sched.go 28.4 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"
Daniel Hiltgen's avatar
Daniel Hiltgen committed
11
	"sort"
Michael Yang's avatar
int  
Michael Yang committed
12
	"strconv"
Daniel Hiltgen's avatar
Daniel Hiltgen committed
13
14
15
16
17
	"strings"
	"sync"
	"time"

	"github.com/ollama/ollama/api"
18
	"github.com/ollama/ollama/envconfig"
Daniel Hiltgen's avatar
Daniel Hiltgen committed
19
20
21
22
23
24
25
26
27
	"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
Daniel Hiltgen's avatar
Daniel Hiltgen committed
28
	origNumCtx      int // Track the initial ctx request
29
	sessionDuration *api.Duration
Daniel Hiltgen's avatar
Daniel Hiltgen committed
30
31
	successCh       chan *runnerRef
	errCh           chan error
32
	schedAttempts   uint
Daniel Hiltgen's avatar
Daniel Hiltgen committed
33
34
35
36
37
38
39
40
41
42
43
}

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

	loaded   map[string]*runnerRef
	loadedMu sync.Mutex

Daniel Hiltgen's avatar
Daniel Hiltgen committed
44
45
	loadFn       func(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel int)
	newServerFn  func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error)
46
47
48
	getGpuFn     func() gpu.GpuInfoList
	getCpuFn     func() gpu.GpuInfoList
	reschedDelay time.Duration
Daniel Hiltgen's avatar
Daniel Hiltgen committed
49
50
}

51
52
53
54
55
56
57
58
59
60
// 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
// Model will still need to fit in VRAM.  If this setting wont fit
// we'll back off down to 1 to try to get it to fit
var defaultParallel = 4

Michael Yang's avatar
lint  
Michael Yang committed
61
var ErrMaxQueue = errors.New("server busy, please try again.  maximum pending requests exceeded")
Daniel Hiltgen's avatar
Daniel Hiltgen committed
62
63

func InitScheduler(ctx context.Context) *Scheduler {
Michael Yang's avatar
int  
Michael Yang committed
64
	maxQueue := envconfig.MaxQueue()
Daniel Hiltgen's avatar
Daniel Hiltgen committed
65
	sched := &Scheduler{
Michael Yang's avatar
int  
Michael Yang committed
66
67
68
69
		pendingReqCh:  make(chan *LlmRequest, maxQueue),
		finishedReqCh: make(chan *LlmRequest, maxQueue),
		expiredCh:     make(chan *runnerRef, maxQueue),
		unloadedCh:    make(chan interface{}, maxQueue),
Daniel Hiltgen's avatar
Daniel Hiltgen committed
70
71
72
		loaded:        make(map[string]*runnerRef),
		newServerFn:   llm.NewLlamaServer,
		getGpuFn:      gpu.GetGPUInfo,
73
		getCpuFn:      gpu.GetCPUInfo,
74
		reschedDelay:  250 * time.Millisecond,
Daniel Hiltgen's avatar
Daniel Hiltgen committed
75
76
77
78
79
80
	}
	sched.loadFn = sched.load
	return sched
}

// context must be canceled to decrement ref count and release the runner
81
func (s *Scheduler) GetRunner(c context.Context, model *Model, opts api.Options, sessionDuration *api.Duration) (chan *runnerRef, chan error) {
82
83
84
85
	if opts.NumCtx < 4 {
		opts.NumCtx = 4
	}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
86
87
88
89
90
91
92
93
	req := &LlmRequest{
		ctx:             c,
		model:           model,
		opts:            opts,
		sessionDuration: sessionDuration,
		successCh:       make(chan *runnerRef),
		errCh:           make(chan error, 1),
	}
94

Daniel Hiltgen's avatar
Daniel Hiltgen committed
95
96
97
	select {
	case s.pendingReqCh <- req:
	default:
98
		req.errCh <- ErrMaxQueue
Daniel Hiltgen's avatar
Daniel Hiltgen committed
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
	}
	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
123
			pending.schedAttempts++
Daniel Hiltgen's avatar
Daniel Hiltgen committed
124
125
			if pending.origNumCtx == 0 {
				pending.origNumCtx = pending.opts.NumCtx
Daniel Hiltgen's avatar
Daniel Hiltgen committed
126
			}
127
128
129
130
131

			if pending.ctx.Err() != nil {
				slog.Debug("pending request cancelled or timed out, skipping scheduling")
				continue
			}
Michael Yang's avatar
Michael Yang committed
132
			numParallel := int(envconfig.NumParallel())
Daniel Hiltgen's avatar
Daniel Hiltgen committed
133
134
135
136
137
138
			// TODO (jmorganca): multimodal models don't support parallel yet
			// see https://github.com/ollama/ollama/issues/4165
			if len(pending.model.ProjectorPaths) > 0 && numParallel != 1 {
				numParallel = 1
				slog.Warn("multimodal models don't support parallel requests yet")
			}
139

Daniel Hiltgen's avatar
Daniel Hiltgen committed
140
141
142
143
144
145
146
147
148
149
150
151
152
153
			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
					}
Michael Yang's avatar
Michael Yang committed
154
				} else if envconfig.MaxRunners() > 0 && loadedCount >= int(envconfig.MaxRunners()) {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
155
					slog.Debug("max runners achieved, unloading one to make room", "runner_count", loadedCount)
156
					runnerToExpire = s.findRunnerToUnload()
Daniel Hiltgen's avatar
Daniel Hiltgen committed
157
				} else {
158
					// Either no models are loaded or below envconfig.MaxRunners
Daniel Hiltgen's avatar
Daniel Hiltgen committed
159
					// Get a refreshed GPU list
160
161
162
163
164
165
					var gpus gpu.GpuInfoList
					if pending.opts.NumGPU == 0 {
						gpus = s.getCpuFn()
					} else {
						gpus = s.getGpuFn()
					}
166

Michael Yang's avatar
int  
Michael Yang committed
167
					if envconfig.MaxRunners() <= 0 {
168
169
170
171
172
173
174
175
176
177
178
						// 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
179
180
							// HACK
							os.Setenv("OLLAMA_MAX_LOADED_MODELS", strconv.Itoa(defaultModelsPerGPU*len(gpus)))
181
182
							slog.Debug("updating default concurrency", "OLLAMA_MAX_LOADED_MODELS", envconfig.MaxRunners, "gpu_count", len(gpus))
						} else {
Michael Yang's avatar
int  
Michael Yang committed
183
184
							// HACK
							os.Setenv("OLLAMA_MAX_LOADED_MODELS", strconv.Itoa(len(gpus)))
185
186
187
							slog.Info("one or more GPUs detected that are unable to accurately report free memory - disabling default concurrency")
						}
					}
188

189
					// Load model for fitting
190
					ggml, err := llm.LoadModel(pending.model.ModelPath, 0)
191
192
193
194
					if err != nil {
						pending.errCh <- err
						break
					}
195

196
197
					// 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
198
199
200
201
202
						// simplifying assumption of defaultParallel when in CPU mode
						if numParallel <= 0 {
							numParallel = defaultParallel
						}

203
204
						pending.opts.NumCtx = pending.origNumCtx * numParallel

205
206
						if loadedCount == 0 {
							slog.Debug("cpu mode with first model, loading")
Daniel Hiltgen's avatar
Daniel Hiltgen committed
207
							s.loadFn(pending, ggml, gpus, numParallel)
208
209
210
211
212
							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
213
							s.loadFn(pending, ggml, gpus, numParallel)
214
215
216
217
218
							break
						}
						// else we need to expire a runner
					} else if loadedCount == 0 {
						// No models loaded. Load the model but prefer the best fit.
219
						slog.Debug("loading first model", "model", pending.model.ModelPath)
220
						g := pickBestFullFitByLibrary(pending, ggml, gpus, &numParallel)
221
222
						if g != nil {
							gpus = g
223
224
225
						} else {
							// Only allow partial loads when this is the first model
							gpus = pickBestPartialFitByLibrary(pending, ggml, gpus, &numParallel)
226
						}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
227
						s.loadFn(pending, ggml, gpus, numParallel)
228
229
230
						break
					}

231
					if runnerToExpire == nil {
232
233
234
235
236
237
						// 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
238
						availGpus := s.filterGPUsWithoutLoadingModels(gpus)
239

240
						// Update free memory from currently loaded models
241
						s.updateFreeSpace(availGpus)
242
						fitGpus := pickBestFullFitByLibrary(pending, ggml, availGpus, &numParallel)
243
						if fitGpus != nil {
244
							slog.Debug("new model fits with existing models, loading")
Daniel Hiltgen's avatar
Daniel Hiltgen committed
245
							s.loadFn(pending, ggml, fitGpus, numParallel)
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
							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
							}()
265
266
267
							break
						}
						runnerToExpire = s.findRunnerToUnload()
Daniel Hiltgen's avatar
Daniel Hiltgen committed
268
269
270
271
272
273
274
275
276
277
					}
				}

				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()
278
				slog.Debug("resetting model to expire immediately to make room", "modelPath", runnerToExpire.modelPath, "refCount", runnerToExpire.refCount)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
279
280
281
282
283
284
285
286
287
288
289
290
				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.
291
				slog.Debug("waiting for pending requests to complete and unload to occur", "modelPath", runnerToExpire.modelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
292
293
294
295
296
				select {
				case <-ctx.Done():
					slog.Debug("shutting down scheduler pending loop")
					return
				case <-s.unloadedCh:
297
					slog.Debug("unload completed", "modelPath", runnerToExpire.modelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
					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 {
320
				slog.Error("finished request signal received after model unloaded", "modelPath", finished.model.ModelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
321
322
323
324
325
326
				continue
			}
			runner.refMu.Lock()
			runner.refCount--
			if runner.refCount <= 0 {
				if runner.sessionDuration <= 0 {
327
					slog.Debug("runner with zero duration has gone idle, expiring to unload", "modelPath", runner.modelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
328
329
330
331
332
333
					if runner.expireTimer != nil {
						runner.expireTimer.Stop()
						runner.expireTimer = nil
					}
					s.expiredCh <- runner
				} else if runner.expireTimer == nil {
334
					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
335
					runner.expireTimer = time.AfterFunc(runner.sessionDuration, func() {
336
						slog.Debug("timer expired, expiring to unload", "modelPath", runner.modelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
337
338
339
340
						runner.refMu.Lock()
						defer runner.refMu.Unlock()
						if runner.expireTimer != nil {
							runner.expireTimer.Stop()
341
							runner.expireTimer = nil
Daniel Hiltgen's avatar
Daniel Hiltgen committed
342
343
344
						}
						s.expiredCh <- runner
					})
345
					runner.expiresAt = time.Now().Add(runner.sessionDuration)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
346
				} else {
347
					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
348
					runner.expireTimer.Reset(runner.sessionDuration)
349
					runner.expiresAt = time.Now().Add(runner.sessionDuration)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
350
351
				}
			}
352
			slog.Debug("after processing request finished event", "modelPath", runner.modelPath, "refCount", runner.refCount)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
353
354
			runner.refMu.Unlock()
		case runner := <-s.expiredCh:
355
			slog.Debug("runner expired event received", "modelPath", runner.modelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
356
357
358
			runner.refMu.Lock()
			if runner.refCount > 0 {
				// Shouldn't happen, but safeguard to ensure no leaked runners
359
				slog.Debug("expired event with positive ref count, retrying", "modelPath", runner.modelPath, "refCount", runner.refCount)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
360
361
362
363
364
365
366
367
368
369
				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
			}

370
			s.loadedMu.Lock()
371
			slog.Debug("got lock to unload", "modelPath", runner.modelPath)
372
			finished := runner.waitForVRAMRecovery()
Daniel Hiltgen's avatar
Daniel Hiltgen committed
373
			runner.unload()
374
			delete(s.loaded, runner.modelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
375
			s.loadedMu.Unlock()
376
			slog.Debug("runner released", "modelPath", runner.modelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
377
			runner.refMu.Unlock()
378
379

			<-finished
380
			slog.Debug("sending an unloaded event", "modelPath", runner.modelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
381
382
383
384
385
386
387
388
389
390
391
392
			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++
393
394
395
396
	if runner.expireTimer != nil {
		runner.expireTimer.Stop()
		runner.expireTimer = nil
	}
397
398
399
	if pending.sessionDuration != nil {
		runner.sessionDuration = pending.sessionDuration.Duration
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
400
401
402
403
404
405
406
407
	pending.successCh <- runner
	go func() {
		<-pending.ctx.Done()
		slog.Debug("context for request finished")
		finished <- pending
	}()
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
408
409
410
411
func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel int) {
	if numParallel < 1 {
		numParallel = 1
	}
Michael Yang's avatar
Michael Yang committed
412
	sessionDuration := envconfig.KeepAlive()
413
414
415
	if req.sessionDuration != nil {
		sessionDuration = req.sessionDuration.Duration
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
416
	llama, err := s.newServerFn(gpus, req.model.ModelPath, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, numParallel)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
417
418
419
420
	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
lint  
Michael Yang committed
421
		if errors.Is(err, llm.ErrUnsupportedFormat) || strings.Contains(err.Error(), "failed to load model") {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
422
423
424
425
426
427
			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
	}
428
429
430
431
432
	runner := &runnerRef{
		model:           req.model,
		modelPath:       req.model.ModelPath,
		llama:           llama,
		Options:         &req.opts,
433
		sessionDuration: sessionDuration,
434
435
436
437
438
439
		gpus:            gpus,
		estimatedVRAM:   llama.EstimatedVRAM(),
		estimatedTotal:  llama.EstimatedTotal(),
		loading:         true,
		refCount:        1,
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
440
	runner.numParallel = numParallel
Daniel Hiltgen's avatar
Daniel Hiltgen committed
441
	runner.refMu.Lock()
442

Daniel Hiltgen's avatar
Daniel Hiltgen committed
443
444
445
446
447
448
449
450
451
452
453
	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
454
			slog.Debug("triggering expiration for failed load", "model", runner.modelPath)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
			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 {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
480
				predMap[predKey{gpu.Library, gpu.ID}] += r.llama.EstimatedVRAMByGPU(gpu.ID)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
			}
		} 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
			}
503
504
505
506
507
508
509
510
511
			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)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
512
func (s *Scheduler) filterGPUsWithoutLoadingModels(allGpus gpu.GpuInfoList) gpu.GpuInfoList {
513
514
515
516
517
518
519
520
521
522
523
524
525
526
	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
527
528
		}
	}
529
	return ret
Daniel Hiltgen's avatar
Daniel Hiltgen committed
530
531
}

532
// TODO consolidate sched_types.go
Daniel Hiltgen's avatar
Daniel Hiltgen committed
533
534
535
536
537
538
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

539
540
541
542
543
	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
544
545
546

	sessionDuration time.Duration
	expireTimer     *time.Timer
547
	expiresAt       time.Time
Daniel Hiltgen's avatar
Daniel Hiltgen committed
548

Daniel Hiltgen's avatar
Daniel Hiltgen committed
549
550
551
	model       *Model
	modelPath   string
	numParallel int
Daniel Hiltgen's avatar
Daniel Hiltgen committed
552
553
554
555
556
	*api.Options
}

// The refMu must already be held when calling unload
func (runner *runnerRef) unload() {
557
558
559
560
	if runner.expireTimer != nil {
		runner.expireTimer.Stop()
		runner.expireTimer = nil
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
561
562
563
	if runner.llama != nil {
		runner.llama.Close()
	}
564
	runner.model = nil
Daniel Hiltgen's avatar
Daniel Hiltgen committed
565
566
567
568
569
570
571
572
573
	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()
574

Daniel Hiltgen's avatar
Daniel Hiltgen committed
575
576
577
578
	timeout := 10 * time.Second
	if runner.loading {
		timeout = 2 * time.Minute // Initial load can take a long time for big models on slow systems...
	}
579

580
581
582
583
	if runner.Options == nil {
		return true
	}

584
585
586
587
588
589
590
591
	// 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
592
593
594
	// Normalize the NumCtx for parallelism
	optsExisting.NumCtx = optsExisting.NumCtx / runner.numParallel

595
	ctx, cancel := context.WithTimeout(ctx, timeout)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
596
	defer cancel()
597
598
	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
599
600
601
602
		!reflect.DeepEqual(optsExisting, optsNew) || // have the runner options changed?
		runner.llama.Ping(ctx) != nil {
		return true
	}
603

Daniel Hiltgen's avatar
Daniel Hiltgen committed
604
605
606
	return false
}

607
608
609
610
611
612
613
614
615
616
617
// 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)

618
619
	// CPU or Metal don't need checking, so no waiting required
	// windows can page VRAM, only cuda currently can report accurate used vram usage
620
621
	if len(runner.gpus) == 0 ||
		(len(runner.gpus) == 1 && (runner.gpus[0].Library == "cpu" || runner.gpus[0].Library == "metal")) ||
622
		(runtime.GOOS == "windows" && runner.gpus[0].Library != "cuda") {
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
		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) {
642
				slog.Warn("gpu VRAM usage didn't recover within timeout", "seconds", time.Since(start).Seconds(), "model", runner.modelPath)
643
644
645
646
647
648
649
650
651
652
653
654
				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 {
655
				slog.Debug(fmt.Sprintf("gpu VRAM free memory converged after %0.2f seconds", time.Since(start).Seconds()), "model", runner.modelPath)
656
657
658
659
660
661
662
663
				finished <- struct{}{}
				return
			}
		}
	}()
	return finished
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
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 }

679
680
// 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
681
// If the model can not be fit fully within the available GPU(s) nil is returned
Daniel Hiltgen's avatar
Daniel Hiltgen committed
682
683
// If numParallel is <= 0, this will attempt try to optimize parallism based on available VRAM, and adjust
// opts.NumCtx accordingly
684
func pickBestFullFitByLibrary(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel *int) gpu.GpuInfoList {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
685
	var estimatedVRAM uint64
Daniel Hiltgen's avatar
Daniel Hiltgen committed
686
687
688
689

	var numParallelToTry []int
	if *numParallel <= 0 {
		// If no specific parallel setting was provided, try larger then smaller, always end with 1
690
		numParallelToTry = append(numParallelToTry, defaultParallel, 1)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
691
692
693
694
	} else {
		numParallelToTry = []int{*numParallel}
	}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
695
696
697
698
699
	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.
Daniel Hiltgen's avatar
Daniel Hiltgen committed
700
		// TODO - Eliminate any GPUs that already have envconfig.MaxRunners loaded on them
Daniel Hiltgen's avatar
Daniel Hiltgen committed
701
702
703
704
		// 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
Daniel Hiltgen's avatar
Daniel Hiltgen committed
705
		for _, p := range numParallelToTry {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
706
			req.opts.NumCtx = req.origNumCtx * p
Michael Yang's avatar
bool  
Michael Yang committed
707
			if !envconfig.SchedSpread() {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
708
709
710
711
712
713
				for _, g := range sgl {
					if ok, estimatedVRAM = llm.PredictServerFit([]gpu.GpuInfo{g}, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
						slog.Info("new model will fit in available VRAM in single GPU, loading", "model", req.model.ModelPath, "gpu", g.ID, "parallel", p, "available", g.FreeMemory, "required", format.HumanBytes2(estimatedVRAM))
						*numParallel = p
						return []gpu.GpuInfo{g}
					}
714
				}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
715
716
717
718
719
720
721
722
			}
		}

		// 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
Daniel Hiltgen's avatar
Daniel Hiltgen committed
723
		for _, p := range numParallelToTry {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
724
			req.opts.NumCtx = req.origNumCtx * p
Daniel Hiltgen's avatar
Daniel Hiltgen committed
725
726
727
728
729
			if ok, estimatedVRAM = llm.PredictServerFit(sgl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); 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))
				*numParallel = p
				return sgl
			}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
730
731
732
733
734
		}
	}
	return nil
}

735
736
// If multiple Libraries are detected, pick the Library which loads the most layers for the model
func pickBestPartialFitByLibrary(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel *int) gpu.GpuInfoList {
737
738
	if *numParallel <= 0 {
		*numParallel = 1
Richard Lyons's avatar
Richard Lyons committed
739
                req.opts.NumCtx = req.origNumCtx
740
        }
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
	byLibrary := gpus.ByLibrary()
	if len(byLibrary) <= 1 {
		return gpus
	}
	var bestEstimate uint64
	var bestFit int
	for i, gl := range byLibrary {
		_, estimatedVRAM := llm.PredictServerFit(gl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts)
		if estimatedVRAM > bestEstimate {
			bestEstimate = estimatedVRAM
			bestFit = i
		}
	}
	return byLibrary[bestFit]
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
757
// findRunnerToUnload finds a runner to unload to make room for a new model
758
func (s *Scheduler) findRunnerToUnload() *runnerRef {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
759
760
761
762
763
764
	s.loadedMu.Lock()
	runnerList := make([]*runnerRef, 0, len(s.loaded))
	for _, r := range s.loaded {
		runnerList = append(runnerList, r)
	}
	s.loadedMu.Unlock()
765
766
767
768
	if len(runnerList) == 0 {
		slog.Debug("no loaded runner to unload")
		return nil
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798

	// 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()
		}
	}
}
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813

// 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()
}