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
3 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
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
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
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262
  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. } `json:"flavor"`
  149. }
  150. type Engine struct {
  151. Info []struct {
  152. ID int `json:"id"`
  153. Value string `json:"value"`
  154. } `json:"engine"`
  155. }
  156. type ResourcePool struct {
  157. Info []struct {
  158. ID string `json:"id"`
  159. Value string `json:"value"`
  160. } `json:"resource_pool"`
  161. }
  162. type MultiNodes struct{
  163. Info []OrgMultiNode `json:"multinode"`
  164. }
  165. type OrgMultiNode struct{
  166. Org string `json:"org"`
  167. Node []int `json:"node"`
  168. }
  169. // type Parameter struct {
  170. // Label string `json:"label"`
  171. // Value string `json:"value"`
  172. // }
  173. // type Parameters struct {
  174. // Parameter []Parameter `json:"parameter"`
  175. // }
  176. type Parameters struct {
  177. Parameter []struct {
  178. Label string `json:"label"`
  179. Value string `json:"value"`
  180. } `json:"parameter"`
  181. }
  182. func GenerateTask(ctx *context.Context, jobName, uuid, description, flavor string) error {
  183. var dataActualPath string
  184. if uuid != "" {
  185. dataActualPath = setting.Bucket + "/" + setting.BasePath + path.Join(uuid[0:1], uuid[1:2]) + "/" + uuid + "/"
  186. } else {
  187. userPath := setting.UserBasePath + ctx.User.Name + "/"
  188. isExist, err := storage.ObsHasObject(userPath)
  189. if err != nil {
  190. log.Error("ObsHasObject failed:%v", err.Error(), ctx.Data["MsgID"])
  191. return err
  192. }
  193. if !isExist {
  194. if err = storage.ObsCreateObject(userPath); err != nil {
  195. log.Error("ObsCreateObject failed:%v", err.Error(), ctx.Data["MsgID"])
  196. return err
  197. }
  198. }
  199. dataActualPath = setting.Bucket + "/" + userPath
  200. }
  201. if poolInfos == nil {
  202. json.Unmarshal([]byte(setting.PoolInfos), &poolInfos)
  203. }
  204. createTime := timeutil.TimeStampNow()
  205. jobResult, err := CreateJob(models.CreateNotebookParams{
  206. JobName: jobName,
  207. Description: description,
  208. ProfileID: setting.ProfileID,
  209. Flavor: flavor,
  210. Pool: models.Pool{
  211. ID: poolInfos.PoolInfo[0].PoolId,
  212. Name: poolInfos.PoolInfo[0].PoolName,
  213. Type: poolInfos.PoolInfo[0].PoolType,
  214. },
  215. Spec: models.Spec{
  216. Storage: models.Storage{
  217. Type: storageTypeOBS,
  218. Location: models.Location{
  219. Path: dataActualPath,
  220. },
  221. },
  222. AutoStop: models.AutoStop{
  223. Enable: true,
  224. Duration: autoStopDuration,
  225. },
  226. },
  227. })
  228. if err != nil {
  229. log.Error("CreateJob failed: %v", err.Error())
  230. return err
  231. }
  232. err = models.CreateCloudbrain(&models.Cloudbrain{
  233. Status: string(models.JobWaiting),
  234. UserID: ctx.User.ID,
  235. RepoID: ctx.Repo.Repository.ID,
  236. JobID: jobResult.ID,
  237. JobName: jobName,
  238. JobType: string(models.JobTypeDebug),
  239. Type: models.TypeCloudBrainTwo,
  240. Uuid: uuid,
  241. ComputeResource: models.NPUResource,
  242. CreatedUnix: createTime,
  243. UpdatedUnix: createTime,
  244. })
  245. if err != nil {
  246. return err
  247. }
  248. notification.NotifyOtherTask(ctx.User, ctx.Repo.Repository, jobResult.ID, jobName, models.ActionCreateDebugNPUTask)
  249. return nil
  250. }
  251. func GenerateNotebook2(ctx *context.Context, displayJobName, jobName, uuid, description, imageId string, spec *models.Specification) error {
  252. if poolInfos == nil {
  253. json.Unmarshal([]byte(setting.PoolInfos), &poolInfos)
  254. }
  255. imageName, err := GetNotebookImageName(imageId)
  256. if err != nil {
  257. log.Error("GetNotebookImageName failed: %v", err.Error())
  258. return err
  259. }
  260. createTime := timeutil.TimeStampNow()
  261. jobResult, err := createNotebook2(models.CreateNotebook2Params{
  262. JobName: jobName,
  263. Description: description,
  264. Flavor: spec.SourceSpecId,
  265. Duration: autoStopDurationMs,
  266. ImageID: imageId,
  267. PoolID: poolInfos.PoolInfo[0].PoolId,
  268. Feature: models.NotebookFeature,
  269. Volume: models.VolumeReq{
  270. Capacity: setting.Capacity,
  271. Category: models.EVSCategory,
  272. Ownership: models.ManagedOwnership,
  273. },
  274. WorkspaceID: "0",
  275. })
  276. if err != nil {
  277. log.Error("createNotebook2 failed: %v", err.Error())
  278. if strings.HasPrefix(err.Error(), UnknownErrorPrefix) {
  279. log.Info("(%s)unknown error, set temp status", displayJobName)
  280. errTemp := models.InsertCloudbrainTemp(&models.CloudbrainTemp{
  281. JobID: models.TempJobId,
  282. VersionID: models.TempVersionId,
  283. Status: models.TempJobStatus,
  284. Type: models.TypeCloudBrainTwo,
  285. JobName: jobName,
  286. JobType: string(models.JobTypeDebug),
  287. })
  288. if errTemp != nil {
  289. log.Error("InsertCloudbrainTemp failed: %v", errTemp.Error())
  290. return errTemp
  291. }
  292. }
  293. return err
  294. }
  295. task := &models.Cloudbrain{
  296. Status: jobResult.Status,
  297. UserID: ctx.User.ID,
  298. RepoID: ctx.Repo.Repository.ID,
  299. JobID: jobResult.ID,
  300. JobName: jobName,
  301. FlavorCode: spec.SourceSpecId,
  302. DisplayJobName: displayJobName,
  303. JobType: string(models.JobTypeDebug),
  304. Type: models.TypeCloudBrainTwo,
  305. Uuid: uuid,
  306. ComputeResource: models.NPUResource,
  307. Image: imageName,
  308. Description: description,
  309. CreatedUnix: createTime,
  310. UpdatedUnix: createTime,
  311. Spec: spec,
  312. }
  313. err = models.CreateCloudbrain(task)
  314. if err != nil {
  315. return err
  316. }
  317. stringId := strconv.FormatInt(task.ID, 10)
  318. notification.NotifyOtherTask(ctx.User, ctx.Repo.Repository, stringId, displayJobName, models.ActionCreateDebugNPUTask)
  319. return nil
  320. }
  321. func GenerateTrainJob(ctx *context.Context, req *GenerateTrainJobReq) (err error) {
  322. createTime := timeutil.TimeStampNow()
  323. var jobResult *models.CreateTrainJobResult
  324. var createErr error
  325. if req.EngineID < 0 {
  326. jobResult, createErr = createTrainJobUserImage(models.CreateUserImageTrainJobParams{
  327. JobName: req.JobName,
  328. Description: req.Description,
  329. Config: models.UserImageConfig{
  330. WorkServerNum: req.WorkServerNumber,
  331. AppUrl: req.CodeObsPath,
  332. BootFileUrl: req.BootFileUrl,
  333. DataUrl: req.DataUrl,
  334. TrainUrl: req.TrainUrl,
  335. LogUrl: req.LogUrl,
  336. PoolID: req.PoolID,
  337. CreateVersion: true,
  338. Flavor: models.Flavor{
  339. Code: req.Spec.SourceSpecId,
  340. },
  341. Parameter: req.Parameters,
  342. UserImageUrl: req.UserImageUrl,
  343. UserCommand: req.UserCommand,
  344. },
  345. })
  346. } else {
  347. jobResult, createErr = createTrainJob(models.CreateTrainJobParams{
  348. JobName: req.JobName,
  349. Description: req.Description,
  350. Config: models.Config{
  351. WorkServerNum: req.WorkServerNumber,
  352. AppUrl: req.CodeObsPath,
  353. BootFileUrl: req.BootFileUrl,
  354. DataUrl: req.DataUrl,
  355. EngineID: req.EngineID,
  356. TrainUrl: req.TrainUrl,
  357. LogUrl: req.LogUrl,
  358. PoolID: req.PoolID,
  359. CreateVersion: true,
  360. Flavor: models.Flavor{
  361. Code: req.Spec.SourceSpecId,
  362. },
  363. Parameter: req.Parameters,
  364. },
  365. })
  366. }
  367. if createErr != nil {
  368. log.Error("createTrainJob failed: %v", createErr.Error())
  369. if strings.HasPrefix(createErr.Error(), UnknownErrorPrefix) {
  370. log.Info("(%s)unknown error, set temp status", req.DisplayJobName)
  371. errTemp := models.InsertCloudbrainTemp(&models.CloudbrainTemp{
  372. JobID: models.TempJobId,
  373. VersionID: models.TempVersionId,
  374. Status: models.TempJobStatus,
  375. Type: models.TypeCloudBrainTwo,
  376. JobName: req.JobName,
  377. JobType: string(models.JobTypeTrain),
  378. })
  379. if errTemp != nil {
  380. log.Error("InsertCloudbrainTemp failed: %v", errTemp.Error())
  381. return errTemp
  382. }
  383. }
  384. return createErr
  385. }
  386. jobId := strconv.FormatInt(jobResult.JobID, 10)
  387. createErr = models.CreateCloudbrain(&models.Cloudbrain{
  388. Status: TransTrainJobStatus(jobResult.Status),
  389. UserID: ctx.User.ID,
  390. RepoID: ctx.Repo.Repository.ID,
  391. JobID: jobId,
  392. JobName: req.JobName,
  393. DisplayJobName: req.DisplayJobName,
  394. JobType: string(models.JobTypeTrain),
  395. Type: models.TypeCloudBrainTwo,
  396. VersionID: jobResult.VersionID,
  397. VersionName: jobResult.VersionName,
  398. Uuid: req.Uuid,
  399. DatasetName: req.DatasetName,
  400. CommitID: req.CommitID,
  401. IsLatestVersion: req.IsLatestVersion,
  402. ComputeResource: models.NPUResource,
  403. EngineID: req.EngineID,
  404. TrainUrl: req.TrainUrl,
  405. BranchName: req.BranchName,
  406. Parameters: req.Params,
  407. BootFile: req.BootFile,
  408. DataUrl: req.DataUrl,
  409. LogUrl: req.LogUrl,
  410. FlavorCode: req.Spec.SourceSpecId,
  411. Description: req.Description,
  412. WorkServerNumber: req.WorkServerNumber,
  413. FlavorName: req.FlavorName,
  414. EngineName: req.EngineName,
  415. VersionCount: req.VersionCount,
  416. TotalVersionCount: req.TotalVersionCount,
  417. CreatedUnix: createTime,
  418. UpdatedUnix: createTime,
  419. Spec: req.Spec,
  420. ModelName: req.ModelName,
  421. ModelVersion: req.ModelVersion,
  422. LabelName: req.LabelName,
  423. PreTrainModelUrl: req.PreTrainModelUrl,
  424. CkptName: req.CkptName,
  425. })
  426. if createErr != nil {
  427. log.Error("CreateCloudbrain(%s) failed:%v", req.DisplayJobName, createErr.Error())
  428. return createErr
  429. }
  430. notification.NotifyOtherTask(ctx.User, ctx.Repo.Repository, jobId, req.DisplayJobName, models.ActionCreateTrainTask)
  431. return nil
  432. }
  433. func GenerateModelConvertTrainJob(req *GenerateTrainJobReq) (*models.CreateTrainJobResult, error) {
  434. return createTrainJobUserImage(models.CreateUserImageTrainJobParams{
  435. JobName: req.JobName,
  436. Description: req.Description,
  437. Config: models.UserImageConfig{
  438. WorkServerNum: req.WorkServerNumber,
  439. AppUrl: req.CodeObsPath,
  440. BootFileUrl: req.BootFileUrl,
  441. DataUrl: req.DataUrl,
  442. TrainUrl: req.TrainUrl,
  443. LogUrl: req.LogUrl,
  444. PoolID: req.PoolID,
  445. CreateVersion: true,
  446. Flavor: models.Flavor{
  447. Code: req.FlavorCode,
  448. },
  449. Parameter: req.Parameters,
  450. UserImageUrl: req.UserImageUrl,
  451. UserCommand: req.UserCommand,
  452. },
  453. })
  454. }
  455. func GenerateTrainJobVersion(ctx *context.Context, req *GenerateTrainJobReq, jobId string) (err error) {
  456. createTime := timeutil.TimeStampNow()
  457. var jobResult *models.CreateTrainJobResult
  458. var createErr error
  459. if req.EngineID < 0 {
  460. jobResult, createErr = createTrainJobVersionUserImage(models.CreateTrainJobVersionUserImageParams{
  461. Description: req.Description,
  462. Config: models.TrainJobVersionUserImageConfig{
  463. WorkServerNum: req.WorkServerNumber,
  464. AppUrl: req.CodeObsPath,
  465. BootFileUrl: req.BootFileUrl,
  466. DataUrl: req.DataUrl,
  467. TrainUrl: req.TrainUrl,
  468. LogUrl: req.LogUrl,
  469. PoolID: req.PoolID,
  470. Flavor: models.Flavor{
  471. Code: req.Spec.SourceSpecId,
  472. },
  473. Parameter: req.Parameters,
  474. PreVersionId: req.PreVersionId,
  475. UserImageUrl: req.UserImageUrl,
  476. UserCommand: req.UserCommand,
  477. },
  478. }, jobId)
  479. } else {
  480. jobResult, createErr = createTrainJobVersion(models.CreateTrainJobVersionParams{
  481. Description: req.Description,
  482. Config: models.TrainJobVersionConfig{
  483. WorkServerNum: req.WorkServerNumber,
  484. AppUrl: req.CodeObsPath,
  485. BootFileUrl: req.BootFileUrl,
  486. DataUrl: req.DataUrl,
  487. EngineID: req.EngineID,
  488. TrainUrl: req.TrainUrl,
  489. LogUrl: req.LogUrl,
  490. PoolID: req.PoolID,
  491. Flavor: models.Flavor{
  492. Code: req.Spec.SourceSpecId,
  493. },
  494. Parameter: req.Parameters,
  495. PreVersionId: req.PreVersionId,
  496. },
  497. }, jobId)
  498. }
  499. if createErr != nil {
  500. log.Error("createTrainJobVersion failed: %v", createErr.Error())
  501. if strings.HasPrefix(createErr.Error(), UnknownErrorPrefix) {
  502. log.Info("(%s)unknown error, set temp status", req.DisplayJobName)
  503. errTemp := models.InsertCloudbrainTemp(&models.CloudbrainTemp{
  504. JobID: jobId,
  505. VersionID: models.TempVersionId,
  506. Status: models.TempJobStatus,
  507. Type: models.TypeCloudBrainTwo,
  508. JobName: req.JobName,
  509. JobType: string(models.JobTypeTrain),
  510. })
  511. if errTemp != nil {
  512. log.Error("InsertCloudbrainTemp failed: %v", errTemp.Error())
  513. return errTemp
  514. }
  515. }
  516. return createErr
  517. }
  518. var jobTypes []string
  519. jobTypes = append(jobTypes, string(models.JobTypeTrain))
  520. repo := ctx.Repo.Repository
  521. VersionTaskList, VersionListCount, createErr := models.CloudbrainsVersionList(&models.CloudbrainsOptions{
  522. RepoID: repo.ID,
  523. Type: models.TypeCloudBrainTwo,
  524. JobTypes: jobTypes,
  525. JobID: strconv.FormatInt(jobResult.JobID, 10),
  526. })
  527. if createErr != nil {
  528. ctx.ServerError("Cloudbrain", createErr)
  529. return createErr
  530. }
  531. //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
  532. createErr = models.CreateCloudbrain(&models.Cloudbrain{
  533. Status: TransTrainJobStatus(jobResult.Status),
  534. UserID: ctx.User.ID,
  535. RepoID: ctx.Repo.Repository.ID,
  536. JobID: strconv.FormatInt(jobResult.JobID, 10),
  537. JobName: req.JobName,
  538. DisplayJobName: req.DisplayJobName,
  539. JobType: string(models.JobTypeTrain),
  540. Type: models.TypeCloudBrainTwo,
  541. VersionID: jobResult.VersionID,
  542. VersionName: jobResult.VersionName,
  543. Uuid: req.Uuid,
  544. DatasetName: req.DatasetName,
  545. CommitID: req.CommitID,
  546. IsLatestVersion: req.IsLatestVersion,
  547. PreVersionName: req.PreVersionName,
  548. ComputeResource: models.NPUResource,
  549. EngineID: req.EngineID,
  550. TrainUrl: req.TrainUrl,
  551. BranchName: req.BranchName,
  552. Parameters: req.Params,
  553. BootFile: req.BootFile,
  554. DataUrl: req.DataUrl,
  555. LogUrl: req.LogUrl,
  556. PreVersionId: req.PreVersionId,
  557. FlavorCode: req.Spec.SourceSpecId,
  558. Description: req.Description,
  559. WorkServerNumber: req.WorkServerNumber,
  560. FlavorName: req.FlavorName,
  561. EngineName: req.EngineName,
  562. TotalVersionCount: VersionTaskList[0].TotalVersionCount + 1,
  563. VersionCount: VersionListCount + 1,
  564. CreatedUnix: createTime,
  565. UpdatedUnix: createTime,
  566. Spec: req.Spec,
  567. ModelName: req.ModelName,
  568. ModelVersion: req.ModelVersion,
  569. LabelName: req.LabelName,
  570. PreTrainModelUrl: req.PreTrainModelUrl,
  571. CkptName: req.CkptName,
  572. })
  573. if createErr != nil {
  574. log.Error("CreateCloudbrain(%s) failed:%v", req.JobName, createErr.Error())
  575. return createErr
  576. }
  577. //将训练任务的上一版本的isLatestVersion设置为"0"
  578. createErr = models.SetVersionCountAndLatestVersion(strconv.FormatInt(jobResult.JobID, 10), VersionTaskList[0].VersionName, VersionCountOne, NotLatestVersion, TotalVersionCount)
  579. if createErr != nil {
  580. ctx.ServerError("Update IsLatestVersion failed", createErr)
  581. return createErr
  582. }
  583. return createErr
  584. }
  585. func TransTrainJobStatus(status int) string {
  586. switch status {
  587. case 0:
  588. return "UNKNOWN"
  589. case 1:
  590. return "INIT"
  591. case 2:
  592. return "IMAGE_CREATING"
  593. case 3:
  594. return "IMAGE_FAILED"
  595. case 4:
  596. return "SUBMIT_TRYING"
  597. case 5:
  598. return "SUBMIT_FAILED"
  599. case 6:
  600. return "DELETE_FAILED"
  601. case 7:
  602. return "WAITING"
  603. case 8:
  604. return "RUNNING"
  605. case 9:
  606. return "KILLING"
  607. case 10:
  608. return "COMPLETED"
  609. case 11:
  610. return "FAILED"
  611. case 12:
  612. return "KILLED"
  613. case 13:
  614. return "CANCELED"
  615. case 14:
  616. return "LOST"
  617. case 15:
  618. return "SCALING"
  619. case 16:
  620. return "SUBMIT_MODEL_FAILED"
  621. case 17:
  622. return "DEPLOY_SERVICE_FAILED"
  623. case 18:
  624. return "CHECK_INIT"
  625. case 19:
  626. return "CHECK_RUNNING"
  627. case 20:
  628. return "CHECK_RUNNING_COMPLETED"
  629. case 21:
  630. return "CHECK_FAILED"
  631. default:
  632. return strconv.Itoa(status)
  633. }
  634. }
  635. func GetOutputPathByCount(TotalVersionCount int) (VersionOutputPath string) {
  636. talVersionCountToString := fmt.Sprintf("%04d", TotalVersionCount)
  637. VersionOutputPath = "V" + talVersionCountToString
  638. return VersionOutputPath
  639. }
  640. func GenerateInferenceJob(ctx *context.Context, req *GenerateInferenceJobReq) (err error) {
  641. createTime := timeutil.TimeStampNow()
  642. jobResult, err := createInferenceJob(models.CreateInferenceJobParams{
  643. JobName: req.JobName,
  644. Description: req.Description,
  645. InfConfig: models.InfConfig{
  646. WorkServerNum: req.WorkServerNumber,
  647. AppUrl: req.CodeObsPath,
  648. BootFileUrl: req.BootFileUrl,
  649. DataUrl: req.DataUrl,
  650. EngineID: req.EngineID,
  651. // TrainUrl: req.TrainUrl,
  652. LogUrl: req.LogUrl,
  653. PoolID: req.PoolID,
  654. CreateVersion: true,
  655. Flavor: models.Flavor{
  656. Code: req.Spec.SourceSpecId,
  657. },
  658. Parameter: req.Parameters,
  659. },
  660. })
  661. if err != nil {
  662. log.Error("createInferenceJob failed: %v", err.Error())
  663. if strings.HasPrefix(err.Error(), UnknownErrorPrefix) {
  664. log.Info("(%s)unknown error, set temp status", req.DisplayJobName)
  665. err = models.InsertCloudbrainTemp(&models.CloudbrainTemp{
  666. JobID: models.TempJobId,
  667. VersionID: models.TempVersionId,
  668. Status: models.TempJobStatus,
  669. Type: models.TypeCloudBrainTwo,
  670. JobName: req.JobName,
  671. JobType: string(models.JobTypeInference),
  672. })
  673. if err != nil {
  674. log.Error("InsertCloudbrainTemp failed: %v", err.Error())
  675. return err
  676. }
  677. }
  678. return err
  679. }
  680. // attach, err := models.GetAttachmentByUUID(req.Uuid)
  681. // if err != nil {
  682. // log.Error("GetAttachmentByUUID(%s) failed:%v", strconv.FormatInt(jobResult.JobID, 10), err.Error())
  683. // return err
  684. // }
  685. jobID := strconv.FormatInt(jobResult.JobID, 10)
  686. err = models.CreateCloudbrain(&models.Cloudbrain{
  687. Status: TransTrainJobStatus(jobResult.Status),
  688. UserID: ctx.User.ID,
  689. RepoID: ctx.Repo.Repository.ID,
  690. JobID: jobID,
  691. JobName: req.JobName,
  692. DisplayJobName: req.DisplayJobName,
  693. JobType: string(models.JobTypeInference),
  694. Type: models.TypeCloudBrainTwo,
  695. VersionID: jobResult.VersionID,
  696. VersionName: jobResult.VersionName,
  697. Uuid: req.Uuid,
  698. DatasetName: req.DatasetName,
  699. CommitID: req.CommitID,
  700. EngineID: req.EngineID,
  701. TrainUrl: req.TrainUrl,
  702. BranchName: req.BranchName,
  703. Parameters: req.Params,
  704. BootFile: req.BootFile,
  705. DataUrl: req.DataUrl,
  706. LogUrl: req.LogUrl,
  707. FlavorCode: req.Spec.SourceSpecId,
  708. Description: req.Description,
  709. WorkServerNumber: req.WorkServerNumber,
  710. FlavorName: req.FlavorName,
  711. EngineName: req.EngineName,
  712. LabelName: req.LabelName,
  713. IsLatestVersion: req.IsLatestVersion,
  714. ComputeResource: models.NPUResource,
  715. VersionCount: req.VersionCount,
  716. TotalVersionCount: req.TotalVersionCount,
  717. ModelName: req.ModelName,
  718. ModelVersion: req.ModelVersion,
  719. CkptName: req.CkptName,
  720. ResultUrl: req.ResultUrl,
  721. CreatedUnix: createTime,
  722. UpdatedUnix: createTime,
  723. Spec: req.Spec,
  724. })
  725. if err != nil {
  726. log.Error("CreateCloudbrain(%s) failed:%v", req.JobName, err.Error())
  727. return err
  728. }
  729. notification.NotifyOtherTask(ctx.User, ctx.Repo.Repository, jobID, req.DisplayJobName, models.ActionCreateInferenceTask)
  730. return nil
  731. }
  732. func GetNotebookImageName(imageId string) (string, error) {
  733. var validImage = false
  734. var imageName = ""
  735. for _, imageInfo := range setting.StImageInfos.ImageInfo {
  736. if imageInfo.Id == imageId {
  737. validImage = true
  738. imageName = imageInfo.Value
  739. }
  740. }
  741. if !validImage {
  742. log.Error("the image id(%s) is invalid", imageId)
  743. return imageName, errors.New("the image id is invalid")
  744. }
  745. return imageName, nil
  746. }
  747. func InitSpecialPool() {
  748. if SpecialPools == nil && setting.ModelArtsSpecialPools != "" {
  749. json.Unmarshal([]byte(setting.ModelArtsSpecialPools), &SpecialPools)
  750. }
  751. }
  752. func InitMultiNode(){
  753. if MultiNodeConfig ==nil && setting.ModelArtsMultiNode!=""{
  754. json.Unmarshal([]byte(setting.ModelArtsMultiNode), &MultiNodeConfig)
  755. }
  756. }
  757. func HandleTrainJobInfo(task *models.Cloudbrain) error {
  758. result, err := GetTrainJob(task.JobID, strconv.FormatInt(task.VersionID, 10))
  759. if err != nil {
  760. log.Error("GetTrainJob(%s) failed:%v", task.DisplayJobName, err)
  761. return err
  762. }
  763. if result != nil {
  764. oldStatus := task.Status
  765. task.Status = TransTrainJobStatus(result.IntStatus)
  766. task.Duration = result.Duration / 1000
  767. task.TrainJobDuration = result.TrainJobDuration
  768. if task.StartTime == 0 && result.StartTime > 0 {
  769. task.StartTime = timeutil.TimeStamp(result.StartTime / 1000)
  770. }
  771. task.TrainJobDuration = models.ConvertDurationToStr(task.Duration)
  772. if task.EndTime == 0 && models.IsTrainJobTerminal(task.Status) && task.StartTime > 0 {
  773. task.EndTime = task.StartTime.Add(task.Duration)
  774. }
  775. task.CorrectCreateUnix()
  776. if oldStatus != task.Status {
  777. notification.NotifyChangeCloudbrainStatus(task, oldStatus)
  778. }
  779. err = models.UpdateJob(task)
  780. if err != nil {
  781. log.Error("UpdateJob(%s) failed:%v", task.JobName, err)
  782. return err
  783. }
  784. }
  785. return nil
  786. }
  787. func HandleNotebookInfo(task *models.Cloudbrain) error {
  788. var result *models.GetNotebook2Result
  789. var err error
  790. if task.Type == models.TypeCloudBrainTwo {
  791. result, err = GetNotebook2(task.JobID)
  792. } else if task.Type == models.TypeCDCenter {
  793. result, err = modelarts_cd.GetNotebook(task.JobID)
  794. }
  795. if err != nil {
  796. log.Error("GetNotebook2(%s) failed:%v", task.DisplayJobName, err)
  797. return err
  798. }
  799. if result != nil {
  800. oldStatus := task.Status
  801. task.Status = result.Status
  802. if task.StartTime == 0 && result.Lease.UpdateTime > 0 {
  803. task.StartTime = timeutil.TimeStamp(result.Lease.UpdateTime / 1000)
  804. }
  805. if task.EndTime == 0 && models.IsModelArtsDebugJobTerminal(task.Status) {
  806. task.EndTime = timeutil.TimeStampNow()
  807. }
  808. task.CorrectCreateUnix()
  809. task.ComputeAndSetDuration()
  810. if oldStatus != task.Status {
  811. notification.NotifyChangeCloudbrainStatus(task, oldStatus)
  812. }
  813. if task.FlavorCode == "" {
  814. task.FlavorCode = result.Flavor
  815. }
  816. err = models.UpdateJob(task)
  817. if err != nil {
  818. log.Error("UpdateJob(%s) failed:%v", task.DisplayJobName, err)
  819. return err
  820. }
  821. }
  822. return nil
  823. }
  824. func SyncTempStatusJob() {
  825. jobs, err := models.GetCloudBrainTempJobs()
  826. if err != nil {
  827. log.Error("GetCloudBrainTempJobs failed:%v", err.Error())
  828. return
  829. }
  830. for _, temp := range jobs {
  831. log.Info("start to handle record: %s", temp.JobName)
  832. if temp.Type == models.TypeCloudBrainTwo {
  833. if temp.JobType == string(models.JobTypeDebug) {
  834. err = handleNotebook(temp)
  835. if err != nil {
  836. log.Error("handleNotebook falied:%v", err)
  837. break
  838. }
  839. } else if temp.JobType == string(models.JobTypeTrain) || temp.JobType == string(models.JobTypeInference) {
  840. _, err = models.GetCloudbrainByJobID(temp.JobID)
  841. if err != nil {
  842. //one version
  843. err = handleTrainJob(temp)
  844. if err != nil {
  845. log.Error("handleTrainJob falied:%v", err)
  846. break
  847. }
  848. } else {
  849. //multi version
  850. err = handleTrainJobMultiVersion(temp)
  851. if err != nil {
  852. log.Error("handleTrainJobMultiVersion falied:%v", err)
  853. break
  854. }
  855. }
  856. }
  857. }
  858. }
  859. return
  860. }
  861. func handleNotebook(temp *models.CloudbrainTemp) error {
  862. if temp.Status == models.TempJobStatus {
  863. err := handleTempNotebook(temp)
  864. if err != nil {
  865. log.Error("handleTempNotebook failed:%v", err)
  866. return err
  867. }
  868. } else if temp.Status == string(models.ModelArtsStopping) {
  869. res, err := GetNotebook2(temp.JobID)
  870. if err != nil {
  871. log.Error("GetNotebook2 failed:%v", err)
  872. return err
  873. }
  874. temp.Status = res.Status
  875. if temp.Status == string(models.ModelArtsStopped) {
  876. err = models.UpdateCloudbrainTemp(temp)
  877. if err != nil {
  878. log.Error("UpdateCloudbrainTemp failed:%v", err)
  879. return err
  880. }
  881. _, err := DelNotebook2(temp.JobID)
  882. if err != nil {
  883. log.Error("DelNotebook2 failed:%v", err)
  884. return err
  885. }
  886. temp.Status = string(models.ModelArtsDeleted)
  887. err = models.UpdateCloudbrainTemp(temp)
  888. if err != nil {
  889. log.Error("UpdateCloudbrainTemp failed:%v", err)
  890. return err
  891. }
  892. }
  893. }
  894. return nil
  895. }
  896. func handleTempNotebook(temp *models.CloudbrainTemp) error {
  897. var err error
  898. var isExist bool
  899. for {
  900. result, err := GetNotebookList(1000, 0, "createTime", "DESC", temp.JobName)
  901. if err != nil {
  902. log.Error("GetNotebookList failed:%v", err)
  903. break
  904. }
  905. temp.QueryTimes++
  906. err = models.UpdateCloudbrainTemp(temp)
  907. if err != nil {
  908. log.Error("UpdateCloudbrainTemp failed:%v", err)
  909. }
  910. if result != nil {
  911. for _, notebook := range result.NotebookList {
  912. if temp.JobID == models.TempJobId {
  913. //new notebook
  914. if notebook.JobName == temp.JobName {
  915. isExist = true
  916. temp.Status = notebook.Status
  917. temp.JobID = notebook.JobID
  918. break
  919. }
  920. } else {
  921. //restart: always can find one record
  922. if notebook.JobName == temp.JobName {
  923. if notebook.Status != string(models.ModelArtsStopped) {
  924. isExist = true
  925. temp.Status = notebook.Status
  926. temp.JobID = notebook.JobID
  927. break
  928. }
  929. }
  930. }
  931. }
  932. if isExist {
  933. log.Info("find the record(%s), status(%s)", temp.JobName, temp.Status)
  934. if temp.Status == string(models.ModelArtsCreateFailed) {
  935. err = models.UpdateCloudbrainTemp(temp)
  936. if err != nil {
  937. log.Error("UpdateCloudbrainTemp failed:%v", err)
  938. break
  939. }
  940. _, err := DelNotebook2(temp.JobID)
  941. if err != nil {
  942. log.Error("DelNotebook2(%s) failed:%v", temp.JobName, err)
  943. break
  944. }
  945. temp.Status = string(models.ModelArtsDeleted)
  946. } else {
  947. _, err := ManageNotebook2(temp.JobID, models.NotebookAction{Action: models.ActionStop})
  948. if err != nil {
  949. log.Error("ManageNotebook2(%s) failed:%v", temp.JobName, err)
  950. break
  951. }
  952. temp.Status = string(models.ModelArtsStopping)
  953. }
  954. models.UpdateCloudbrainTemp(temp)
  955. } else {
  956. log.Error("can not find the record(%s) till now", temp.JobName)
  957. err = errors.New("not found")
  958. break
  959. }
  960. } else {
  961. log.Error("can not find the record(%s) till now", temp.JobName)
  962. err = errors.New("not found")
  963. break
  964. }
  965. break
  966. }
  967. if temp.QueryTimes >= setting.MaxTempQueryTimes && !isExist {
  968. log.Info("reach MaxTempQueryTimes, set the job failed")
  969. temp.Status = string(models.ModelArtsTrainJobFailed)
  970. err = models.UpdateCloudbrainTemp(temp)
  971. if err != nil {
  972. log.Error("UpdateCloudbrainTemp(%s) failed:%v", temp.JobName, err)
  973. return err
  974. }
  975. }
  976. return err
  977. }
  978. func handleTrainJob(temp *models.CloudbrainTemp) error {
  979. if temp.Status == models.TempJobStatus {
  980. err := handleTempTrainJob(temp)
  981. if err != nil {
  982. log.Error("handleTempTrainJob failed:%v", err)
  983. return err
  984. }
  985. } else if temp.Status == string(models.ModelArtsTrainJobKilling) {
  986. res, err := GetTrainJob(temp.JobID, temp.VersionID)
  987. if err != nil {
  988. log.Error("GetTrainJob failed:%v", err)
  989. return err
  990. }
  991. temp.Status = TransTrainJobStatus(res.IntStatus)
  992. if temp.Status == string(models.ModelArtsTrainJobKilled) {
  993. err = models.UpdateCloudbrainTemp(temp)
  994. if err != nil {
  995. log.Error("UpdateCloudbrainTemp failed:%v", err)
  996. return err
  997. }
  998. _, err := DelTrainJob(temp.JobID)
  999. if err != nil {
  1000. log.Error("DelTrainJob failed:%v", err)
  1001. return err
  1002. }
  1003. temp.Status = string(models.ModelArtsDeleted)
  1004. err = models.UpdateCloudbrainTemp(temp)
  1005. if err != nil {
  1006. log.Error("UpdateCloudbrainTemp failed:%v", err)
  1007. return err
  1008. }
  1009. }
  1010. }
  1011. return nil
  1012. }
  1013. func handleTrainJobMultiVersion(temp *models.CloudbrainTemp) error {
  1014. if temp.Status == models.TempJobStatus {
  1015. err := handleTempTrainJobMultiVersion(temp)
  1016. if err != nil {
  1017. log.Error("handleTempTrainJobMultiVersion failed:%v", err)
  1018. return err
  1019. }
  1020. } else if temp.Status == string(models.ModelArtsTrainJobKilling) {
  1021. res, err := GetTrainJob(temp.JobID, temp.VersionID)
  1022. if err != nil {
  1023. log.Error("GetTrainJob failed:%v", err)
  1024. return err
  1025. }
  1026. temp.Status = TransTrainJobStatus(res.IntStatus)
  1027. if temp.Status == string(models.ModelArtsTrainJobKilled) {
  1028. err = models.UpdateCloudbrainTemp(temp)
  1029. if err != nil {
  1030. log.Error("UpdateCloudbrainTemp failed:%v", err)
  1031. return err
  1032. }
  1033. _, err := DelTrainJobVersion(temp.JobID, temp.VersionID)
  1034. if err != nil {
  1035. log.Error("DelTrainJob failed:%v", err)
  1036. return err
  1037. }
  1038. temp.Status = string(models.ModelArtsDeleted)
  1039. err = models.UpdateCloudbrainTemp(temp)
  1040. if err != nil {
  1041. log.Error("UpdateCloudbrainTemp failed:%v", err)
  1042. return err
  1043. }
  1044. }
  1045. }
  1046. return nil
  1047. }
  1048. func handleTempTrainJobMultiVersion(temp *models.CloudbrainTemp) error {
  1049. var err error
  1050. var isExist bool
  1051. for {
  1052. result, err := GetTrainJobVersionList(1000, 1, temp.JobID)
  1053. if err != nil {
  1054. log.Error("GetTrainJobVersionList failed:%v", err)
  1055. break
  1056. }
  1057. temp.QueryTimes++
  1058. err = models.UpdateCloudbrainTemp(temp)
  1059. if err != nil {
  1060. log.Error("UpdateCloudbrainTemp failed:%v", err)
  1061. }
  1062. if result != nil {
  1063. count, _ := models.GetCloudbrainCountByJobName(temp.JobName, temp.JobType, temp.Type)
  1064. if result.VersionCount == int64(count+1) {
  1065. isExist = true
  1066. temp.Status = TransTrainJobStatus(result.JobVersionList[0].IntStatus)
  1067. temp.VersionID = strconv.FormatInt(result.JobVersionList[0].VersionID, 10)
  1068. log.Info("find the record(%s), status(%s)", temp.JobName, temp.Status)
  1069. _, err := StopTrainJob(temp.JobID, temp.VersionID)
  1070. if err != nil {
  1071. log.Error("StopTrainJob failed:%v", err)
  1072. break
  1073. }
  1074. temp.Status = string(models.ModelArtsTrainJobKilling)
  1075. err = models.UpdateCloudbrainTemp(temp)
  1076. if err != nil {
  1077. log.Error("UpdateCloudbrainTemp(%s) failed:%v", temp.JobName, err)
  1078. break
  1079. }
  1080. } else {
  1081. log.Error("can not find the record(%s) till now", temp.JobName)
  1082. err = errors.New("not found")
  1083. break
  1084. }
  1085. }
  1086. break
  1087. }
  1088. if temp.QueryTimes >= setting.MaxTempQueryTimes && !isExist {
  1089. log.Info("reach MaxTempQueryTimes, set the job failed")
  1090. temp.Status = string(models.ModelArtsTrainJobFailed)
  1091. err = models.UpdateCloudbrainTemp(temp)
  1092. if err != nil {
  1093. log.Error("UpdateCloudbrainTemp(%s) failed:%v", temp.JobName, err)
  1094. return err
  1095. }
  1096. }
  1097. return err
  1098. }
  1099. func handleTempTrainJob(temp *models.CloudbrainTemp) error {
  1100. var err error
  1101. var isExist bool
  1102. for {
  1103. result, err := GetTrainJobList(1000, 1, "create_time", "desc", temp.JobName)
  1104. if err != nil {
  1105. log.Error("GetTrainJobList failed:%v", err)
  1106. break
  1107. }
  1108. temp.QueryTimes++
  1109. err = models.UpdateCloudbrainTemp(temp)
  1110. if err != nil {
  1111. log.Error("UpdateCloudbrainTemp failed:%v", err)
  1112. }
  1113. if result != nil {
  1114. for _, job := range result.JobList {
  1115. if temp.JobName == job.JobName && TransTrainJobStatus(job.IntStatus) != string(models.ModelArtsTrainJobFailed) {
  1116. isExist = true
  1117. temp.Status = TransTrainJobStatus(job.IntStatus)
  1118. temp.JobID = strconv.FormatInt(job.JobID, 10)
  1119. temp.VersionID = strconv.FormatInt(job.VersionID, 10)
  1120. log.Info("find the record(%s), status(%s)", temp.JobName, temp.Status)
  1121. _, err = StopTrainJob(temp.JobID, temp.VersionID)
  1122. if err != nil {
  1123. log.Error("StopTrainJob(%s) failed:%v", temp.JobName, err)
  1124. break
  1125. }
  1126. temp.Status = string(models.ModelArtsTrainJobKilling)
  1127. err = models.UpdateCloudbrainTemp(temp)
  1128. if err != nil {
  1129. log.Error("UpdateCloudbrainTemp(%s) failed:%v", temp.JobName, err)
  1130. break
  1131. }
  1132. }
  1133. }
  1134. if !isExist {
  1135. log.Error("can not find the record(%s) till now", temp.JobName)
  1136. err = errors.New("not found")
  1137. break
  1138. }
  1139. }
  1140. break
  1141. }
  1142. if temp.QueryTimes >= setting.MaxTempQueryTimes && !isExist {
  1143. log.Info("reach MaxTempQueryTimes, set the job failed")
  1144. temp.Status = string(models.ModelArtsTrainJobFailed)
  1145. err = models.UpdateCloudbrainTemp(temp)
  1146. if err != nil {
  1147. log.Error("UpdateCloudbrainTemp(%s) failed:%v", temp.JobName, err)
  1148. return err
  1149. }
  1150. }
  1151. return err
  1152. }