloudBrainTwo {
  792. result, err = GetNotebook2(task.JobID)
  793. } else if task.Type == models.TypeCDCenter {
  794. result, err = modelarts_cd.GetNotebook(task.JobID)
  795. }
  796. if err != nil {
  797. log.Error("GetNotebook2(%s) failed:%v", task.DisplayJobName, err)
  798. return err
  799. }
  800. if result != nil {
  801. oldStatus := task.Status
  802. task.Status = result.Status
  803. if task.StartTime == 0 && result.Lease.UpdateTime > 0 {
  804. task.StartTime = timeutil.TimeStamp(result.Lease.UpdateTime / 1000)
  805. }
  806. if task.EndTime == 0 && models.IsModelArtsDebugJobTerminal(task.Status) {
  807. task.EndTime = timeutil.TimeStampNow()
  808. }
  809. task.CorrectCreateUnix()
  810. task.ComputeAndSetDuration()
  811. if oldStatus != task.Status {
  812. notification.NotifyChangeCloudbrainStatus(task, oldStatus)
  813. }
  814. if task.FlavorCode == "" {
  815. task.FlavorCode = result.Flavor
  816. }
  817. err = models.UpdateJob(task)
  818. if err != nil {
  819. log.Error("UpdateJob(%s) failed:%v", task.DisplayJobName, err)
  820. return err
  821. }
  822. }
  823. return nil
  824. }
  825. func SyncTempStatusJob() {
  826. jobs, err := models.GetCloudBrainTempJobs()
  827. if err != nil {
  828. log.Error("GetCloudBrainTempJobs failed:%v", err.Error())
  829. return
  830. }
  831. for _, temp := range jobs {
  832. log.Info("start to handle record: %s", temp.JobName)
  833. if temp.Type == models.TypeCloudBrainTwo {
  834. if temp.JobType == string(models.JobTypeDebug) {
  835. err = handleNotebook(temp)
  836. if err != nil {
  837. log.Error("handleNotebook falied:%v", err)
  838. break
  839. }
  840. } else if temp.JobType == string(models.JobTypeTrain) || temp.JobType == string(models.JobTypeInference) {
  841. _, err = models.GetCloudbrainByJobID(temp.JobID)
  842. if err != nil {
  843. //one version
  844. err = handleTrainJob(temp)
  845. if err != nil {
  846. log.Error("handleTrainJob falied:%v", err)
  847. break
  848. }
  849. } else {
  850. //multi version
  851. err = handleTrainJobMultiVersion(temp)
  852. if err != nil {
  853. log.Error("handleTrainJobMultiVersion falied:%v", err)
  854. break
  855. }
  856. }
  857. }
  858. }
  859. }
  860. return
  861. }
  862. func handleNotebook(temp *models.CloudbrainTemp) error {
  863. if temp.Status == models.TempJobStatus {
  864. err := handleTempNotebook(temp)
  865. if err != nil {
  866. log.Error("handleTempNotebook failed:%v", err)
  867. return err
  868. }
  869. } else if temp.Status == string(models.ModelArtsStopping) {
  870. res, err := GetNotebook2(temp.JobID)
  871. if err != nil {
  872. log.Error("GetNotebook2 failed:%v", err)
  873. return err
  874. }
  875. temp.Status = res.Status
  876. if temp.Status == string(models.ModelArtsStopped) {
  877. err = models.UpdateCloudbrainTemp(temp)
  878. if err != nil {
  879. log.Error("UpdateCloudbrainTemp failed:%v", err)
  880. return err
  881. }
  882. _, err := DelNotebook2(temp.JobID)
  883. if err != nil {
  884. log.Error("DelNotebook2 failed:%v", err)
  885. return err
  886. }
  887. temp.Status = string(models.ModelArtsDeleted)
  888. err = models.UpdateCloudbrainTemp(temp)
  889. if err != nil {
  890. log.Error("UpdateCloudbrainTemp failed:%v", err)
  891. return err
  892. }
  893. }
  894. }
  895. return nil
  896. }
  897. func handleTempNotebook(temp *models.CloudbrainTemp) error {
  898. var err error
  899. var isExist bool
  900. for {
  901. result, err := GetNotebookList(1000, 0, "createTime", "DESC", temp.JobName)
  902. if err != nil {
  903. log.Error("GetNotebookList failed:%v", err)
  904. break
  905. }
  906. temp.QueryTimes++
  907. err = models.UpdateCloudbrainTemp(temp)
  908. if err != nil {
  909. log.Error("UpdateCloudbrainTemp failed:%v", err)
  910. }
  911. if result != nil {
  912. for _, notebook := range result.NotebookList {
  913. if temp.JobID == models.TempJobId {
  914. //new notebook
  915. if notebook.JobName == temp.JobName {
  916. isExist = true
  917. temp.Status = notebook.Status
  918. temp.JobID = notebook.JobID
  919. break
  920. }
  921. } else {
  922. //restart: always can find one record
  923. if notebook.JobName == temp.JobName {
  924. if notebook.Status != string(models.ModelArtsStopped) {
  925. isExist = true
  926. temp.Status = notebook.Status
  927. temp.JobID = notebook.JobID
  928. break
  929. }
  930. }
  931. }
  932. }
  933. if isExist {
  934. log.Info("find the record(%s), status(%s)", temp.JobName, temp.Status)
  935. if temp.Status == string(models.ModelArtsCreateFailed) {
  936. err = models.UpdateCloudbrainTemp(temp)
  937. if err != nil {
  938. log.Error("UpdateCloudbrainTemp failed:%v", err)
  939. break
  940. }
  941. _, err := DelNotebook2(temp.JobID)
  942. if err != nil {
  943. log.Error("DelNotebook2(%s) failed:%v", temp.JobName, err)
  944. break
  945. }
  946. temp.Status = string(models.ModelArtsDeleted)
  947. } else {
  948. _, err := ManageNotebook2(temp.JobID, models.NotebookAction{Action: models.ActionStop})
  949. if err != nil {
  950. log.Error("ManageNotebook2(%s) failed:%v", temp.JobName, err)
  951. break
  952. }
  953. temp.Status = string(models.ModelArtsStopping)
  954. }
  955. models.UpdateCloudbrainTemp(temp)
  956. } else {
  957. log.Error("can not find the record(%s) till now", temp.JobName)
  958. err = errors.New("not found")
  959. break
  960. }
  961. } else {
  962. log.Error("can not find the record(%s) till now", temp.JobName)
  963. err = errors.New("not found")
  964. break
  965. }
  966. break
  967. }
  968. if temp.QueryTimes >= setting.MaxTempQueryTimes && !isExist {
  969. log.Info("reach MaxTempQueryTimes, set the job failed")
  970. temp.Status = string(models.ModelArtsTrainJobFailed)
  971. err = models.UpdateCloudbrainTemp(temp)
  972. if err != nil {
  973. log.Error("UpdateCloudbrainTemp(%s) failed:%v", temp.JobName, err)
  974. return err
  975. }
  976. }
  977. return err
  978. }
  979. func handleTrainJob(temp *models.CloudbrainTemp) error {
  980. if temp.Status == models.TempJobStatus {
  981. err := handleTempTrainJob(temp)
  982. if err != nil {
  983. log.Error("handleTempTrainJob failed:%v", err)
  984. return err
  985. }
  986. } else if temp.Status == string(models.ModelArtsTrainJobKilling) {
  987. res, err := GetTrainJob(temp.JobID, temp.VersionID)
  988. if err != nil {
  989. log.Error("GetTrainJob failed:%v", err)
  990. return err
  991. }
  992. temp.Status = TransTrainJobStatus(res.IntStatus)
  993. if temp.Status == string(models.ModelArtsTrainJobKilled) {
  994. err = models.UpdateCloudbrainTemp(temp)
  995. if err != nil {
  996. log.Error("UpdateCloudbrainTemp failed:%v", err)
  997. return err
  998. }
  999. _, err := DelTrainJob(temp.JobID)
  1000. if err != nil {
  1001. log.Error("DelTrainJob failed:%v", err)
  1002. return err
  1003. }
  1004. temp.Status = string(models.ModelArtsDeleted)
  1005. err = models.UpdateCloudbrainTemp(temp)
  1006. if err != nil {
  1007. log.Error("UpdateCloudbrainTemp failed:%v", err)
  1008. return err
  1009. }
  1010. }
  1011. }
  1012. return nil
  1013. }
  1014. func handleTrainJobMultiVersion(temp *models.CloudbrainTemp) error {
  1015. if temp.Status == models.TempJobStatus {
  1016. err := handleTempTrainJobMultiVersion(temp)
  1017. if err != nil {
  1018. log.Error("handleTempTrainJobMultiVersion failed:%v", err)
  1019. return err
  1020. }
  1021. } else if temp.Status == string(models.ModelArtsTrainJobKilling) {
  1022. res, err := GetTrainJob(temp.JobID, temp.VersionID)
  1023. if err != nil {
  1024. log.Error("GetTrainJob failed:%v", err)
  1025. return err
  1026. }
  1027. temp.Status = TransTrainJobStatus(res.IntStatus)
  1028. if temp.Status == string(models.ModelArtsTrainJobKilled) {
  1029. err = models.UpdateCloudbrainTemp(temp)
  1030. if err != nil {
  1031. log.Error("UpdateCloudbrainTemp failed:%v", err)
  1032. return err
  1033. }
  1034. _, err := DelTrainJobVersion(temp.JobID, temp.VersionID)
  1035. if err != nil {
  1036. log.Error("DelTrainJob failed:%v", err)
  1037. return err
  1038. }
  1039. temp.Status = string(models.ModelArtsDeleted)
  1040. err = models.UpdateCloudbrainTemp(temp)
  1041. if err != nil {
  1042. log.Error("UpdateCloudbrainTemp failed:%v", err)
  1043. return err
  1044. }
  1045. }
  1046. }
  1047. return nil
  1048. }
  1049. func handleTempTrainJobMultiVersion(temp *models.CloudbrainTemp) error {
  1050. var err error
  1051. var isExist bool
  1052. for {
  1053. result, err := GetTrainJobVersionList(1000, 1, temp.JobID)
  1054. if err != nil {
  1055. log.Error("GetTrainJobVersionList failed:%v", err)
  1056. break
  1057. }
  1058. temp.QueryTimes++
  1059. err = models.UpdateCloudbrainTemp(temp)
  1060. if err != nil {
  1061. log.Error("UpdateCloudbrainTemp failed:%v", err)
  1062. }
  1063. if result != nil {
  1064. count, _ := models.GetCloudbrainCountByJobName(temp.JobName, temp.JobType, temp.Type)
  1065. if result.VersionCount == int64(count+1) {
  1066. isExist = true
  1067. temp.Status = TransTrainJobStatus(result.JobVersionList[0].IntStatus)
  1068. temp.VersionID = strconv.FormatInt(result.JobVersionList[0].VersionID, 10)
  1069. log.Info("find the record(%s), status(%s)", temp.JobName, temp.Status)
  1070. _, err := StopTrainJob(temp.JobID, temp.VersionID)
  1071. if err != nil {
  1072. log.Error("StopTrainJob failed:%v", err)
  1073. break
  1074. }
  1075. temp.Status = string(models.ModelArtsTrainJobKilling)
  1076. err = models.UpdateCloudbrainTemp(temp)
  1077. if err != nil {
  1078. log.Error("UpdateCloudbrainTemp(%s) failed:%v", temp.JobName, err)
  1079. break
  1080. }
  1081. } else {
  1082. log.Error("can not find the record(%s) till now", temp.JobName)
  1083. err = errors.New("not found")
  1084. break
  1085. }
  1086. }
  1087. break
  1088. }
  1089. if temp.QueryTimes >= setting.MaxTempQueryTimes && !isExist {
  1090. log.Info("reach MaxTempQueryTimes, set the job failed")
  1091. temp.Status = string(models.ModelArtsTrainJobFailed)
  1092. err = models.UpdateCloudbrainTemp(temp)
  1093. if err != nil {
  1094. log.Error("UpdateCloudbrainTemp(%s) failed:%v", temp.JobName, err)
  1095. return err
  1096. }
  1097. }
  1098. return err
  1099. }
  1100. func handleTempTrainJob(temp *models.CloudbrainTemp) error {
  1101. var err error
  1102. var isExist bool
  1103. for {
  1104. result, err := GetTrainJobList(1000, 1, "create_time", "desc", temp.JobName)
  1105. if err != nil {
  1106. log.Error("GetTrainJobList failed:%v", err)
  1107. break
  1108. }
  1109. temp.QueryTimes++
  1110. err = models.UpdateCloudbrainTemp(temp)
  1111. if err != nil {
  1112. log.Error("UpdateCloudbrainTemp failed:%v", err)
  1113. }
  1114. if result != nil {
  1115. for _, job := range result.JobList {
  1116. if temp.JobName == job.JobName && TransTrainJobStatus(job.IntStatus) != string(models.ModelArtsTrainJobFailed) {
  1117. isExist = true
  1118. temp.Status = TransTrainJobStatus(job.IntStatus)
  1119. temp.JobID = strconv.FormatInt(job.JobID, 10)
  1120. temp.VersionID = strconv.FormatInt(job.VersionID, 10)
  1121. log.Info("find the record(%s), status(%s)", temp.JobName, temp.Status)
  1122. _, err = StopTrainJob(temp.JobID, temp.VersionID)
  1123. if err != nil {
  1124. log.Error("StopTrainJob(%s) failed:%v", temp.JobName, err)
  1125. break
  1126. }
  1127. temp.Status = string(models.ModelArtsTrainJobKilling)
  1128. err = models.UpdateCloudbrainTemp(temp)
  1129. if err != nil {
  1130. log.Error("UpdateCloudbrainTemp(%s) failed:%v", temp.JobName, err)
  1131. break
  1132. }
  1133. }
  1134. }
  1135. if !isExist {
  1136. log.Error("can not find the record(%s) till now", temp.JobName)
  1137. err = errors.New("not found")
  1138. break
  1139. }
  1140. }
  1141. break
  1142. }
  1143. if temp.QueryTimes >= setting.MaxTempQueryTimes && !isExist {
  1144. log.Info("reach MaxTempQueryTimes, set the job failed")
  1145. temp.Status = string(models.ModelArtsTrainJobFailed)
  1146. err = models.UpdateCloudbrainTemp(temp)
  1147. if err != nil {
  1148. log.Error("UpdateCloudbrainTemp(%s) failed:%v", temp.JobName, err)
  1149. return err
  1150. }
  1151. }
  1152. return err
  1153. }