sched.go 18 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543
  1. package server
  2. import (
  3. "context"
  4. "errors"
  5. "fmt"
  6. "log/slog"
  7. "reflect"
  8. "sort"
  9. "strings"
  10. "sync"
  11. "time"
  12. "github.com/ollama/ollama/api"
  13. "github.com/ollama/ollama/format"
  14. "github.com/ollama/ollama/gpu"
  15. "github.com/ollama/ollama/llm"
  16. "github.com/ollama/ollama/server/envconfig"
  17. "golang.org/x/exp/slices"
  18. )
  19. type LlmRequest struct {
  20. ctx context.Context //nolint:containedctx
  21. model *Model
  22. opts api.Options
  23. sessionDuration time.Duration
  24. successCh chan *runnerRef
  25. errCh chan error
  26. }
  27. type Scheduler struct {
  28. pendingReqCh chan *LlmRequest
  29. finishedReqCh chan *LlmRequest
  30. expiredCh chan *runnerRef
  31. unloadedCh chan interface{}
  32. loaded map[string]*runnerRef
  33. loadedMu sync.Mutex
  34. loadFn func(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList)
  35. newServerFn func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options) (llm.LlamaServer, error)
  36. getGpuFn func() gpu.GpuInfoList
  37. }
  38. var ErrMaxQueue = fmt.Errorf("server busy, please try again. maximum pending requests exceeded")
  39. func InitScheduler(ctx context.Context) *Scheduler {
  40. sched := &Scheduler{
  41. pendingReqCh: make(chan *LlmRequest, envconfig.MaxQueuedRequests),
  42. finishedReqCh: make(chan *LlmRequest, envconfig.MaxQueuedRequests),
  43. expiredCh: make(chan *runnerRef, envconfig.MaxQueuedRequests),
  44. unloadedCh: make(chan interface{}, envconfig.MaxQueuedRequests),
  45. loaded: make(map[string]*runnerRef),
  46. newServerFn: llm.NewLlamaServer,
  47. getGpuFn: gpu.GetGPUInfo,
  48. }
  49. sched.loadFn = sched.load
  50. return sched
  51. }
  52. // context must be canceled to decrement ref count and release the runner
  53. func (s *Scheduler) GetRunner(c context.Context, model *Model, opts api.Options, sessionDuration time.Duration) (chan *runnerRef, chan error) {
  54. // allocate a large enough kv cache for all parallel requests
  55. opts.NumCtx = opts.NumCtx * envconfig.NumParallel
  56. req := &LlmRequest{
  57. ctx: c,
  58. model: model,
  59. opts: opts,
  60. sessionDuration: sessionDuration,
  61. successCh: make(chan *runnerRef),
  62. errCh: make(chan error, 1),
  63. }
  64. select {
  65. case s.pendingReqCh <- req:
  66. default:
  67. req.errCh <- ErrMaxQueue
  68. }
  69. return req.successCh, req.errCh
  70. }
  71. // Returns immediately, spawns go routines for the scheduler which will shutdown when ctx is done
  72. func (s *Scheduler) Run(ctx context.Context) {
  73. slog.Debug("starting llm scheduler")
  74. go func() {
  75. s.processPending(ctx)
  76. }()
  77. go func() {
  78. s.processCompleted(ctx)
  79. }()
  80. }
  81. func (s *Scheduler) processPending(ctx context.Context) {
  82. for {
  83. select {
  84. case <-ctx.Done():
  85. slog.Debug("shutting down scheduler pending loop")
  86. return
  87. case pending := <-s.pendingReqCh:
  88. // Block other requests until we get this pending request running
  89. for {
  90. var runnerToExpire *runnerRef
  91. s.loadedMu.Lock()
  92. runner := s.loaded[pending.model.ModelPath]
  93. loadedCount := len(s.loaded)
  94. s.loadedMu.Unlock()
  95. if runner != nil {
  96. if runner.needsReload(ctx, pending) {
  97. runnerToExpire = runner
  98. } else {
  99. // Runner is usable, return it
  100. pending.useLoadedRunner(runner, s.finishedReqCh)
  101. break
  102. }
  103. } else if envconfig.MaxRunners > 0 && loadedCount >= envconfig.MaxRunners {
  104. slog.Debug("max runners achieved, unloading one to make room", "runner_count", loadedCount)
  105. runnerToExpire = s.findRunnerToUnload()
  106. } else {
  107. // Either no models are loaded or below envconfig.MaxRunners
  108. // Get a refreshed GPU list
  109. gpus := s.getGpuFn()
  110. // Load model for fitting
  111. ggml, err := llm.LoadModel(pending.model.ModelPath)
  112. if err != nil {
  113. pending.errCh <- err
  114. break
  115. }
  116. // If we're CPU only mode, just limit by envconfig.MaxRunners above
  117. // TODO handle system memory exhaustion
  118. if (len(gpus) == 1 && gpus[0].Library == "cpu") || pending.opts.NumGPU == 0 {
  119. slog.Debug("cpu mode with existing models, loading")
  120. s.loadFn(pending, ggml, gpus)
  121. break
  122. }
  123. // No models loaded. Load the model but prefer the best fit.
  124. if loadedCount == 0 {
  125. slog.Debug("loading first model", "model", pending.model.ModelPath)
  126. g := pickBestFitGPUs(pending, ggml, gpus)
  127. if g != nil {
  128. gpus = g
  129. }
  130. s.loadFn(pending, ggml, gpus)
  131. break
  132. }
  133. // More than one loaded model, so we have to see if the new one fits
  134. // Update free memory from currently loaded models
  135. s.updateFreeSpace(gpus)
  136. gpus = pickBestFitGPUs(pending, ggml, gpus)
  137. if gpus != nil {
  138. slog.Debug("new model fits with existing models, loading")
  139. s.loadFn(pending, ggml, gpus)
  140. break
  141. }
  142. runnerToExpire = s.findRunnerToUnload()
  143. }
  144. if runnerToExpire == nil {
  145. // Shouildn't happen
  146. slog.Error("runner to expire was nil!")
  147. continue
  148. }
  149. // Trigger an expiration to unload once it's done
  150. runnerToExpire.refMu.Lock()
  151. slog.Debug("resetting model to expire immediately to make room", "model", runnerToExpire.model, "refCount", runnerToExpire.refCount)
  152. if runnerToExpire.expireTimer != nil {
  153. runnerToExpire.expireTimer.Stop()
  154. runnerToExpire.expireTimer = nil
  155. }
  156. runnerToExpire.sessionDuration = 0
  157. if runnerToExpire.refCount <= 0 {
  158. s.expiredCh <- runnerToExpire
  159. }
  160. runnerToExpire.refMu.Unlock()
  161. // Wait for the unload to happen
  162. // Note: at this point we're queueing up all incoming requests, even if they were for
  163. // a different model that's loaded and not scheduled to be removed.
  164. slog.Debug("waiting for pending requests to complete and unload to occur", "model", runnerToExpire.model)
  165. select {
  166. case <-ctx.Done():
  167. slog.Debug("shutting down scheduler pending loop")
  168. return
  169. case <-s.unloadedCh:
  170. slog.Debug("unload completed", "model", runnerToExpire.model)
  171. continue
  172. }
  173. }
  174. case <-s.unloadedCh:
  175. // An unload request when there are no pending request can be ignored
  176. slog.Debug("ignoring unload event with no pending requests")
  177. }
  178. }
  179. }
  180. func (s *Scheduler) processCompleted(ctx context.Context) {
  181. // Process completed requests, expired timers, and unloading models
  182. for {
  183. select {
  184. case <-ctx.Done():
  185. slog.Debug("shutting down scheduler completed loop")
  186. return
  187. case finished := <-s.finishedReqCh:
  188. s.loadedMu.Lock()
  189. runner := s.loaded[finished.model.ModelPath]
  190. s.loadedMu.Unlock()
  191. if runner == nil {
  192. slog.Error("finished requeset signal received after model unloaded", "model", finished.model.ModelPath)
  193. continue
  194. }
  195. runner.refMu.Lock()
  196. runner.refCount--
  197. if runner.refCount <= 0 {
  198. if runner.sessionDuration <= 0 {
  199. slog.Debug("runner with zero duration has gone idle, expiring to unload", "model", runner.model)
  200. if runner.expireTimer != nil {
  201. runner.expireTimer.Stop()
  202. runner.expireTimer = nil
  203. }
  204. s.expiredCh <- runner
  205. } else if runner.expireTimer == nil {
  206. slog.Debug("runner with non-zero duration has gone idle, adding timer", "model", runner.model, "duration", runner.sessionDuration)
  207. runner.expireTimer = time.AfterFunc(runner.sessionDuration, func() {
  208. slog.Debug("timer expired, expiring to unload", "model", runner.model)
  209. runner.refMu.Lock()
  210. defer runner.refMu.Unlock()
  211. if runner.expireTimer != nil {
  212. runner.expireTimer.Stop()
  213. runner.expireTimer = nil
  214. }
  215. s.expiredCh <- runner
  216. })
  217. } else {
  218. slog.Debug("runner with non-zero duration has gone idle, resetting timer", "model", runner.model, "duration", runner.sessionDuration)
  219. runner.expireTimer.Reset(runner.sessionDuration)
  220. }
  221. }
  222. slog.Debug("after processing request finished event", "model", runner.model, "refCount", runner.refCount)
  223. runner.refMu.Unlock()
  224. case runner := <-s.expiredCh:
  225. slog.Debug("runner expired event received", "model", runner.model)
  226. runner.refMu.Lock()
  227. if runner.refCount > 0 {
  228. // Shouldn't happen, but safeguard to ensure no leaked runners
  229. slog.Debug("expired event with positive ref count, retrying", "model", runner.model, "refCount", runner.refCount)
  230. go func(runner *runnerRef) {
  231. // We can't unload yet, but want to as soon as the current request completes
  232. // So queue up another expired event
  233. time.Sleep(10 * time.Millisecond)
  234. s.expiredCh <- runner
  235. }(runner)
  236. runner.refMu.Unlock()
  237. continue
  238. }
  239. s.loadedMu.Lock()
  240. slog.Debug("got lock to unload", "model", runner.model)
  241. runner.unload()
  242. delete(s.loaded, runner.model)
  243. s.loadedMu.Unlock()
  244. slog.Debug("runner released", "model", runner.model)
  245. runner.refMu.Unlock()
  246. slog.Debug("sending an unloaded event", "model", runner.model)
  247. s.unloadedCh <- struct{}{}
  248. }
  249. }
  250. }
  251. // Complete the pending request and send the runner back to the requester
  252. // Wires up a finished event after the request context is completed
  253. // Updates session duration, and resets expiration timer
  254. func (pending *LlmRequest) useLoadedRunner(runner *runnerRef, finished chan *LlmRequest) {
  255. runner.refMu.Lock()
  256. defer runner.refMu.Unlock()
  257. runner.refCount++
  258. if runner.expireTimer != nil {
  259. runner.expireTimer.Stop()
  260. runner.expireTimer = nil
  261. }
  262. runner.sessionDuration = pending.sessionDuration
  263. pending.successCh <- runner
  264. go func() {
  265. <-pending.ctx.Done()
  266. slog.Debug("context for request finished")
  267. finished <- pending
  268. }()
  269. }
  270. func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) {
  271. llama, err := s.newServerFn(gpus, req.model.ModelPath, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts)
  272. if err != nil {
  273. // some older models are not compatible with newer versions of llama.cpp
  274. // show a generalized compatibility error until there is a better way to
  275. // check for model compatibility
  276. if errors.Is(llm.ErrUnsupportedFormat, err) || strings.Contains(err.Error(), "failed to load model") {
  277. err = fmt.Errorf("%v: this model may be incompatible with your version of Ollama. If you previously pulled this model, try updating it by running `ollama pull %s`", err, req.model.ShortName)
  278. }
  279. slog.Info("NewLlamaServer failed", "model", req.model.ModelPath, "error", err)
  280. req.errCh <- err
  281. return
  282. }
  283. runner := &runnerRef{}
  284. runner.model = req.model.ModelPath
  285. runner.adapters = req.model.AdapterPaths
  286. runner.projectors = req.model.ProjectorPaths
  287. runner.llama = llama
  288. runner.Options = &req.opts
  289. runner.sessionDuration = req.sessionDuration
  290. runner.gpus = gpus
  291. runner.estimatedVRAM = llama.EstimatedVRAM()
  292. runner.loading = true
  293. runner.refCount = 1
  294. runner.refMu.Lock()
  295. s.loadedMu.Lock()
  296. s.loaded[req.model.ModelPath] = runner
  297. slog.Info("loaded runners", "count", len(s.loaded))
  298. s.loadedMu.Unlock()
  299. go func() {
  300. defer runner.refMu.Unlock()
  301. if err = llama.WaitUntilRunning(req.ctx); err != nil {
  302. slog.Error("error loading llama server", "error", err)
  303. runner.refCount--
  304. req.errCh <- err
  305. slog.Debug("triggering expiration for failed load", "model", runner.model)
  306. s.expiredCh <- runner
  307. return
  308. }
  309. slog.Debug("finished setting up runner", "model", req.model.ModelPath)
  310. runner.loading = false
  311. go func() {
  312. <-req.ctx.Done()
  313. slog.Debug("context for request finished")
  314. s.finishedReqCh <- req
  315. }()
  316. req.successCh <- runner
  317. }()
  318. }
  319. func (s *Scheduler) updateFreeSpace(allGpus gpu.GpuInfoList) {
  320. type predKey struct {
  321. Library string
  322. ID string
  323. }
  324. predMap := map[predKey]uint64{} // Sum up the total predicted usage per GPU for all runners
  325. s.loadedMu.Lock()
  326. for _, r := range s.loaded {
  327. r.refMu.Lock()
  328. gpuIDs := make([]string, 0, len(r.gpus))
  329. if r.llama != nil {
  330. // TODO this should be broken down by GPU instead of assuming uniform spread
  331. estimatedVRAMPerGPU := r.llama.EstimatedVRAM() / uint64(len(r.gpus))
  332. for _, gpu := range r.gpus {
  333. gpuIDs = append(gpuIDs, gpu.ID)
  334. }
  335. for _, gpu := range allGpus {
  336. if slices.Contains(gpuIDs, gpu.ID) {
  337. predMap[predKey{gpu.Library, gpu.ID}] += estimatedVRAMPerGPU
  338. }
  339. }
  340. } else {
  341. slog.Warn("unexpected nil runner reference, memory prediction may be incorrect")
  342. }
  343. r.refMu.Unlock()
  344. }
  345. s.loadedMu.Unlock()
  346. // Now that we've summed up all the GPU usage predictions across all the loaded runners, update the gpu list
  347. for i := range allGpus {
  348. if p, ok := predMap[predKey{allGpus[i].Library, allGpus[i].ID}]; ok {
  349. slog.Debug("gpu reported", "gpu", allGpus[i].ID, "library", allGpus[i].Library, "available", format.HumanBytes2(allGpus[i].FreeMemory))
  350. if p > allGpus[i].TotalMemory {
  351. // Shouldn't happen
  352. slog.Warn("predicted usage exceeds VRAM", "gpu", allGpus[i].ID, "totalMemory", allGpus[i].TotalMemory, "predicted", p)
  353. allGpus[i].FreeMemory = 0
  354. } else if (allGpus[i].TotalMemory - p) < allGpus[i].FreeMemory { // predicted free is smaller than reported free, use it
  355. // TODO maybe we should just always trust our numbers, since cuda's free memory reporting is laggy
  356. // and we might unload models we didn't actually need to. The risk is if some other GPU intensive app is loaded
  357. // after we start our first runner, then we'll never acount for that, so picking the smallest free value seems prudent.
  358. allGpus[i].FreeMemory = allGpus[i].TotalMemory - p
  359. }
  360. slog.Info("updated VRAM", "gpu", allGpus[i].ID, "library", allGpus[i].Library, "total", format.HumanBytes2(allGpus[i].TotalMemory), "available", format.HumanBytes2(allGpus[i].FreeMemory))
  361. }
  362. }
  363. }
  364. type runnerRef struct {
  365. refMu sync.Mutex
  366. // refCond sync.Cond // Signaled on transition from 1 -> 0 refCount
  367. refCount uint // prevent unloading if > 0
  368. // unloading bool // set to true when we are trying to unload the runner
  369. llama llm.LlamaServer
  370. loading bool // True only during initial load, then false forever
  371. gpus gpu.GpuInfoList // Recorded at time of provisioning
  372. estimatedVRAM uint64
  373. sessionDuration time.Duration
  374. expireTimer *time.Timer
  375. model string
  376. adapters []string
  377. projectors []string
  378. *api.Options
  379. }
  380. // The refMu must already be held when calling unload
  381. func (runner *runnerRef) unload() {
  382. if runner.expireTimer != nil {
  383. runner.expireTimer.Stop()
  384. runner.expireTimer = nil
  385. }
  386. if runner.llama != nil {
  387. runner.llama.Close()
  388. }
  389. runner.llama = nil
  390. runner.adapters = nil
  391. runner.projectors = nil
  392. runner.Options = nil
  393. runner.gpus = nil
  394. }
  395. func (runner *runnerRef) needsReload(ctx context.Context, req *LlmRequest) bool {
  396. slog.Debug("evaluating already loaded", "model", req.model.ModelPath)
  397. runner.refMu.Lock()
  398. defer runner.refMu.Unlock()
  399. timeout := 10 * time.Second
  400. if runner.loading {
  401. timeout = 2 * time.Minute // Initial load can take a long time for big models on slow systems...
  402. }
  403. // Don't reload runner if num_gpu=-1 was provided
  404. optsExisting := runner.Options.Runner
  405. optsNew := req.opts.Runner
  406. if optsNew.NumGPU < 0 {
  407. optsExisting.NumGPU = -1
  408. optsNew.NumGPU = -1
  409. }
  410. ctx, cancel := context.WithTimeout(ctx, timeout)
  411. defer cancel()
  412. if !reflect.DeepEqual(runner.adapters, req.model.AdapterPaths) || // have the adapters changed?
  413. !reflect.DeepEqual(runner.projectors, req.model.ProjectorPaths) || // have the projectors changed?
  414. !reflect.DeepEqual(optsExisting, optsNew) || // have the runner options changed?
  415. runner.llama.Ping(ctx) != nil {
  416. return true
  417. }
  418. return false
  419. }
  420. type ByDuration []*runnerRef
  421. func (a ByDuration) Len() int { return len(a) }
  422. func (a ByDuration) Swap(i, j int) { a[i], a[j] = a[j], a[i] }
  423. func (a ByDuration) Less(i, j int) bool {
  424. // uint64 to turn negative time (never unload) to largest
  425. return uint64(a[i].sessionDuration) < uint64(a[j].sessionDuration)
  426. }
  427. // TODO - future consideration to pick runners based on size
  428. // type BySize []*runnerRef
  429. // func (a BySize) Len() int { return len(a) }
  430. // func (a BySize) Swap(i, j int) { a[i], a[j] = a[j], a[i] }
  431. // func (a BySize) Less(i, j int) bool { return a[i].estimatedVRAM < a[j].estimatedVRAM }
  432. // pickBestFitGPUs will try to find the optimal placement of the model in the available GPUs where the model fully fits
  433. // If the model can not be fit fully within the available GPU(s) nil is returned
  434. func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) gpu.GpuInfoList {
  435. var estimatedVRAM uint64
  436. for _, gl := range gpus.ByLibrary() {
  437. var ok bool
  438. sgl := append(make(gpu.GpuInfoList, 0, len(gl)), gl...)
  439. // TODO - potentially sort by performance capability, existing models loaded, etc.
  440. // Note: at present, this will favor more VRAM over faster GPU speed in mixed setups
  441. sort.Sort(sort.Reverse(gpu.ByFreeMemory(sgl)))
  442. // First attempt to fit the model into a single GPU
  443. for _, g := range sgl {
  444. if ok, estimatedVRAM = llm.PredictServerFit([]gpu.GpuInfo{g}, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
  445. slog.Debug("new model will fit in available VRAM in single GPU, loading", "model", req.model.ModelPath, "gpu", g.ID, "available", g.FreeMemory, "required", format.HumanBytes2(estimatedVRAM))
  446. return []gpu.GpuInfo{g}
  447. }
  448. }
  449. // TODO future refinements
  450. // - if multiple Libraries, see if any single GPU in any Library will fit
  451. // - try subsets of GPUs instead of just falling back to 1 or all in a family
  452. // Now try all the GPUs
  453. if ok, estimatedVRAM = llm.PredictServerFit(gl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
  454. slog.Debug("new model will fit in available VRAM, loading", "model", req.model.ModelPath, "library", gl[0].Library, "required", format.HumanBytes2(estimatedVRAM))
  455. return gl
  456. }
  457. }
  458. return nil
  459. }
  460. // findRunnerToUnload finds a runner to unload to make room for a new model
  461. func (s *Scheduler) findRunnerToUnload() *runnerRef {
  462. s.loadedMu.Lock()
  463. runnerList := make([]*runnerRef, 0, len(s.loaded))
  464. for _, r := range s.loaded {
  465. runnerList = append(runnerList, r)
  466. }
  467. s.loadedMu.Unlock()
  468. // In the future we can enhance the algorithm to be smarter about picking the optimal runner to unload
  469. // e.g., if we have multiple options, will one make room for the request?
  470. sort.Sort(ByDuration(runnerList))
  471. // First try to find a runner that's already idle
  472. for _, runner := range runnerList {
  473. runner.refMu.Lock()
  474. rc := runner.refCount
  475. runner.refMu.Unlock()
  476. if rc == 0 {
  477. slog.Debug("found an idle runner to unload")
  478. return runner
  479. }
  480. }
  481. // None appear idle, just wait for the one with the shortest duration
  482. slog.Debug("no idle runners, picking the shortest duration", "count", len(runnerList))
  483. return runnerList[0]
  484. }
  485. func (s *Scheduler) unloadAllRunners() {
  486. s.loadedMu.Lock()
  487. defer s.loadedMu.Unlock()
  488. for model, runner := range s.loaded {
  489. if runner.llama != nil {
  490. slog.Debug("shutting down runner", "model", model)
  491. runner.llama.Close()
  492. }
  493. }
  494. }