sched.go 24.4 KB
Newer Older
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1
2
3
4
5
6
7
8
package server

import (
	"context"
	"errors"
	"fmt"
	"log/slog"
	"reflect"
9
	"runtime"
Daniel Hiltgen's avatar
Daniel Hiltgen committed
10
11
12
13
14
15
	"sort"
	"strings"
	"sync"
	"time"

	"github.com/ollama/ollama/api"
16
	"github.com/ollama/ollama/envconfig"
Daniel Hiltgen's avatar
Daniel Hiltgen committed
17
18
19
20
21
22
23
24
25
26
27
28
	"github.com/ollama/ollama/format"
	"github.com/ollama/ollama/gpu"
	"github.com/ollama/ollama/llm"
)

type LlmRequest struct {
	ctx             context.Context //nolint:containedctx
	model           *Model
	opts            api.Options
	sessionDuration time.Duration
	successCh       chan *runnerRef
	errCh           chan error
29
	schedAttempts   uint
Daniel Hiltgen's avatar
Daniel Hiltgen committed
30
31
32
33
34
35
36
37
38
39
40
}

type Scheduler struct {
	pendingReqCh  chan *LlmRequest
	finishedReqCh chan *LlmRequest
	expiredCh     chan *runnerRef
	unloadedCh    chan interface{}

	loaded   map[string]*runnerRef
	loadedMu sync.Mutex

41
42
43
44
45
	loadFn       func(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList)
	newServerFn  func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options) (llm.LlamaServer, error)
	getGpuFn     func() gpu.GpuInfoList
	getCpuFn     func() gpu.GpuInfoList
	reschedDelay time.Duration
Daniel Hiltgen's avatar
Daniel Hiltgen committed
46
47
}

48
var ErrMaxQueue = fmt.Errorf("server busy, please try again.  maximum pending requests exceeded")
Daniel Hiltgen's avatar
Daniel Hiltgen committed
49
50
51

func InitScheduler(ctx context.Context) *Scheduler {
	sched := &Scheduler{
52
53
54
55
		pendingReqCh:  make(chan *LlmRequest, envconfig.MaxQueuedRequests),
		finishedReqCh: make(chan *LlmRequest, envconfig.MaxQueuedRequests),
		expiredCh:     make(chan *runnerRef, envconfig.MaxQueuedRequests),
		unloadedCh:    make(chan interface{}, envconfig.MaxQueuedRequests),
Daniel Hiltgen's avatar
Daniel Hiltgen committed
56
57
58
		loaded:        make(map[string]*runnerRef),
		newServerFn:   llm.NewLlamaServer,
		getGpuFn:      gpu.GetGPUInfo,
59
		getCpuFn:      gpu.GetCPUInfo,
60
		reschedDelay:  250 * time.Millisecond,
Daniel Hiltgen's avatar
Daniel Hiltgen committed
61
62
63
64
65
66
67
	}
	sched.loadFn = sched.load
	return sched
}

// context must be canceled to decrement ref count and release the runner
func (s *Scheduler) GetRunner(c context.Context, model *Model, opts api.Options, sessionDuration time.Duration) (chan *runnerRef, chan error) {
68
	// allocate a large enough kv cache for all parallel requests
69
70
71
72
	if opts.NumCtx < 4 {
		opts.NumCtx = 4
	}

Michael Yang's avatar
Michael Yang committed
73
	opts.NumCtx *= envconfig.NumParallel
74

Daniel Hiltgen's avatar
Daniel Hiltgen committed
75
76
77
78
79
80
81
82
	req := &LlmRequest{
		ctx:             c,
		model:           model,
		opts:            opts,
		sessionDuration: sessionDuration,
		successCh:       make(chan *runnerRef),
		errCh:           make(chan error, 1),
	}
83

Daniel Hiltgen's avatar
Daniel Hiltgen committed
84
85
86
	select {
	case s.pendingReqCh <- req:
	default:
87
		req.errCh <- ErrMaxQueue
Daniel Hiltgen's avatar
Daniel Hiltgen committed
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
	}
	return req.successCh, req.errCh
}

// Returns immediately, spawns go routines for the scheduler which will shutdown when ctx is done
func (s *Scheduler) Run(ctx context.Context) {
	slog.Debug("starting llm scheduler")
	go func() {
		s.processPending(ctx)
	}()

	go func() {
		s.processCompleted(ctx)
	}()
}

func (s *Scheduler) processPending(ctx context.Context) {
	for {
		select {
		case <-ctx.Done():
			slog.Debug("shutting down scheduler pending loop")
			return
		case pending := <-s.pendingReqCh:
			// Block other requests until we get this pending request running
112
			pending.schedAttempts++
113
114
115
116
117
118

			if pending.ctx.Err() != nil {
				slog.Debug("pending request cancelled or timed out, skipping scheduling")
				continue
			}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
119
120
121
122
123
124
125
126
127
128
129
130
131
132
			for {
				var runnerToExpire *runnerRef
				s.loadedMu.Lock()
				runner := s.loaded[pending.model.ModelPath]
				loadedCount := len(s.loaded)
				s.loadedMu.Unlock()
				if runner != nil {
					if runner.needsReload(ctx, pending) {
						runnerToExpire = runner
					} else {
						// Runner is usable, return it
						pending.useLoadedRunner(runner, s.finishedReqCh)
						break
					}
133
				} else if envconfig.MaxRunners > 0 && loadedCount >= envconfig.MaxRunners {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
134
					slog.Debug("max runners achieved, unloading one to make room", "runner_count", loadedCount)
135
					runnerToExpire = s.findRunnerToUnload()
Daniel Hiltgen's avatar
Daniel Hiltgen committed
136
				} else {
137
					// Either no models are loaded or below envconfig.MaxRunners
Daniel Hiltgen's avatar
Daniel Hiltgen committed
138
					// Get a refreshed GPU list
139
140
141
142
143
144
					var gpus gpu.GpuInfoList
					if pending.opts.NumGPU == 0 {
						gpus = s.getCpuFn()
					} else {
						gpus = s.getGpuFn()
					}
145

146
					// Load model for fitting
147
148
149
150
151
					ggml, err := llm.LoadModel(pending.model.ModelPath)
					if err != nil {
						pending.errCh <- err
						break
					}
152

153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
					// Evaluate if the model will fit in the available system memory, or if we should unload a model first
					if len(gpus) == 1 && gpus[0].Library == "cpu" {
						if loadedCount == 0 {
							slog.Debug("cpu mode with first model, loading")
							s.loadFn(pending, ggml, gpus)
							break
						}
						runnerToExpire = s.maybeFindCPURunnerToUnload(pending, ggml, gpus)
						if runnerToExpire == nil {
							slog.Debug("cpu mode with available system memory or first model, loading")
							s.loadFn(pending, ggml, gpus)
							break
						}
						// else we need to expire a runner
					} else if loadedCount == 0 {
						// No models loaded. Load the model but prefer the best fit.
169
170
171
172
173
174
175
176
177
						slog.Debug("loading first model", "model", pending.model.ModelPath)
						g := pickBestFitGPUs(pending, ggml, gpus)
						if g != nil {
							gpus = g
						}
						s.loadFn(pending, ggml, gpus)
						break
					}

178
					if runnerToExpire == nil {
179
180
181
182
183
184
185
186
						// More than one loaded model, so we have to see if the
						// new one fits
						//
						// We want to avoid loading on any GPUs that have other
						// models still loading on them to avoid potential races
						// with VRAM consumption ramping up during load
						availGpus := s.filterGPUsWithLoadingModels(gpus)

187
						// Update free memory from currently loaded models
188
189
190
						s.updateFreeSpace(availGpus)
						fitGpus := pickBestFitGPUs(pending, ggml, availGpus)
						if fitGpus != nil {
191
							slog.Debug("new model fits with existing models, loading")
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
							s.loadFn(pending, ggml, fitGpus)
							break
						}

						// We couldn't find a set of GPUs to fully load the new
						// model. If no other models are loading (both GPU lists
						// are the same) then we need to unload another model to
						// make room
						if len(availGpus) < len(gpus) {
							// There are other requests pending, and this one
							// needs more time, so put it on the back of the
							// queue so that we might satisfy other pending
							// requests that aren't blocked
							go func() {
								// Process in a go routine to avoid deadlocking
								// the scheduler if our queue is full
								slog.Debug("delaying scheduling while other models finish loading", "attempts", pending.schedAttempts, "model", pending.model.ModelPath)
								time.Sleep(s.reschedDelay)
								s.pendingReqCh <- pending
							}()
212
213
214
							break
						}
						runnerToExpire = s.findRunnerToUnload()
Daniel Hiltgen's avatar
Daniel Hiltgen committed
215
216
217
218
219
220
221
222
223
224
					}
				}

				if runnerToExpire == nil {
					// Shouildn't happen
					slog.Error("runner to expire was nil!")
					continue
				}
				// Trigger an expiration to unload once it's done
				runnerToExpire.refMu.Lock()
225
				slog.Debug("resetting model to expire immediately to make room", "modelPath", runnerToExpire.modelPath, "refCount", runnerToExpire.refCount)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
226
227
228
229
230
231
232
233
234
235
236
237
				if runnerToExpire.expireTimer != nil {
					runnerToExpire.expireTimer.Stop()
					runnerToExpire.expireTimer = nil
				}
				runnerToExpire.sessionDuration = 0
				if runnerToExpire.refCount <= 0 {
					s.expiredCh <- runnerToExpire
				}
				runnerToExpire.refMu.Unlock()
				// Wait for the unload to happen
				// Note: at this point we're queueing up all incoming requests, even if they were for
				// a different model that's loaded and not scheduled to be removed.
238
				slog.Debug("waiting for pending requests to complete and unload to occur", "modelPath", runnerToExpire.modelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
239
240
241
242
243
				select {
				case <-ctx.Done():
					slog.Debug("shutting down scheduler pending loop")
					return
				case <-s.unloadedCh:
244
					slog.Debug("unload completed", "modelPath", runnerToExpire.modelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
					continue
				}
			}
		case <-s.unloadedCh:
			// An unload request when there are no pending request can be ignored
			slog.Debug("ignoring unload event with no pending requests")
		}
	}
}

func (s *Scheduler) processCompleted(ctx context.Context) {
	// Process completed requests, expired timers, and unloading models
	for {
		select {
		case <-ctx.Done():
			slog.Debug("shutting down scheduler completed loop")
			return
		case finished := <-s.finishedReqCh:
			s.loadedMu.Lock()
			runner := s.loaded[finished.model.ModelPath]
			s.loadedMu.Unlock()
			if runner == nil {
267
				slog.Error("finished request signal received after model unloaded", "modelPath", finished.model.ModelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
268
269
270
271
272
273
				continue
			}
			runner.refMu.Lock()
			runner.refCount--
			if runner.refCount <= 0 {
				if runner.sessionDuration <= 0 {
274
					slog.Debug("runner with zero duration has gone idle, expiring to unload", "modelPath", runner.modelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
275
276
277
278
279
280
					if runner.expireTimer != nil {
						runner.expireTimer.Stop()
						runner.expireTimer = nil
					}
					s.expiredCh <- runner
				} else if runner.expireTimer == nil {
281
					slog.Debug("runner with non-zero duration has gone idle, adding timer", "modelPath", runner.modelPath, "duration", runner.sessionDuration)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
282
					runner.expireTimer = time.AfterFunc(runner.sessionDuration, func() {
283
						slog.Debug("timer expired, expiring to unload", "modelPath", runner.modelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
284
285
286
287
						runner.refMu.Lock()
						defer runner.refMu.Unlock()
						if runner.expireTimer != nil {
							runner.expireTimer.Stop()
288
							runner.expireTimer = nil
Daniel Hiltgen's avatar
Daniel Hiltgen committed
289
290
291
						}
						s.expiredCh <- runner
					})
292
					runner.expiresAt = time.Now().Add(runner.sessionDuration)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
293
				} else {
294
					slog.Debug("runner with non-zero duration has gone idle, resetting timer", "modelPath", runner.modelPath, "duration", runner.sessionDuration)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
295
					runner.expireTimer.Reset(runner.sessionDuration)
296
					runner.expiresAt = time.Now().Add(runner.sessionDuration)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
297
298
				}
			}
299
			slog.Debug("after processing request finished event", "modelPath", runner.modelPath, "refCount", runner.refCount)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
300
301
			runner.refMu.Unlock()
		case runner := <-s.expiredCh:
302
			slog.Debug("runner expired event received", "modelPath", runner.modelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
303
304
305
			runner.refMu.Lock()
			if runner.refCount > 0 {
				// Shouldn't happen, but safeguard to ensure no leaked runners
306
				slog.Debug("expired event with positive ref count, retrying", "modelPath", runner.modelPath, "refCount", runner.refCount)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
307
308
309
310
311
312
313
314
315
316
				go func(runner *runnerRef) {
					// We can't unload yet, but want to as soon as the current request completes
					// So queue up another expired event
					time.Sleep(10 * time.Millisecond)
					s.expiredCh <- runner
				}(runner)
				runner.refMu.Unlock()
				continue
			}

317
			s.loadedMu.Lock()
318
			slog.Debug("got lock to unload", "modelPath", runner.modelPath)
319
			finished := runner.waitForVRAMRecovery()
Daniel Hiltgen's avatar
Daniel Hiltgen committed
320
			runner.unload()
321
			delete(s.loaded, runner.modelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
322
			s.loadedMu.Unlock()
323
			slog.Debug("runner released", "modelPath", runner.modelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
324
			runner.refMu.Unlock()
325
326

			<-finished
327
			slog.Debug("sending an unloaded event", "modelPath", runner.modelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
328
329
330
331
332
333
334
335
336
337
338
339
			s.unloadedCh <- struct{}{}
		}
	}
}

// Complete the pending request and send the runner back to the requester
// Wires up a finished event after the request context is completed
// Updates session duration, and resets expiration timer
func (pending *LlmRequest) useLoadedRunner(runner *runnerRef, finished chan *LlmRequest) {
	runner.refMu.Lock()
	defer runner.refMu.Unlock()
	runner.refCount++
340
341
342
343
	if runner.expireTimer != nil {
		runner.expireTimer.Stop()
		runner.expireTimer = nil
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
344
345
346
347
348
349
350
351
352
	runner.sessionDuration = pending.sessionDuration
	pending.successCh <- runner
	go func() {
		<-pending.ctx.Done()
		slog.Debug("context for request finished")
		finished <- pending
	}()
}

353
354
func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) {
	llama, err := s.newServerFn(gpus, req.model.ModelPath, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
355
356
357
358
359
360
361
362
363
364
365
	if err != nil {
		// some older models are not compatible with newer versions of llama.cpp
		// show a generalized compatibility error until there is a better way to
		// check for model compatibility
		if errors.Is(llm.ErrUnsupportedFormat, err) || strings.Contains(err.Error(), "failed to load model") {
			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)
		}
		slog.Info("NewLlamaServer failed", "model", req.model.ModelPath, "error", err)
		req.errCh <- err
		return
	}
366
367
368
369
370
371
372
373
374
375
376
377
	runner := &runnerRef{
		model:           req.model,
		modelPath:       req.model.ModelPath,
		llama:           llama,
		Options:         &req.opts,
		sessionDuration: req.sessionDuration,
		gpus:            gpus,
		estimatedVRAM:   llama.EstimatedVRAM(),
		estimatedTotal:  llama.EstimatedTotal(),
		loading:         true,
		refCount:        1,
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
378
	runner.refMu.Lock()
379

Daniel Hiltgen's avatar
Daniel Hiltgen committed
380
381
382
383
384
385
386
387
388
389
390
	s.loadedMu.Lock()
	s.loaded[req.model.ModelPath] = runner
	slog.Info("loaded runners", "count", len(s.loaded))
	s.loadedMu.Unlock()

	go func() {
		defer runner.refMu.Unlock()
		if err = llama.WaitUntilRunning(req.ctx); err != nil {
			slog.Error("error loading llama server", "error", err)
			runner.refCount--
			req.errCh <- err
391
			slog.Debug("triggering expiration for failed load", "model", runner.modelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
			s.expiredCh <- runner
			return
		}
		slog.Debug("finished setting up runner", "model", req.model.ModelPath)
		runner.loading = false
		go func() {
			<-req.ctx.Done()
			slog.Debug("context for request finished")
			s.finishedReqCh <- req
		}()
		req.successCh <- runner
	}()
}

func (s *Scheduler) updateFreeSpace(allGpus gpu.GpuInfoList) {
	type predKey struct {
		Library string
		ID      string
	}
	predMap := map[predKey]uint64{} // Sum up the total predicted usage per GPU for all runners
	s.loadedMu.Lock()
	for _, r := range s.loaded {
		r.refMu.Lock()
		if r.llama != nil {
			for _, gpu := range allGpus {
417
418
419
				// if slices.Contains(gpuIDs, gpu.ID) {
				predMap[predKey{gpu.Library, gpu.ID}] += r.llama.EstimagedVRAMByGPU(gpu.ID)
				// }
Daniel Hiltgen's avatar
Daniel Hiltgen committed
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
			}
		} else {
			slog.Warn("unexpected nil runner reference, memory prediction may be incorrect")
		}
		r.refMu.Unlock()
	}
	s.loadedMu.Unlock()

	// Now that we've summed up all the GPU usage predictions across all the loaded runners, update the gpu list
	for i := range allGpus {
		if p, ok := predMap[predKey{allGpus[i].Library, allGpus[i].ID}]; ok {
			slog.Debug("gpu reported", "gpu", allGpus[i].ID, "library", allGpus[i].Library, "available", format.HumanBytes2(allGpus[i].FreeMemory))
			if p > allGpus[i].TotalMemory {
				// Shouldn't happen
				slog.Warn("predicted usage exceeds VRAM", "gpu", allGpus[i].ID, "totalMemory", allGpus[i].TotalMemory, "predicted", p)
				allGpus[i].FreeMemory = 0
			} else if (allGpus[i].TotalMemory - p) < allGpus[i].FreeMemory { // predicted free is smaller than reported free, use it
				// TODO maybe we should just always trust our numbers, since cuda's free memory reporting is laggy
				// and we might unload models we didn't actually need to.  The risk is if some other GPU intensive app is loaded
				// after we start our first runner, then we'll never acount for that, so picking the smallest free value seems prudent.
				allGpus[i].FreeMemory = allGpus[i].TotalMemory - p
			}
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
			slog.Info("updated VRAM based on existing loaded models", "gpu", allGpus[i].ID, "library", allGpus[i].Library, "total", format.HumanBytes2(allGpus[i].TotalMemory), "available", format.HumanBytes2(allGpus[i].FreeMemory))
		}
	}
}

// While models are loading the VRAM consumption numbers will be indeterminate, so we have
// to avoid scheduling another model on the same GPU(s) that haven't stabilized.
// This routine returns the set of GPUs that do not have an active loading model.
// If all GPUs have loading models, an empty list will be returned (not a single CPU entry)
func (s *Scheduler) filterGPUsWithLoadingModels(allGpus gpu.GpuInfoList) gpu.GpuInfoList {
	ret := append(gpu.GpuInfoList{}, allGpus...)
	s.loadedMu.Lock()
	defer s.loadedMu.Unlock()
	for _, runner := range s.loaded {
		if runner.loading {
			slog.Debug("overlapping loads detected", "gpus", runner.gpus, "model", runner.modelPath)
			for _, busyGPU := range runner.gpus {
				for i := range ret {
					if ret[i].ID == busyGPU.ID {
						ret = append(ret[:i], ret[i+1:]...)
						break
					}
				}
			}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
466
467
		}
	}
468
	return ret
Daniel Hiltgen's avatar
Daniel Hiltgen committed
469
470
}

471
// TODO consolidate sched_types.go
Daniel Hiltgen's avatar
Daniel Hiltgen committed
472
473
474
475
476
477
type runnerRef struct {
	refMu sync.Mutex
	// refCond   sync.Cond // Signaled on transition from 1 -> 0 refCount
	refCount uint // prevent unloading if > 0
	// unloading bool      // set to true when we are trying to unload the runner

478
479
480
481
482
	llama          llm.LlamaServer
	loading        bool            // True only during initial load, then false forever
	gpus           gpu.GpuInfoList // Recorded at time of provisioning
	estimatedVRAM  uint64
	estimatedTotal uint64
Daniel Hiltgen's avatar
Daniel Hiltgen committed
483
484
485

	sessionDuration time.Duration
	expireTimer     *time.Timer
486
	expiresAt       time.Time
Daniel Hiltgen's avatar
Daniel Hiltgen committed
487

488
489
	model     *Model
	modelPath string
Daniel Hiltgen's avatar
Daniel Hiltgen committed
490
491
492
493
494
	*api.Options
}

// The refMu must already be held when calling unload
func (runner *runnerRef) unload() {
495
496
497
498
	if runner.expireTimer != nil {
		runner.expireTimer.Stop()
		runner.expireTimer = nil
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
499
500
501
	if runner.llama != nil {
		runner.llama.Close()
	}
502
	runner.model = nil
Daniel Hiltgen's avatar
Daniel Hiltgen committed
503
504
505
506
507
508
509
510
511
	runner.llama = nil
	runner.Options = nil
	runner.gpus = nil
}

func (runner *runnerRef) needsReload(ctx context.Context, req *LlmRequest) bool {
	slog.Debug("evaluating already loaded", "model", req.model.ModelPath)
	runner.refMu.Lock()
	defer runner.refMu.Unlock()
512

Daniel Hiltgen's avatar
Daniel Hiltgen committed
513
514
515
516
	timeout := 10 * time.Second
	if runner.loading {
		timeout = 2 * time.Minute // Initial load can take a long time for big models on slow systems...
	}
517

518
519
520
521
	if runner.Options == nil {
		return true
	}

522
523
524
525
526
527
528
529
530
	// Don't reload runner if num_gpu=-1 was provided
	optsExisting := runner.Options.Runner
	optsNew := req.opts.Runner
	if optsNew.NumGPU < 0 {
		optsExisting.NumGPU = -1
		optsNew.NumGPU = -1
	}

	ctx, cancel := context.WithTimeout(ctx, timeout)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
531
	defer cancel()
532
533
	if !reflect.DeepEqual(runner.model.AdapterPaths, req.model.AdapterPaths) || // have the adapters changed?
		!reflect.DeepEqual(runner.model.ProjectorPaths, req.model.ProjectorPaths) || // have the projectors changed?
Daniel Hiltgen's avatar
Daniel Hiltgen committed
534
535
536
537
		!reflect.DeepEqual(optsExisting, optsNew) || // have the runner options changed?
		runner.llama.Ping(ctx) != nil {
		return true
	}
538

Daniel Hiltgen's avatar
Daniel Hiltgen committed
539
540
541
	return false
}

542
543
544
545
546
547
548
549
550
551
552
// Free memory reporting on GPUs can lag for a while even after the runner
// exits, so we have to keep checking until we see the available memory recover,
// otherwise subsequent model loads will get far less layers loaded or worse
// case, may completely fall back to CPU mode.
// This routine must be called before the runner unloads so it can establish
// a before and after GPU memory allocation.  The returned channel
// will be notified when we're done waiting, or have timed out and should
// proceed anyway
func (runner *runnerRef) waitForVRAMRecovery() chan interface{} {
	finished := make(chan interface{}, 1)

553
554
	// CPU or Metal don't need checking, so no waiting required
	// windows can page VRAM, only cuda currently can report accurate used vram usage
555
556
	if len(runner.gpus) == 0 ||
		(len(runner.gpus) == 1 && (runner.gpus[0].Library == "cpu" || runner.gpus[0].Library == "metal")) ||
557
		(runtime.GOOS == "windows" && runner.gpus[0].Library != "cuda") {
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
		finished <- struct{}{}
		return finished
	}
	start := time.Now()

	// Establish a baseline before we unload
	gpusBefore := gpu.GetGPUInfo()
	var totalMemoryBefore, freeMemoryBefore uint64
	for _, gpu := range gpusBefore {
		totalMemoryBefore += gpu.TotalMemory
		freeMemoryBefore += gpu.FreeMemory
	}
	go func() {
		expiresAt := start.Add(5 * time.Second) // typical convergence is 0.5-1.5s
		ticker := time.NewTicker(250 * time.Millisecond)
		defer ticker.Stop()
		for {
			<-ticker.C
			if time.Now().After(expiresAt) {
577
				slog.Warn("gpu VRAM usage didn't recover within timeout", "seconds", time.Since(start).Seconds(), "model", runner.modelPath)
578
579
580
581
582
583
584
585
586
587
588
589
				finished <- struct{}{}
			}

			// Query GPUs, look for free to go back up
			gpusNow := gpu.GetGPUInfo()
			var totalMemoryNow, freeMemoryNow uint64
			for _, gpu := range gpusNow {
				totalMemoryNow += gpu.TotalMemory
				freeMemoryNow += gpu.FreeMemory
			}
			// If we're within ~80% of the estimated memory usage recovered, bail out
			if float32(freeMemoryNow-freeMemoryBefore) > float32(runner.estimatedVRAM)*0.8 {
590
				slog.Debug(fmt.Sprintf("gpu VRAM free memory converged after %0.2f seconds", time.Since(start).Seconds()), "model", runner.modelPath)
591
592
593
594
595
596
597
598
				finished <- struct{}{}
				return
			}
		}
	}()
	return finished
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
type ByDuration []*runnerRef

func (a ByDuration) Len() int      { return len(a) }
func (a ByDuration) Swap(i, j int) { a[i], a[j] = a[j], a[i] }
func (a ByDuration) Less(i, j int) bool {
	// uint64 to turn negative time (never unload) to largest
	return uint64(a[i].sessionDuration) < uint64(a[j].sessionDuration)
}

// TODO - future consideration to pick runners based on size
// type BySize []*runnerRef
// func (a BySize) Len() int           { return len(a) }
// func (a BySize) Swap(i, j int)      { a[i], a[j] = a[j], a[i] }
// func (a BySize) Less(i, j int) bool { return a[i].estimatedVRAM < a[j].estimatedVRAM }

// pickBestFitGPUs will try to find the optimal placement of the model in the available GPUs where the model fully fits
// If the model can not be fit fully within the available GPU(s) nil is returned
616
func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) gpu.GpuInfoList {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
617
618
619
620
621
622
623
624
625
626
	var estimatedVRAM uint64
	for _, gl := range gpus.ByLibrary() {
		var ok bool
		sgl := append(make(gpu.GpuInfoList, 0, len(gl)), gl...)

		// TODO - potentially sort by performance capability, existing models loaded, etc.
		// Note: at present, this will favor more VRAM over faster GPU speed in mixed setups
		sort.Sort(sort.Reverse(gpu.ByFreeMemory(sgl)))

		// First attempt to fit the model into a single GPU
627
628
629
630
631
632
		if !envconfig.SchedSpread {
			for _, g := range sgl {
				if ok, estimatedVRAM = llm.PredictServerFit([]gpu.GpuInfo{g}, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
					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))
					return []gpu.GpuInfo{g}
				}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
633
634
635
636
637
638
639
640
			}
		}

		// TODO future refinements
		// - if multiple Libraries, see if any single GPU in any Library will fit
		// - try subsets of GPUs instead of just falling back to 1 or all in a family

		// Now try all the GPUs
641
642
643
		if ok, estimatedVRAM = llm.PredictServerFit(sgl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
			slog.Debug("new model will fit in available VRAM, loading", "model", req.model.ModelPath, "library", sgl[0].Library, "required", format.HumanBytes2(estimatedVRAM))
			return sgl
Daniel Hiltgen's avatar
Daniel Hiltgen committed
644
645
646
647
648
649
		}
	}
	return nil
}

// findRunnerToUnload finds a runner to unload to make room for a new model
650
func (s *Scheduler) findRunnerToUnload() *runnerRef {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
651
652
653
654
655
656
	s.loadedMu.Lock()
	runnerList := make([]*runnerRef, 0, len(s.loaded))
	for _, r := range s.loaded {
		runnerList = append(runnerList, r)
	}
	s.loadedMu.Unlock()
657
658
659
660
	if len(runnerList) == 0 {
		slog.Debug("no loaded runner to unload")
		return nil
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690

	// In the future we can enhance the algorithm to be smarter about picking the optimal runner to unload
	// e.g., if we have multiple options, will one make room for the request?
	sort.Sort(ByDuration(runnerList))

	// First try to find a runner that's already idle
	for _, runner := range runnerList {
		runner.refMu.Lock()
		rc := runner.refCount
		runner.refMu.Unlock()
		if rc == 0 {
			slog.Debug("found an idle runner to unload")
			return runner
		}
	}
	// None appear idle, just wait for the one with the shortest duration
	slog.Debug("no idle runners, picking the shortest duration", "count", len(runnerList))
	return runnerList[0]
}

func (s *Scheduler) unloadAllRunners() {
	s.loadedMu.Lock()
	defer s.loadedMu.Unlock()
	for model, runner := range s.loaded {
		if runner.llama != nil {
			slog.Debug("shutting down runner", "model", model)
			runner.llama.Close()
		}
	}
}
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706

// If other runners are loaded, make sure the pending request will fit in system memory
// If not, pick a runner to unload, else return nil and the request can be loaded
func (s *Scheduler) maybeFindCPURunnerToUnload(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) *runnerRef {
	slog.Debug("evaluating if CPU model load will fit in available system memory")
	estimate := llm.EstimateGPULayers(gpus, ggml, req.model.ProjectorPaths, req.opts)
	if estimate.TotalSize <= gpus[0].FreeMemory {
		slog.Debug("cpu inference mode, model fits in available system memory", "model", format.HumanBytes2(estimate.TotalSize), "available", format.HumanBytes2(gpus[0].FreeMemory))
		return nil
	}

	// TODO - optimization: try to find CPU only runners first, or partial offloads with enough in system memory to make room

	return s.findRunnerToUnload()

}