sched.go 31.7 KB
Newer Older
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1
2
3
4
5
6
7
package server

import (
	"context"
	"errors"
	"fmt"
	"log/slog"
Michael Yang's avatar
int  
Michael Yang committed
8
	"os"
Daniel Hiltgen's avatar
Daniel Hiltgen committed
9
	"reflect"
10
	"runtime"
11
	"slices"
Daniel Hiltgen's avatar
Daniel Hiltgen committed
12
	"sort"
Michael Yang's avatar
int  
Michael Yang committed
13
	"strconv"
Daniel Hiltgen's avatar
Daniel Hiltgen committed
14
15
16
17
18
	"strings"
	"sync"
	"time"

	"github.com/ollama/ollama/api"
19
	"github.com/ollama/ollama/discover"
20
	"github.com/ollama/ollama/envconfig"
Daniel Hiltgen's avatar
Daniel Hiltgen committed
21
	"github.com/ollama/ollama/format"
Michael Yang's avatar
Michael Yang committed
22
	"github.com/ollama/ollama/fs/ggml"
Daniel Hiltgen's avatar
Daniel Hiltgen committed
23
	"github.com/ollama/ollama/llm"
24
	"github.com/ollama/ollama/types/model"
Daniel Hiltgen's avatar
Daniel Hiltgen committed
25
26
27
28
29
30
)

type LlmRequest struct {
	ctx             context.Context //nolint:containedctx
	model           *Model
	opts            api.Options
Daniel Hiltgen's avatar
Daniel Hiltgen committed
31
	origNumCtx      int // Track the initial ctx request
32
	sessionDuration *api.Duration
Daniel Hiltgen's avatar
Daniel Hiltgen committed
33
34
	successCh       chan *runnerRef
	errCh           chan error
35
	schedAttempts   uint
Daniel Hiltgen's avatar
Daniel Hiltgen committed
36
37
38
39
40
41
}

type Scheduler struct {
	pendingReqCh  chan *LlmRequest
	finishedReqCh chan *LlmRequest
	expiredCh     chan *runnerRef
42
	unloadedCh    chan any
Daniel Hiltgen's avatar
Daniel Hiltgen committed
43
44
45
46

	loaded   map[string]*runnerRef
	loadedMu sync.Mutex

Michael Yang's avatar
Michael Yang committed
47
48
	loadFn       func(req *LlmRequest, f *ggml.GGML, gpus discover.GpuInfoList, numParallel int)
	newServerFn  func(gpus discover.GpuInfoList, model string, f *ggml.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error)
49
50
	getGpuFn     func() discover.GpuInfoList
	getCpuFn     func() discover.GpuInfoList
51
	reschedDelay time.Duration
Daniel Hiltgen's avatar
Daniel Hiltgen committed
52
53
}

54
55
56
57
58
59
// Default automatic value for number of models we allow per GPU
// Model will still need to fit in VRAM, but loading many small models
// on a large GPU can cause stalling
var defaultModelsPerGPU = 3

// Default automatic value for parallel setting
60
var defaultParallel = 1
61

Michael Yang's avatar
lint  
Michael Yang committed
62
var ErrMaxQueue = errors.New("server busy, please try again.  maximum pending requests exceeded")
Daniel Hiltgen's avatar
Daniel Hiltgen committed
63
64

func InitScheduler(ctx context.Context) *Scheduler {
Michael Yang's avatar
int  
Michael Yang committed
65
	maxQueue := envconfig.MaxQueue()
Daniel Hiltgen's avatar
Daniel Hiltgen committed
66
	sched := &Scheduler{
Michael Yang's avatar
int  
Michael Yang committed
67
68
69
		pendingReqCh:  make(chan *LlmRequest, maxQueue),
		finishedReqCh: make(chan *LlmRequest, maxQueue),
		expiredCh:     make(chan *runnerRef, maxQueue),
70
		unloadedCh:    make(chan any, maxQueue),
Daniel Hiltgen's avatar
Daniel Hiltgen committed
71
72
		loaded:        make(map[string]*runnerRef),
		newServerFn:   llm.NewLlamaServer,
73
74
		getGpuFn:      discover.GetGPUInfo,
		getCpuFn:      discover.GetCPUInfo,
75
		reschedDelay:  250 * time.Millisecond,
Daniel Hiltgen's avatar
Daniel Hiltgen committed
76
77
78
79
80
81
	}
	sched.loadFn = sched.load
	return sched
}

// context must be canceled to decrement ref count and release the runner
82
func (s *Scheduler) GetRunner(c context.Context, model *Model, opts api.Options, sessionDuration *api.Duration) (chan *runnerRef, chan error) {
83
84
85
86
	if opts.NumCtx < 4 {
		opts.NumCtx = 4
	}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
87
88
89
90
91
92
93
94
	req := &LlmRequest{
		ctx:             c,
		model:           model,
		opts:            opts,
		sessionDuration: sessionDuration,
		successCh:       make(chan *runnerRef),
		errCh:           make(chan error, 1),
	}
95

Daniel Hiltgen's avatar
Daniel Hiltgen committed
96
97
98
	select {
	case s.pendingReqCh <- req:
	default:
99
		req.errCh <- ErrMaxQueue
Daniel Hiltgen's avatar
Daniel Hiltgen committed
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
	}
	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
124
			pending.schedAttempts++
Daniel Hiltgen's avatar
Daniel Hiltgen committed
125
126
			if pending.origNumCtx == 0 {
				pending.origNumCtx = pending.opts.NumCtx
Daniel Hiltgen's avatar
Daniel Hiltgen committed
127
			}
128
129
130
131
132

			if pending.ctx.Err() != nil {
				slog.Debug("pending request cancelled or timed out, skipping scheduling")
				continue
			}
Michael Yang's avatar
Michael Yang committed
133
			numParallel := int(envconfig.NumParallel())
134
135
136
			// `mllama` is a snowflake and uses an encoder cache which cannot be used with num_parallel > 1
			// ref: https://github.com/ollama/ollama/issues/4165
			if slices.Contains(pending.model.Config.ModelFamilies, "mllama") && numParallel != 1 {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
137
				numParallel = 1
138
				slog.Warn("mllama does not currently support parallel requests")
Daniel Hiltgen's avatar
Daniel Hiltgen committed
139
			}
140

Daniel Hiltgen's avatar
Daniel Hiltgen committed
141
142
143
144
145
146
147
148
			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) {
149
						slog.Debug("reloading", "runner", runner)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
150
151
152
153
154
155
						runnerToExpire = runner
					} else {
						// Runner is usable, return it
						pending.useLoadedRunner(runner, s.finishedReqCh)
						break
					}
Michael Yang's avatar
Michael Yang committed
156
				} else if envconfig.MaxRunners() > 0 && loadedCount >= int(envconfig.MaxRunners()) {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
157
					slog.Debug("max runners achieved, unloading one to make room", "runner_count", loadedCount)
158
					runnerToExpire = s.findRunnerToUnload()
Daniel Hiltgen's avatar
Daniel Hiltgen committed
159
				} else {
160
					// Either no models are loaded or below envconfig.MaxRunners
Daniel Hiltgen's avatar
Daniel Hiltgen committed
161
					// Get a refreshed GPU list
162
					var gpus discover.GpuInfoList
163
164
165
166
167
					if pending.opts.NumGPU == 0 {
						gpus = s.getCpuFn()
					} else {
						gpus = s.getGpuFn()
					}
168

Michael Yang's avatar
int  
Michael Yang committed
169
					if envconfig.MaxRunners() <= 0 {
170
171
172
173
174
175
176
177
178
179
180
						// No user specified MaxRunners, so figure out what automatic setting to use
						// If all GPUs have reliable free memory reporting, defaultModelsPerGPU * the number of GPUs
						// if any GPU has unreliable free memory reporting, 1x the number of GPUs
						allReliable := true
						for _, gpu := range gpus {
							if gpu.UnreliableFreeMemory {
								allReliable = false
								break
							}
						}
						if allReliable {
Michael Yang's avatar
int  
Michael Yang committed
181
182
							// HACK
							os.Setenv("OLLAMA_MAX_LOADED_MODELS", strconv.Itoa(defaultModelsPerGPU*len(gpus)))
183
							slog.Debug("updating default concurrency", "OLLAMA_MAX_LOADED_MODELS", envconfig.MaxRunners(), "gpu_count", len(gpus))
184
						} else {
Michael Yang's avatar
int  
Michael Yang committed
185
186
							// HACK
							os.Setenv("OLLAMA_MAX_LOADED_MODELS", strconv.Itoa(len(gpus)))
187
188
189
							slog.Info("one or more GPUs detected that are unable to accurately report free memory - disabling default concurrency")
						}
					}
190

191
					// Load model for fitting
192
					ggml, err := llm.LoadModel(pending.model.ModelPath, 1024)
193
194
195
196
					if err != nil {
						pending.errCh <- err
						break
					}
197

198
					// Embedding models should always be loaded with parallel=1
199
					if pending.model.CheckCapabilities(model.CapabilityCompletion) != nil {
200
201
202
						numParallel = 1
					}

203
204
					// 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" {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
205
206
207
208
209
						// simplifying assumption of defaultParallel when in CPU mode
						if numParallel <= 0 {
							numParallel = defaultParallel
						}

210
211
						pending.opts.NumCtx = pending.origNumCtx * numParallel

212
213
						if loadedCount == 0 {
							slog.Debug("cpu mode with first model, loading")
Daniel Hiltgen's avatar
Daniel Hiltgen committed
214
							s.loadFn(pending, ggml, gpus, numParallel)
215
216
217
218
219
							break
						}
						runnerToExpire = s.maybeFindCPURunnerToUnload(pending, ggml, gpus)
						if runnerToExpire == nil {
							slog.Debug("cpu mode with available system memory or first model, loading")
Daniel Hiltgen's avatar
Daniel Hiltgen committed
220
							s.loadFn(pending, ggml, gpus, numParallel)
221
222
223
224
225
							break
						}
						// else we need to expire a runner
					} else if loadedCount == 0 {
						// No models loaded. Load the model but prefer the best fit.
226
						slog.Debug("loading first model", "model", pending.model.ModelPath)
227
						g := pickBestFullFitByLibrary(pending, ggml, gpus, &numParallel)
228
229
						if g != nil {
							gpus = g
230
231
232
						} else {
							// Only allow partial loads when this is the first model
							gpus = pickBestPartialFitByLibrary(pending, ggml, gpus, &numParallel)
233
						}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
234
						s.loadFn(pending, ggml, gpus, numParallel)
235
236
237
						break
					}

238
					if runnerToExpire == nil {
239
240
241
242
243
244
						// 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
Daniel Hiltgen's avatar
Daniel Hiltgen committed
245
						availGpus := s.filterGPUsWithoutLoadingModels(gpus)
246

247
						// Update free memory from currently loaded models
248
						s.updateFreeSpace(availGpus)
249
						fitGpus := pickBestFullFitByLibrary(pending, ggml, availGpus, &numParallel)
250
						if fitGpus != nil {
251
							slog.Debug("new model fits with existing models, loading")
Daniel Hiltgen's avatar
Daniel Hiltgen committed
252
							s.loadFn(pending, ggml, fitGpus, numParallel)
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
							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
							}()
272
273
274
							break
						}
						runnerToExpire = s.findRunnerToUnload()
Daniel Hiltgen's avatar
Daniel Hiltgen committed
275
276
277
278
279
280
281
282
283
284
					}
				}

				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()
285
				slog.Debug("resetting model to expire immediately to make room", "runner", runnerToExpire, "refCount", runnerToExpire.refCount)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
286
287
288
289
290
291
292
293
294
295
296
297
				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.
298
				slog.Debug("waiting for pending requests to complete and unload to occur", "runner", runnerToExpire)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
299
300
301
302
303
				select {
				case <-ctx.Done():
					slog.Debug("shutting down scheduler pending loop")
					return
				case <-s.unloadedCh:
304
					slog.Debug("unload completed", "runner", runnerToExpire)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
					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 {
327
				slog.Error("finished request signal received after model unloaded", "modelPath", finished.model.ModelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
328
329
330
331
332
333
				continue
			}
			runner.refMu.Lock()
			runner.refCount--
			if runner.refCount <= 0 {
				if runner.sessionDuration <= 0 {
334
					slog.Debug("runner with zero duration has gone idle, expiring to unload", "runner", runner)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
335
336
337
338
339
340
					if runner.expireTimer != nil {
						runner.expireTimer.Stop()
						runner.expireTimer = nil
					}
					s.expiredCh <- runner
				} else if runner.expireTimer == nil {
341
					slog.Debug("runner with non-zero duration has gone idle, adding timer", "runner", runner, "duration", runner.sessionDuration)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
342
					runner.expireTimer = time.AfterFunc(runner.sessionDuration, func() {
343
						slog.Debug("timer expired, expiring to unload", "runner", runner)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
344
345
346
347
						runner.refMu.Lock()
						defer runner.refMu.Unlock()
						if runner.expireTimer != nil {
							runner.expireTimer.Stop()
348
							runner.expireTimer = nil
Daniel Hiltgen's avatar
Daniel Hiltgen committed
349
350
351
						}
						s.expiredCh <- runner
					})
352
					runner.expiresAt = time.Now().Add(runner.sessionDuration)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
353
				} else {
354
					slog.Debug("runner with non-zero duration has gone idle, resetting timer", "runner", runner, "duration", runner.sessionDuration)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
355
					runner.expireTimer.Reset(runner.sessionDuration)
356
					runner.expiresAt = time.Now().Add(runner.sessionDuration)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
357
358
				}
			}
359
			slog.Debug("after processing request finished event", "runner", runner, "refCount", runner.refCount)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
360
361
			runner.refMu.Unlock()
		case runner := <-s.expiredCh:
362
			slog.Debug("runner expired event received", "runner", runner)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
363
364
			runner.refMu.Lock()
			if runner.refCount > 0 {
365
				slog.Debug("expired event with positive ref count, retrying", "runner", runner, "refCount", runner.refCount)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
366
367
368
369
370
371
372
373
374
375
				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
			}

376
			s.loadedMu.Lock()
377
378
379
380
381
382
383
384
385
386
387
			slog.Debug("got lock to unload expired event", "runner", runner)
			runnerToUnload := s.loaded[runner.modelPath]
			if runnerToUnload == nil {
				// If runnerToUnload is nil, we already processed an event and
				// unloaded it. This double unload can happen if the initial
				// request is canceled and we're trying to load another model
				// that requires this one to be evicted, or the settings change
				// and require a reload
				s.loadedMu.Unlock()
				runner.refMu.Unlock()
				slog.Debug("duplicate expired event, ignoring", "runner", runner)
388
389
390
391
392
393
394
395
396
397
398
			} else if runner.pid != runnerToUnload.pid {
				// If the pids do not match, we likely had multiple load
				// failures for the same model in quick succession due to
				// request context canceled and are draining the queue of
				// events. Ensure the orphaned runner is properly shut down, but
				// do not delete the mismatched loaded runner, or wait for VRAM
				// convergence.
				slog.Debug("orphaned runner shutting down", "orphan", runner, "loaded", runnerToUnload)
				runner.unload()
				s.loadedMu.Unlock()
				runner.refMu.Unlock()
399
400
401
402
403
404
405
406
407
408
409
410
			} else {
				slog.Debug("starting background wait for VRAM recovery", "runner", runner)
				finished := runner.waitForVRAMRecovery()
				runner.unload()
				delete(s.loaded, runner.modelPath)
				s.loadedMu.Unlock()
				slog.Debug("runner terminated and removed from list, blocking for VRAM recovery", "runner", runner)
				<-finished
				runner.refMu.Unlock()
				slog.Debug("sending an unloaded event", "runner", runner)
				s.unloadedCh <- struct{}{}
			}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
411
412
413
414
415
416
417
418
419
420
421
		}
	}
}

// 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++
422
423
424
425
	if runner.expireTimer != nil {
		runner.expireTimer.Stop()
		runner.expireTimer = nil
	}
426
427
428
	if pending.sessionDuration != nil {
		runner.sessionDuration = pending.sessionDuration.Duration
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
429
430
431
	pending.successCh <- runner
	go func() {
		<-pending.ctx.Done()
432
		slog.Debug("context for request finished", "runner", runner)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
433
434
435
436
		finished <- pending
	}()
}

Michael Yang's avatar
Michael Yang committed
437
func (s *Scheduler) load(req *LlmRequest, f *ggml.GGML, gpus discover.GpuInfoList, numParallel int) {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
438
439
440
	if numParallel < 1 {
		numParallel = 1
	}
Michael Yang's avatar
Michael Yang committed
441
	sessionDuration := envconfig.KeepAlive()
442
443
444
	if req.sessionDuration != nil {
		sessionDuration = req.sessionDuration.Duration
	}
Michael Yang's avatar
Michael Yang committed
445
	llama, err := s.newServerFn(gpus, req.model.ModelPath, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, numParallel)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
446
447
448
449
	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
Michael Yang's avatar
Michael Yang committed
450
		if errors.Is(err, ggml.ErrUnsupportedFormat) || strings.Contains(err.Error(), "failed to load model") {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
451
452
453
454
455
456
			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
	}
457
458
459
460
461
	runner := &runnerRef{
		model:           req.model,
		modelPath:       req.model.ModelPath,
		llama:           llama,
		Options:         &req.opts,
462
		sessionDuration: sessionDuration,
463
464
465
466
		gpus:            gpus,
		estimatedVRAM:   llama.EstimatedVRAM(),
		estimatedTotal:  llama.EstimatedTotal(),
		loading:         true,
467
		pid:             llama.Pid(),
468
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
469
	runner.numParallel = numParallel
470
	runner.refMu.Lock() // hold lock until running or aborted
471

Daniel Hiltgen's avatar
Daniel Hiltgen committed
472
	s.loadedMu.Lock()
473
474
475
476
477
478
479
	if oldRunner, ok := s.loaded[req.model.ModelPath]; ok {
		// Shouldn't happen, but safeguard against leaking a runner
		slog.Warn("model was still loaded", "old_runner", oldRunner, "new_runner", runner)
		oldRunner.refMu.Lock()
		oldRunner.unload()
		oldRunner.refMu.Unlock()
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
480
481
482
483
484
485
486
487
488
	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)
			req.errCh <- err
489
			slog.Debug("triggering expiration for failed load", "runner", runner)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
490
491
492
			s.expiredCh <- runner
			return
		}
493
		slog.Debug("finished setting up", "runner", runner)
494
495
496
		if runner.pid < 0 {
			runner.pid = llama.Pid()
		}
497
		runner.refCount++
Daniel Hiltgen's avatar
Daniel Hiltgen committed
498
499
500
501
502
503
504
505
506
507
		runner.loading = false
		go func() {
			<-req.ctx.Done()
			slog.Debug("context for request finished")
			s.finishedReqCh <- req
		}()
		req.successCh <- runner
	}()
}

508
func (s *Scheduler) updateFreeSpace(allGpus discover.GpuInfoList) {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
509
510
511
512
513
514
	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()
515
	runners := make([]*runnerRef, 0, len(s.loaded))
Daniel Hiltgen's avatar
Daniel Hiltgen committed
516
	for _, r := range s.loaded {
517
518
519
520
		runners = append(runners, r)
	}
	s.loadedMu.Unlock()
	for _, r := range runners {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
521
522
523
		r.refMu.Lock()
		if r.llama != nil {
			for _, gpu := range allGpus {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
524
				predMap[predKey{gpu.Library, gpu.ID}] += r.llama.EstimatedVRAMByGPU(gpu.ID)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
			}
		} else {
			slog.Warn("unexpected nil runner reference, memory prediction may be incorrect")
		}
		r.refMu.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
543
				// after we start our first runner, then we'll never account for that, so picking the smallest free value seems prudent.
Daniel Hiltgen's avatar
Daniel Hiltgen committed
544
545
				allGpus[i].FreeMemory = allGpus[i].TotalMemory - p
			}
546
547
548
549
550
551
552
553
554
			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)
555
556
func (s *Scheduler) filterGPUsWithoutLoadingModels(allGpus discover.GpuInfoList) discover.GpuInfoList {
	ret := append(discover.GpuInfoList{}, allGpus...)
557
558
559
560
561
562
563
564
565
566
567
568
569
	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
570
571
		}
	}
572
	return ret
Daniel Hiltgen's avatar
Daniel Hiltgen committed
573
574
}

575
// TODO consolidate sched_types.go
Daniel Hiltgen's avatar
Daniel Hiltgen committed
576
type runnerRef struct {
577
	refMu    sync.Mutex
Daniel Hiltgen's avatar
Daniel Hiltgen committed
578
579
	refCount uint // prevent unloading if > 0

580
	llama          llm.LlamaServer
581
	pid            int
582
583
	loading        bool                 // True only during initial load, then false forever
	gpus           discover.GpuInfoList // Recorded at time of provisioning
584
585
	estimatedVRAM  uint64
	estimatedTotal uint64
Daniel Hiltgen's avatar
Daniel Hiltgen committed
586
587
588

	sessionDuration time.Duration
	expireTimer     *time.Timer
589
	expiresAt       time.Time
Daniel Hiltgen's avatar
Daniel Hiltgen committed
590

Daniel Hiltgen's avatar
Daniel Hiltgen committed
591
592
593
	model       *Model
	modelPath   string
	numParallel int
Daniel Hiltgen's avatar
Daniel Hiltgen committed
594
595
596
597
598
	*api.Options
}

// The refMu must already be held when calling unload
func (runner *runnerRef) unload() {
599
600
601
602
	if runner.expireTimer != nil {
		runner.expireTimer.Stop()
		runner.expireTimer = nil
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
603
604
605
	if runner.llama != nil {
		runner.llama.Close()
	}
606
	runner.model = nil
Daniel Hiltgen's avatar
Daniel Hiltgen committed
607
608
609
610
611
612
613
614
615
	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()
616

Daniel Hiltgen's avatar
Daniel Hiltgen committed
617
618
619
620
	timeout := 10 * time.Second
	if runner.loading {
		timeout = 2 * time.Minute // Initial load can take a long time for big models on slow systems...
	}
621

622
623
624
625
	if runner.Options == nil {
		return true
	}

626
627
628
629
630
631
632
633
	// 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
	}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
634
635
636
	// Normalize the NumCtx for parallelism
	optsExisting.NumCtx = optsExisting.NumCtx / runner.numParallel

637
	ctx, cancel := context.WithTimeout(ctx, timeout)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
638
	defer cancel()
639
640
	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
641
642
643
644
		!reflect.DeepEqual(optsExisting, optsNew) || // have the runner options changed?
		runner.llama.Ping(ctx) != nil {
		return true
	}
645

Daniel Hiltgen's avatar
Daniel Hiltgen committed
646
647
648
	return false
}

649
650
651
652
653
654
655
656
// 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
657
658
func (runner *runnerRef) waitForVRAMRecovery() chan any {
	finished := make(chan any, 1)
659

660
661
	// CPU or Metal don't need checking, so no waiting required
	// windows can page VRAM, only cuda currently can report accurate used vram usage
662
663
	if len(runner.gpus) == 0 ||
		(len(runner.gpus) == 1 && (runner.gpus[0].Library == "cpu" || runner.gpus[0].Library == "metal")) ||
664
		(runtime.GOOS == "windows" && runner.gpus[0].Library != "cuda") {
665
		finished <- struct{}{}
666
		slog.Debug("no need to wait for VRAM recovery", "runner", runner)
667
668
669
670
671
		return finished
	}
	start := time.Now()

	// Establish a baseline before we unload
672
	gpusBefore := discover.GetGPUInfo()
673
674
675
676
677
678
679
680
681
682
683
684
	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) {
685
				slog.Warn("gpu VRAM usage didn't recover within timeout", "seconds", time.Since(start).Seconds(), "runner", runner)
686
687
688
689
				finished <- struct{}{}
			}

			// Query GPUs, look for free to go back up
690
			gpusNow := discover.GetGPUInfo()
691
692
693
694
695
696
697
			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 {
698
				slog.Debug(fmt.Sprintf("gpu VRAM free memory converged after %0.2f seconds", time.Since(start).Seconds()), "runner", runner)
699
700
701
702
703
704
705
706
				finished <- struct{}{}
				return
			}
		}
	}()
	return finished
}

707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
func (runner *runnerRef) LogValue() slog.Value {
	if runner == nil {
		return slog.StringValue("nil")
	}
	attrs := []slog.Attr{}
	if runner.model != nil {
		attrs = append(attrs, slog.String("name", runner.model.Name))
	}
	if len(runner.gpus) > 0 {
		attrs = append(attrs,
			slog.String("inference", runner.gpus[0].Library),
			slog.Int("devices", len(runner.gpus)),
		)
	}
	attrs = append(attrs,
		slog.String("size", format.HumanBytes2(runner.estimatedTotal)),
		slog.String("vram", format.HumanBytes2(runner.estimatedVRAM)),
		slog.Int("parallel", runner.numParallel),
		slog.Int("pid", runner.pid),
		slog.String("model", runner.modelPath),
	)
728
729
730
	if runner.Options != nil {
		attrs = append(attrs, slog.Int("num_ctx", runner.Options.NumCtx))
	}
731
732
733
	return slog.GroupValue(attrs...)
}

734
735
736
737
738
739
740
741
742
743
744
745
746
type ByDurationAndName []*runnerRef

func (a ByDurationAndName) Len() int      { return len(a) }
func (a ByDurationAndName) Swap(i, j int) { a[i], a[j] = a[j], a[i] }
func (a ByDurationAndName) Less(i, j int) bool {
	// Primary sort by session duration (uint64 to handle negatives)
	d1 := uint64(a[i].sessionDuration)
	d2 := uint64(a[j].sessionDuration)
	if d1 != d2 {
		return d1 < d2
	}
	// Secondary sort by model path lex order
	return a[i].modelPath < a[j].modelPath
Daniel Hiltgen's avatar
Daniel Hiltgen committed
747
748
749
750
751
752
753
754
}

// 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 }

755
756
// pickBestFullFitByLibrary will try to find the optimal placement of the model in the available GPUs where the model fully fits
// The list of GPUs returned will always be the same brand (library)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
757
// If the model can not be fit fully within the available GPU(s) nil is returned
758
// If numParallel is <= 0, this will attempt try to optimize parallelism based on available VRAM, and adjust
Daniel Hiltgen's avatar
Daniel Hiltgen committed
759
// opts.NumCtx accordingly
Michael Yang's avatar
Michael Yang committed
760
func pickBestFullFitByLibrary(req *LlmRequest, f *ggml.GGML, gpus discover.GpuInfoList, numParallel *int) discover.GpuInfoList {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
761
762
763
	var numParallelToTry []int
	if *numParallel <= 0 {
		// If no specific parallel setting was provided, try larger then smaller, always end with 1
764
		numParallelToTry = append(numParallelToTry, defaultParallel, 1)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
765
766
767
768
	} else {
		numParallelToTry = []int{*numParallel}
	}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
769
	for _, gl := range gpus.ByLibrary() {
770
		sgl := append(make(discover.GpuInfoList, 0, len(gl)), gl...)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
771
772

		// TODO - potentially sort by performance capability, existing models loaded, etc.
Daniel Hiltgen's avatar
Daniel Hiltgen committed
773
		// TODO - Eliminate any GPUs that already have envconfig.MaxRunners loaded on them
774
		// Note: at present, this will favor most current available VRAM descending and ignoring faster GPU speed in mixed setups
775
		sort.Sort(sort.Reverse(discover.ByFreeMemory(sgl)))
Daniel Hiltgen's avatar
Daniel Hiltgen committed
776

777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
		if !envconfig.SchedSpread() {
			for _, p := range numParallelToTry {
				req.opts.NumCtx = req.origNumCtx * p
				// Try to pack into as few GPUs as possible, starting from 1 GPU
				for numGPUs := 1; numGPUs <= len(sgl); numGPUs++ {
					gpuSubset := sgl[:numGPUs]
					ok, estimatedVRAM := llm.PredictServerFit(gpuSubset, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, p)

					if ok {
						slog.Info("new model will fit in available VRAM across minimum required GPUs, loading",
							"model", req.model.ModelPath,
							"library", sgl[0].Library,
							"parallel", p,
							"required", format.HumanBytes2(estimatedVRAM),
							"gpus", numGPUs)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
792
						*numParallel = p
793
						return gpuSubset
Daniel Hiltgen's avatar
Daniel Hiltgen committed
794
					}
795
				}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
796
			}
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
		} else {
			// 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 (OLLAMA_SCHED_SPREAD is set)
			for _, p := range numParallelToTry {
				req.opts.NumCtx = req.origNumCtx * p
				if ok, estimatedVRAM := llm.PredictServerFit(sgl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, p); ok {
					slog.Info("new model will fit in available VRAM, loading",
						"model", req.model.ModelPath,
						"library", sgl[0].Library,
						"parallel", p,
						"required", format.HumanBytes2(estimatedVRAM),
						"gpus", len(sgl))
					*numParallel = p
					return sgl
				}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
815
			}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
816
817
818
819
820
		}
	}
	return nil
}

821
// If multiple Libraries are detected, pick the Library which loads the most layers for the model
Michael Yang's avatar
Michael Yang committed
822
func pickBestPartialFitByLibrary(req *LlmRequest, f *ggml.GGML, gpus discover.GpuInfoList, numParallel *int) discover.GpuInfoList {
823
824
	if *numParallel <= 0 {
		*numParallel = 1
Richard Lyons's avatar
Richard Lyons committed
825
826
		req.opts.NumCtx = req.origNumCtx
	}
827
828
829
830
831
832
833
	byLibrary := gpus.ByLibrary()
	if len(byLibrary) <= 1 {
		return gpus
	}
	var bestEstimate uint64
	var bestFit int
	for i, gl := range byLibrary {
834
		_, estimatedVRAM := llm.PredictServerFit(gl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, *numParallel)
835
836
837
838
839
840
841
842
		if estimatedVRAM > bestEstimate {
			bestEstimate = estimatedVRAM
			bestFit = i
		}
	}
	return byLibrary[bestFit]
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
843
// findRunnerToUnload finds a runner to unload to make room for a new model
844
func (s *Scheduler) findRunnerToUnload() *runnerRef {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
845
846
847
848
849
850
	s.loadedMu.Lock()
	runnerList := make([]*runnerRef, 0, len(s.loaded))
	for _, r := range s.loaded {
		runnerList = append(runnerList, r)
	}
	s.loadedMu.Unlock()
851
852
853
854
	if len(runnerList) == 0 {
		slog.Debug("no loaded runner to unload")
		return nil
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
855
856
857

	// 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?
858
	sort.Sort(ByDurationAndName(runnerList))
Daniel Hiltgen's avatar
Daniel Hiltgen committed
859
860
861
862
863
864
865

	// 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 {
866
			slog.Debug("found an idle runner to unload", "runner", runner)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
867
868
869
870
			return runner
		}
	}
	// None appear idle, just wait for the one with the shortest duration
871
	slog.Debug("no idle runners, picking the shortest duration", "runner_count", len(runnerList), "runner", runnerList[0])
Daniel Hiltgen's avatar
Daniel Hiltgen committed
872
873
874
875
876
877
878
879
880
881
882
883
884
	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()
		}
	}
}
885

Patrick Devine's avatar
Patrick Devine committed
886
887
888
func (s *Scheduler) expireRunner(model *Model) {
	s.loadedMu.Lock()
	runner, ok := s.loaded[model.ModelPath]
889
	s.loadedMu.Unlock()
Patrick Devine's avatar
Patrick Devine committed
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
	if ok {
		runner.refMu.Lock()
		runner.expiresAt = time.Now()
		if runner.expireTimer != nil {
			runner.expireTimer.Stop()
			runner.expireTimer = nil
		}
		runner.sessionDuration = 0
		if runner.refCount <= 0 {
			s.expiredCh <- runner
		}
		runner.refMu.Unlock()
	}
}

905
906
// 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
Michael Yang's avatar
Michael Yang committed
907
func (s *Scheduler) maybeFindCPURunnerToUnload(req *LlmRequest, f *ggml.GGML, gpus discover.GpuInfoList) *runnerRef {
908
	slog.Debug("evaluating if CPU model load will fit in available system memory")
909
	estimate := llm.EstimateGPULayers(gpus, f, req.model.ProjectorPaths, req.opts, req.opts.NumCtx/req.origNumCtx)
910
911
912
913
914
915
916
917
918
	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()
}