server.go 53.9 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
package llm

import (
	"bufio"
	"bytes"
	"context"
	"encoding/json"
	"errors"
	"fmt"
	"io"
	"log"
	"log/slog"
	"math/rand"
	"net"
	"net/http"
	"os"
	"os/exec"
	"path/filepath"
	"runtime"
20
	"slices"
Jesse Gross's avatar
Jesse Gross committed
21
	"sort"
22
23
	"strconv"
	"strings"
24
	"sync"
25
26
	"time"

Daniel Hiltgen's avatar
Daniel Hiltgen committed
27
28
	"golang.org/x/sync/semaphore"

29
	"github.com/ollama/ollama/api"
30
	"github.com/ollama/ollama/discover"
31
	"github.com/ollama/ollama/envconfig"
32
	"github.com/ollama/ollama/format"
Michael Yang's avatar
Michael Yang committed
33
	"github.com/ollama/ollama/fs/ggml"
34
	"github.com/ollama/ollama/llama"
35
	"github.com/ollama/ollama/logutil"
Jesse Gross's avatar
Jesse Gross committed
36
	"github.com/ollama/ollama/ml"
37
	"github.com/ollama/ollama/model"
38
39
)

40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
type filteredEnv []string

func (e filteredEnv) LogValue() slog.Value {
	var attrs []slog.Attr
	for _, env := range e {
		if key, value, ok := strings.Cut(env, "="); ok {
			switch {
			case strings.HasPrefix(key, "OLLAMA_"),
				strings.HasPrefix(key, "CUDA_"),
				strings.HasPrefix(key, "ROCR_"),
				strings.HasPrefix(key, "ROCM_"),
				strings.HasPrefix(key, "HIP_"),
				strings.HasPrefix(key, "GPU_"),
				strings.HasPrefix(key, "HSA_"),
				strings.HasPrefix(key, "GGML_"),
				slices.Contains([]string{
					"PATH",
					"LD_LIBRARY_PATH",
					"DYLD_LIBRARY_PATH",
				}, key):
				attrs = append(attrs, slog.String(key, value))
			}
		}
	}
	return slog.GroupValue(attrs...)
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
67
type LlamaServer interface {
Jesse Gross's avatar
Jesse Gross committed
68
	ModelPath() string
69
	Load(ctx context.Context, gpus discover.GpuInfoList, requireFull bool) ([]ml.DeviceID, error)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
70
71
72
	Ping(ctx context.Context) error
	WaitUntilRunning(ctx context.Context) error
	Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error
73
	Embedding(ctx context.Context, input string) ([]float32, error)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
74
75
76
	Tokenize(ctx context.Context, content string) ([]int, error)
	Detokenize(ctx context.Context, tokens []int) (string, error)
	Close() error
Jesse Gross's avatar
Jesse Gross committed
77
78
	VRAMSize() uint64 // Total VRAM across all GPUs
	TotalSize() uint64
79
	VRAMByGPU(id ml.DeviceID) uint64
80
	Pid() int
81
82
83
	GetPort() int
	GetDeviceInfos(ctx context.Context) []ml.DeviceInfo
	HasExited() bool
Daniel Hiltgen's avatar
Daniel Hiltgen committed
84
85
}

Jesse Gross's avatar
Jesse Gross committed
86
// llmServer is an instance of a runner hosting a single model
Daniel Hiltgen's avatar
Daniel Hiltgen committed
87
type llmServer struct {
88
89
90
91
92
93
	port        int
	cmd         *exec.Cmd
	done        chan error // Channel to signal when the process exits
	status      *StatusWriter
	options     api.Options
	numParallel int
94
	modelPath   string
95

Jesse Gross's avatar
Jesse Gross committed
96
97
	loadRequest LoadRequest // Parameters used to initialize the runner

98
99
100
	// llamaModel is an instance of the cgo llama.cpp model definition
	// nil if this server is running the new engine
	llamaModel     *llama.Model
Jesse Gross's avatar
Jesse Gross committed
101
	llamaModelLock *sync.Mutex
102
103
104
105

	// textProcessor handles text encoding/decoding for the model in the Ollama engine
	// nil if this server is running the llama.cpp based engine
	textProcessor model.TextProcessor
Daniel Hiltgen's avatar
Daniel Hiltgen committed
106

Jesse Gross's avatar
Jesse Gross committed
107
108
	totalLayers  uint64
	loadStart    time.Time // Record how long it took the model to load
109
	loadProgress float32
Daniel Hiltgen's avatar
Daniel Hiltgen committed
110
111

	sem *semaphore.Weighted
112
113
}

Jesse Gross's avatar
Jesse Gross committed
114
115
116
117
118
119
120
121
122
123
124
125
126
127
type llamaServer struct {
	llmServer

	ggml     *ggml.GGML
	gpus     discover.GpuInfoList // The set of GPUs covered by the memory estimate
	estimate MemoryEstimate
}

type ollamaServer struct {
	llmServer

	mem *ml.BackendMemory
}

128
129
130
131
132
// LoadModel will load a model from disk. The model must be in the GGML format.
//
// It collects array values for arrays with a size less than or equal to
// maxArraySize. If maxArraySize is 0, the default value of 1024 is used. If
// the maxArraySize is negative, all arrays are collected.
Michael Yang's avatar
Michael Yang committed
133
func LoadModel(model string, maxArraySize int) (*ggml.GGML, error) {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
134
135
136
137
	if _, err := os.Stat(model); err != nil {
		return nil, err
	}

138
139
140
141
142
143
	f, err := os.Open(model)
	if err != nil {
		return nil, err
	}
	defer f.Close()

144
	ggml, err := ggml.Decode(f, maxArraySize)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
145
146
	return ggml, err
}
147

Daniel Hiltgen's avatar
Daniel Hiltgen committed
148
// NewLlamaServer will run a server for the given GPUs
149
func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, adapters, projectors []string, opts api.Options, numParallel int) (LlamaServer, error) {
Jesse Gross's avatar
Jesse Gross committed
150
151
152
153
	var llamaModel *llama.Model
	var textProcessor model.TextProcessor
	var err error
	if envconfig.NewEngine() || f.KV().OllamaEngineRequired() {
154
155
156
157
158
		if len(projectors) == 0 {
			textProcessor, err = model.NewTextProcessor(modelPath)
		} else {
			err = errors.New("split vision models aren't supported")
		}
Jesse Gross's avatar
Jesse Gross committed
159
160
161
162
		if err != nil {
			// To prepare for opt-out mode, instead of treating this as an error, we fallback to the old runner
			slog.Debug("model not yet supported by Ollama engine, switching to compatibility mode", "model", modelPath, "error", err)
		}
163
	}
Jesse Gross's avatar
Jesse Gross committed
164
165
166
167
168
	if textProcessor == nil {
		llamaModel, err = llama.LoadModelFromFile(modelPath, llama.ModelParams{VocabOnly: true})
		if err != nil {
			return nil, err
		}
169
170
	}

Jesse Gross's avatar
Jesse Gross committed
171
172
173
174
175
	// Verify the requested context size is <= the model training size
	trainCtx := f.KV().ContextLength()
	if opts.NumCtx > int(trainCtx) && trainCtx > 0 {
		slog.Warn("requested context size too large for model", "num_ctx", opts.NumCtx, "n_ctx_train", trainCtx)
		opts.NumCtx = int(trainCtx)
176
177
	}

178
179
	opts.NumBatch = min(opts.NumBatch, opts.NumCtx)

Jesse Gross's avatar
Jesse Gross committed
180
	loadRequest := LoadRequest{LoraPath: adapters, KvSize: opts.NumCtx * numParallel, BatchSize: opts.NumBatch, Parallel: numParallel, MultiUserCache: envconfig.MultiUserCache()}
181

Jesse Gross's avatar
Jesse Gross committed
182
183
184
185
186
	defaultThreads := discover.GetSystemInfo().GetOptimalThreadCount()
	if opts.NumThread > 0 {
		loadRequest.NumThreads = opts.NumThread
	} else if defaultThreads > 0 {
		loadRequest.NumThreads = defaultThreads
187
	}
Michael Yang's avatar
Michael Yang committed
188

Jesse Gross's avatar
Jesse Gross committed
189
	// TODO - NUMA support currently doesn't work properly
190
191

	if opts.MainGPU > 0 {
Jesse Gross's avatar
Jesse Gross committed
192
		loadRequest.MainGPU = opts.MainGPU
193
194
	}

Jesse Gross's avatar
Jesse Gross committed
195
196
	if len(projectors) > 0 && llamaModel != nil {
		loadRequest.ProjectorPath = projectors[0]
197
198
	}

199
200
	fa := envconfig.FlashAttention(f.FlashAttention())

Jesse Gross's avatar
Jesse Gross committed
201
202
	// This will disable flash attention unless all GPUs on the system support it, even if we end up selecting a subset
	// that can handle it.
203
204
205
206
	if fa && !gpus.FlashAttentionSupported() {
		slog.Warn("flash attention enabled but not supported by gpu")
		fa = false
	}
Sam's avatar
Sam committed
207

Michael Yang's avatar
Michael Yang committed
208
	if fa && !f.SupportsFlashAttention() {
209
210
211
212
		slog.Warn("flash attention enabled but not supported by model")
		fa = false
	}

213
	kvct := strings.ToLower(envconfig.KvCacheType())
214
215
216

	if fa {
		slog.Info("enabling flash attention")
Jesse Gross's avatar
Jesse Gross committed
217
		loadRequest.FlashAttention = true
218
219
220

		// Flash Attention also supports kv cache quantization
		// Enable if the requested and kv cache type is supported by the model
221
		if f.SupportsKVCacheType(kvct) {
Jesse Gross's avatar
Jesse Gross committed
222
			loadRequest.KvCacheType = kvct
223
224
		} else {
			slog.Warn("kv cache type not supported by model", "type", kvct)
Sam's avatar
Sam committed
225
		}
226
227
228
	} else if kvct != "" && kvct != "f16" {
		slog.Warn("quantized kv cache requested but flash attention disabled", "type", kvct)
	}
229

Jesse Gross's avatar
Jesse Gross committed
230
	availableLibs := make(map[string]string)
231
232
	if entries, err := os.ReadDir(discover.LibOllamaPath); err == nil {
		for _, entry := range entries {
Jesse Gross's avatar
Jesse Gross committed
233
			availableLibs[entry.Name()] = filepath.Join(discover.LibOllamaPath, entry.Name())
Michael Yang's avatar
Michael Yang committed
234
235
236
		}
	}

Jesse Gross's avatar
Jesse Gross committed
237
238
239
240
241
	var gpuLibs []string
	for _, gpu := range gpus {
		gpuLibs = append(gpuLibs, gpu.RunnerName())
	}

Michael Yang's avatar
Michael Yang committed
242
	requested := envconfig.LLMLibrary()
Jesse Gross's avatar
Jesse Gross committed
243
	if availableLibs[requested] != "" {
Michael Yang's avatar
Michael Yang committed
244
		slog.Info("using requested gpu library", "requested", requested)
Jesse Gross's avatar
Jesse Gross committed
245
		gpuLibs = []string{requested}
Michael Yang's avatar
Michael Yang committed
246
247
248
	}

	var compatible []string
Jesse Gross's avatar
Jesse Gross committed
249
250
251
252
253
254
255
256
257
258
259
260
261
	for _, gpuLib := range gpuLibs {
		var matchingLibs []string
		for k := range availableLibs {
			// exact match first
			if k == gpuLib {
				matchingLibs = append([]string{k}, matchingLibs...)
				continue
			}

			// then match the family (e.g. 'cuda')
			if strings.Split(k, "_")[0] == strings.Split(gpuLib, "_")[0] {
				matchingLibs = append(matchingLibs, k)
			}
262
		}
263

Jesse Gross's avatar
Jesse Gross committed
264
265
		if len(matchingLibs) > 0 {
			compatible = append(compatible, matchingLibs[0])
Daniel Hiltgen's avatar
Daniel Hiltgen committed
266
		}
Michael Yang's avatar
Michael Yang committed
267
	}
Jesse Gross's avatar
Jesse Gross committed
268

269
270
271
272
273
274
275
276
277
	exe, err := os.Executable()
	if err != nil {
		return nil, fmt.Errorf("unable to lookup executable path: %w", err)
	}

	if eval, err := filepath.EvalSymlinks(exe); err == nil {
		exe = eval
	}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
278
	// iterate through compatible GPU libraries such as 'cuda_v12', 'rocm', etc.
Michael Yang's avatar
Michael Yang committed
279
280
281
	// adding each library's respective path to the LD_LIBRARY_PATH, until finally running
	// without any LD_LIBRARY_PATH flags
	for {
282
283
284
285
286
287
288
289
290
		port := 0
		if a, err := net.ResolveTCPAddr("tcp", "localhost:0"); err == nil {
			var l *net.TCPListener
			if l, err = net.ListenTCP("tcp", a); err == nil {
				port = l.Addr().(*net.TCPAddr).Port
				l.Close()
			}
		}
		if port == 0 {
291
			slog.Debug("ResolveTCPAddr failed, using random port")
292
293
			port = rand.Intn(65535-49152) + 49152 // get a random port in the ephemeral range
		}
Jesse Gross's avatar
Jesse Gross committed
294
		params := []string{"runner"}
295
296
297
		if textProcessor != nil {
			// New engine
			// TODO - if we have failure to load scenarios, add logic to retry with the old runner
Jesse Gross's avatar
Jesse Gross committed
298
			params = append(params, "--ollama-engine")
Jesse Gross's avatar
Jesse Gross committed
299
		}
Jesse Gross's avatar
Jesse Gross committed
300
301
		params = append(params, "--model", modelPath)
		params = append(params, "--port", strconv.Itoa(port))
302

303
304
305
		var pathEnv string
		switch runtime.GOOS {
		case "windows":
306
			pathEnv = "PATH"
307
308
309
310
		case "darwin":
			pathEnv = "DYLD_LIBRARY_PATH"
		default:
			pathEnv = "LD_LIBRARY_PATH"
311
		}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
312

313
314
315
		// Note: we always put our dependency paths first
		// since these are the exact version we compiled/linked against
		libraryPaths := []string{discover.LibOllamaPath}
316
		if libraryPath, ok := os.LookupEnv(pathEnv); ok {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
317
			libraryPaths = append(libraryPaths, filepath.SplitList(libraryPath)...)
318
319
		}

320
		ggmlPaths := []string{discover.LibOllamaPath}
Jesse Gross's avatar
Jesse Gross committed
321
322
		for _, c := range compatible {
			if libpath, ok := availableLibs[c]; ok {
Michael Yang's avatar
Michael Yang committed
323
				slog.Debug("adding gpu library", "path", libpath)
324
				libraryPaths = append([]string{libpath}, libraryPaths...)
325
				ggmlPaths = append(ggmlPaths, libpath)
Michael Yang's avatar
Michael Yang committed
326
327
328
			}
		}

Jesse Gross's avatar
Jesse Gross committed
329
330
331
332
		for _, gpu := range gpus {
			if gpu.DependencyPath != nil {
				slog.Debug("adding gpu dependency paths", "paths", gpu.DependencyPath)
				libraryPaths = append(gpu.DependencyPath, libraryPaths...)
333
				ggmlPaths = append(ggmlPaths, gpu.DependencyPath...)
Jesse Gross's avatar
Jesse Gross committed
334
			}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
335
336
		}

Michael Yang's avatar
Michael Yang committed
337
338
339
		// finally, add the root library path
		libraryPaths = append(libraryPaths, discover.LibOllamaPath)

Jesse Gross's avatar
Jesse Gross committed
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
		s := llmServer{
			port:           port,
			cmd:            exec.Command(exe, params...),
			status:         NewStatusWriter(os.Stderr),
			options:        opts,
			modelPath:      modelPath,
			loadRequest:    loadRequest,
			llamaModel:     llamaModel,
			llamaModelLock: &sync.Mutex{},
			textProcessor:  textProcessor,
			numParallel:    numParallel,
			sem:            semaphore.NewWeighted(int64(numParallel)),
			totalLayers:    f.KV().BlockCount() + 1,
			loadStart:      time.Now(),
			done:           make(chan error, 1),
355
		}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
356

357
		s.cmd.Env = os.Environ()
358
359
		s.cmd.Stdout = os.Stdout
		s.cmd.Stderr = s.status
360
		s.cmd.SysProcAttr = LlamaServerSysProcAttr
361

362
		// Always filter down the set of GPUs in case there are any unsupported devices that might crash
363
		envWorkarounds := gpus.GetVisibleDevicesEnv()
364
365
		pathEnvVal := strings.Join(libraryPaths, string(filepath.ListSeparator))

Jesse Gross's avatar
Jesse Gross committed
366
		// Update or add the path variable with our adjusted version
367
		pathNeeded := true
368
		ollamaPathNeeded := true
369
		envWorkaroundDone := make([]bool, len(envWorkarounds))
370
371
372
373
374
		for i := range s.cmd.Env {
			cmp := strings.SplitN(s.cmd.Env[i], "=", 2)
			if strings.EqualFold(cmp[0], pathEnv) {
				s.cmd.Env[i] = pathEnv + "=" + pathEnvVal
				pathNeeded = false
375
376
377
			} else if strings.EqualFold(cmp[0], "OLLAMA_LIBRARY_PATH") {
				s.cmd.Env[i] = "OLLAMA_LIBRARY_PATH=" + strings.Join(ggmlPaths, string(filepath.ListSeparator))
				ollamaPathNeeded = false
Daniel Hiltgen's avatar
Daniel Hiltgen committed
378
			} else if len(envWorkarounds) != 0 {
379
380
381
382
383
				for j, kv := range envWorkarounds {
					tmp := strings.SplitN(kv, "=", 2)
					if strings.EqualFold(cmp[0], tmp[0]) {
						s.cmd.Env[i] = kv
						envWorkaroundDone[j] = true
Daniel Hiltgen's avatar
Daniel Hiltgen committed
384
385
					}
				}
386
			}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
387
		}
388
389
		if pathNeeded {
			s.cmd.Env = append(s.cmd.Env, pathEnv+"="+pathEnvVal)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
390
		}
391
392
393
		if ollamaPathNeeded {
			s.cmd.Env = append(s.cmd.Env, "OLLAMA_LIBRARY_PATH="+strings.Join(ggmlPaths, string(filepath.ListSeparator)))
		}
394
395
396
397
398
		for i, done := range envWorkaroundDone {
			if !done {
				s.cmd.Env = append(s.cmd.Env, envWorkarounds[i])
			}
		}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
399

Jesse Gross's avatar
Jesse Gross committed
400
		slog.Info("starting runner", "cmd", s.cmd)
401
		slog.Debug("subprocess", "", filteredEnv(s.cmd.Env))
402
403

		if err = s.cmd.Start(); err != nil {
Michael Yang's avatar
Michael Yang committed
404
			var msg string
405
406
407
			if s.status != nil && s.status.LastErrMsg != "" {
				msg = s.status.LastErrMsg
			}
Michael Yang's avatar
Michael Yang committed
408
409
			err := fmt.Errorf("error starting runner: %v %s", err, msg)
			if len(compatible) == 0 {
410
411
412
				if llamaModel != nil {
					llama.FreeModel(llamaModel)
				}
Michael Yang's avatar
Michael Yang committed
413
414
415
416
417
				return nil, err
			}

			slog.Warn("unable to start runner with compatible gpu", "error", err, "compatible", compatible)
			compatible = compatible[1:]
418
419
420
			continue
		}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
421
422
		// reap subprocess when it exits
		go func() {
423
424
425
			err := s.cmd.Wait()
			// Favor a more detailed message over the process exit status
			if err != nil && s.status != nil && s.status.LastErrMsg != "" {
Michael Yang's avatar
Michael Yang committed
426
				slog.Error("llama runner terminated", "error", err)
427
428
429
				if strings.Contains(s.status.LastErrMsg, "unknown model") {
					s.status.LastErrMsg = "this model is not supported by your version of Ollama. You may need to upgrade"
				}
Michael Yang's avatar
lint  
Michael Yang committed
430
				s.done <- errors.New(s.status.LastErrMsg)
431
432
433
			} else {
				s.done <- err
			}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
434
435
		}()

436
		if textProcessor != nil {
Jesse Gross's avatar
Jesse Gross committed
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
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
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
			return &ollamaServer{llmServer: s}, nil
		} else {
			return &llamaServer{llmServer: s, ggml: f}, nil
		}
	}
}

func (s *llmServer) ModelPath() string {
	return s.modelPath
}

type LoadOperation int

// The order of these constants are significant because we iterate over the operations. They
// should be in order of increasingly loading the model.
const (
	LoadOperationFit    LoadOperation = iota // Return memory requirements but do not allocate
	LoadOperationAlloc                       // Allocate memory but do not load the weights
	LoadOperationCommit                      // Load weights - further changes cannot be made after this
	LoadOperationClose                       // Close model and free memory
)

func (o LoadOperation) String() string {
	switch o {
	case LoadOperationFit:
		return "fit"
	case LoadOperationAlloc:
		return "alloc"
	case LoadOperationCommit:
		return "commit"
	case LoadOperationClose:
		return "close"
	default:
		return "unknown"
	}
}

type LoadRequest struct {
	Operation LoadOperation

	LoraPath       []string
	Parallel       int
	BatchSize      int
	FlashAttention bool
	KvSize         int
	KvCacheType    string
	NumThreads     int
	GPULayers      ml.GPULayersList
	MultiUserCache bool

	// Legacy fields - not used with the Ollama engine
	ProjectorPath string
	MainGPU       int
	UseMmap       bool
}

type LoadResponse struct {
	Success bool
	Memory  ml.BackendMemory
}

var ErrLoadRequiredFull = errors.New("unable to load full model on GPU")

500
func (s *llamaServer) Load(ctx context.Context, gpus discover.GpuInfoList, requireFull bool) ([]ml.DeviceID, error) {
Jesse Gross's avatar
Jesse Gross committed
501
502
503
504
505
506
507
508
509
510
511
	systemInfo := discover.GetSystemInfo()
	systemTotalMemory := systemInfo.System.TotalMemory
	systemFreeMemory := systemInfo.System.FreeMemory
	systemSwapFreeMemory := systemInfo.System.FreeSwap
	slog.Info("system memory", "total", format.HumanBytes2(systemTotalMemory), "free", format.HumanBytes2(systemFreeMemory), "free_swap", format.HumanBytes2(systemSwapFreeMemory))

	g := pickBestFullFitByLibrary(s.ggml, s.modelPath, []string{s.loadRequest.ProjectorPath}, s.loadRequest.LoraPath, s.options, gpus, s.numParallel)
	if g == nil {
		if !requireFull {
			g = pickBestPartialFitByLibrary(s.ggml, []string{s.loadRequest.ProjectorPath}, s.loadRequest.LoraPath, s.options, gpus, s.numParallel)
		} else {
512
			slog.Info("model requires more memory than is currently available, evicting a model to make space", "estimate", s.estimate)
513
			return nil, ErrLoadRequiredFull
Jesse Gross's avatar
Jesse Gross committed
514
515
516
517
518
519
520
521
		}
	}

	gpus = g
	s.estimate = estimateGPULayers(gpus, s.ggml, []string{s.loadRequest.ProjectorPath}, s.options, s.numParallel)

	if len(gpus) > 1 || gpus[0].Library != "cpu" {
		switch {
522
		case gpus[0].Library == "Metal" && s.estimate.VRAMSize > systemInfo.System.TotalMemory:
Jesse Gross's avatar
Jesse Gross committed
523
524
525
			// disable partial offloading when model is greater than total system memory as this
			// can lead to locking up the system
			s.options.NumGPU = 0
526
		case gpus[0].Library != "Metal" && s.estimate.Layers == 0:
Jesse Gross's avatar
Jesse Gross committed
527
			// Don't bother loading into the GPU if no layers can fit
528
			gpus = discover.GpuInfoList{discover.GetCPUInfo()}
Jesse Gross's avatar
Jesse Gross committed
529
530
531
532
533
534
535
536
537
538
539
540
		case s.options.NumGPU < 0 && s.estimate.Layers > 0 && gpus[0].Library != "cpu":
			s.options.NumGPU = s.estimate.Layers
		}
	}

	// On linux and windows, over-allocating CPU memory will almost always result in an error
	// Darwin has fully dynamic swap so has no direct concept of free swap space
	if runtime.GOOS != "darwin" {
		systemMemoryRequired := s.estimate.TotalSize - s.estimate.VRAMSize
		available := systemInfo.System.FreeMemory + systemInfo.System.FreeSwap
		if systemMemoryRequired > available {
			slog.Warn("model request too large for system", "requested", format.HumanBytes2(systemMemoryRequired), "available", format.HumanBytes2(available), "total", format.HumanBytes2(systemInfo.System.TotalMemory), "free", format.HumanBytes2(systemInfo.System.FreeMemory), "swap", format.HumanBytes2(systemInfo.System.FreeSwap))
541
			return nil, fmt.Errorf("model requires more system memory (%s) than is available (%s)", format.HumanBytes2(systemMemoryRequired), format.HumanBytes2(available))
Jesse Gross's avatar
Jesse Gross committed
542
543
544
545
546
547
548
549
550
551
552
553
554
555
		}
	}

	slog.Info("offload", "", s.estimate)

	s.gpus = gpus
	s.loadRequest.GPULayers = createGPULayers(s.estimate, s.ggml, gpus, s.options.NumGPU)

	// Mmap is only supported on the llama engine
	if s.textProcessor == nil {
		s.loadRequest.UseMmap = true

		// mmap has issues with partial offloading on metal
		for _, g := range gpus {
556
			if g.Library == "Metal" &&
Jesse Gross's avatar
Jesse Gross committed
557
558
559
560
561
562
563
564
565
566
				uint64(s.options.NumGPU) > 0 &&
				uint64(s.options.NumGPU) < s.ggml.KV().BlockCount()+1 {
				s.options.UseMMap = new(bool)
				*s.options.UseMMap = false
			}
		}

		// Windows CUDA should not use mmap for best performance
		// Linux  with a model larger than free space, mmap leads to thrashing
		// For CPU loads we want the memory to be allocated, not FS cache
567
		if (runtime.GOOS == "windows" && gpus[0].Library == "CUDA" && s.options.UseMMap == nil) ||
Jesse Gross's avatar
Jesse Gross committed
568
569
570
571
572
573
574
575
			(runtime.GOOS == "linux" && systemInfo.System.FreeMemory < s.estimate.TotalSize && s.options.UseMMap == nil) ||
			(gpus[0].Library == "cpu" && s.options.UseMMap == nil) ||
			(s.options.UseMMap != nil && !*s.options.UseMMap) {
			s.loadRequest.UseMmap = false
		}
	}

	if err := s.waitUntilRunnerLaunched(ctx); err != nil {
576
		return nil, err
Jesse Gross's avatar
Jesse Gross committed
577
578
579
580
	}

	resp, err := s.initModel(ctx, s.loadRequest, LoadOperationCommit)
	if err != nil {
581
		return nil, err
Jesse Gross's avatar
Jesse Gross committed
582
583
584
585
586
587
588
589
590
591
	}

	// On the Ollama engine, we can print out a summary of the memory allocations.
	// We don't have this for the llama engine but it does something similar itself.
	if s.textProcessor != nil {
		resp.Memory.Log(slog.LevelInfo)
	}

	if !resp.Success {
		slog.Warn("failed to allocate memory for model", "memory", resp.Memory)
592
		return nil, errors.New("failed to allocate memory for model")
Jesse Gross's avatar
Jesse Gross committed
593
594
595
596
597
598
	}

	// The llama engine does its memory allocations together with model loading, so we
	// need to wait until it is done to ensure that we have accurate memory data before
	// loading the next model
	if s.textProcessor == nil {
599
		return uniqueDeviceIDs(s.loadRequest.GPULayers), s.WaitUntilRunning(ctx)
Jesse Gross's avatar
Jesse Gross committed
600
	} else {
601
		return uniqueDeviceIDs(s.loadRequest.GPULayers), nil
Jesse Gross's avatar
Jesse Gross committed
602
603
604
605
606
607
608
609
	}
}

// createGPULayers maps from the tensor splits assigned by the memory estimates to explicit assignment
// of particular layers onto GPUs
func createGPULayers(estimate MemoryEstimate, ggml *ggml.GGML, gpus discover.GpuInfoList, numGPU int) ml.GPULayersList {
	if numGPU <= 0 {
		return nil
610
	}
Jesse Gross's avatar
Jesse Gross committed
611
612
613

	gpuLayers := make(ml.GPULayersList, len(gpus))
	for i := range gpuLayers {
614
		gpuLayers[i].DeviceID = gpus[i].DeviceID
Jesse Gross's avatar
Jesse Gross committed
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
	}

	var sum float32
	splits := make([]float32, len(estimate.TensorSplit))
	// cumulative sum of all splits
	for i := range splits {
		sum += float32(estimate.TensorSplit[i])
		splits[i] = sum
	}

	if sum <= 0 {
		return nil
	}

	// normalize splits
	for i := range splits {
		splits[i] /= sum
	}

	blocks := int(ggml.KV().BlockCount())
	gpuRangeStart := max(0, blocks-numGPU)
	gpuRangeStop := min(gpuRangeStart+numGPU, blocks+1)
	for i := range blocks + 1 {
		if i < gpuRangeStart || i >= gpuRangeStop {
			continue
		}

		index := slices.IndexFunc(splits, func(f float32) bool { return float32(i-gpuRangeStart)/float32(gpuRangeStop-gpuRangeStart) < f })
		if index < 0 || index >= len(gpus) {
			continue
		}

		gpuLayers[index].Layers = append(gpuLayers[index].Layers, i)
	}

	return gpuLayers
}

// Load finds the optimal layout of layers to offload on GPUs based on no initial information about the size of the model
// It does this by:
// 1. Assigning the full model to the GPU with the largest available free memory
// 2. Attempting to allocate the layout and receiving the memory requirements in response
// 3. Creating a new layout based on the updated memory information
// 4. Going back to step 2 and looping until we either stabilize on a particular layout or discover that we have entered a cycle
//
// This process is repeated for higher levels of loading the model (fit, allocate, commit). The earlier levels are quicker,
// allowing for faster iteration, but may return less information.
662
663
664
//
// Returns the list of GPU IDs that were used in the final allocation on success
func (s *ollamaServer) Load(ctx context.Context, gpus discover.GpuInfoList, requireFull bool) ([]ml.DeviceID, error) {
Jesse Gross's avatar
Jesse Gross committed
665
666
667
668
669
	var success bool
	defer func() {
		if !success {
			s.initModel(ctx, LoadRequest{}, LoadOperationClose)
		}
670
671
672
		if s.mem != nil {
			s.mem.Log(slog.LevelInfo)
		}
Jesse Gross's avatar
Jesse Gross committed
673
674
675
676
677
678
679
680
681
682
683
684
	}()

	slog.Info("loading model", "model layers", s.totalLayers, "requested", s.options.NumGPU)

	systemInfo := discover.GetSystemInfo()
	systemTotalMemory := systemInfo.System.TotalMemory
	systemFreeMemory := systemInfo.System.FreeMemory
	systemSwapFreeMemory := systemInfo.System.FreeSwap
	slog.Info("system memory", "total", format.HumanBytes2(systemTotalMemory), "free", format.HumanBytes2(systemFreeMemory), "free_swap", format.HumanBytes2(systemSwapFreeMemory))

	if !(len(gpus) == 1 && gpus[0].Library == "cpu") {
		for _, gpu := range gpus {
685
686
687
688
			available := gpu.FreeMemory - envconfig.GpuOverhead() - gpu.MinimumMemory
			if gpu.FreeMemory < envconfig.GpuOverhead()+gpu.MinimumMemory {
				available = 0
			}
689
			slog.Info("gpu memory", "id", gpu.ID, "library", gpu.Library,
690
				"available", format.HumanBytes2(available),
Jesse Gross's avatar
Jesse Gross committed
691
692
693
694
695
696
697
698
699
700
701
				"free", format.HumanBytes2(gpu.FreeMemory),
				"minimum", format.HumanBytes2(gpu.MinimumMemory),
				"overhead", format.HumanBytes2(envconfig.GpuOverhead()))
		}
	}

	pastAllocations := make(map[uint64]struct{})
	var backoff float32

	gpuLayers, err := s.createLayout(systemInfo, gpus, s.mem, requireFull, backoff)
	if err != nil {
702
		return nil, err
Jesse Gross's avatar
Jesse Gross committed
703
704
705
	}

	if err := s.waitUntilRunnerLaunched(ctx); err != nil {
706
		return nil, err
Jesse Gross's avatar
Jesse Gross committed
707
708
709
710
711
712
713
714
715
	}

nextOperation:
	for operation := LoadOperationFit; operation < LoadOperationCommit; operation++ {
	nextLoad:
		for {
			s.loadRequest.GPULayers = gpuLayers
			resp, err := s.initModel(ctx, s.loadRequest, operation)
			if err != nil {
716
				return nil, err
Jesse Gross's avatar
Jesse Gross committed
717
718
719
720
721
722
723
724
725
726
727
			}

			resp.Memory.Log(slog.LevelDebug)
			slog.Debug("memory", "success", resp.Success, "required", resp.Memory)

			pastAllocations[gpuLayers.Hash()] = struct{}{}
			s.mem = &resp.Memory

			for {
				newGPULayers, err := s.createLayout(systemInfo, gpus, s.mem, requireFull, backoff)
				if err != nil {
728
					return nil, err
Jesse Gross's avatar
Jesse Gross committed
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
				}

				slog.Debug("new layout created", "layers", newGPULayers)

				// We get additional memory information over time, which will reduce the number of
				// layers that can fit, so fewer layers is actually better. As long as we haven't seen
				// this layout before and it doesn't have more layers than the last one, we can keep
				// trying to see if we can do better.
				if _, ok := pastAllocations[newGPULayers.Hash()]; !ok && newGPULayers.Sum() <= gpuLayers.Sum() {
					gpuLayers = newGPULayers
					continue nextLoad
				}

				// If we are looping around a few different layouts due to graphs moving off and on
				// GPUs, make sure that we try out the intermediate states. For example, if we are
				// looping between offloading 39 and 41 layers, we should also check 40.
				//
				// This switches strategies to force an incremental number of layers to be offloaded
				// and checking the memory layout. If the allocation succeeds and creating a new layout
				// without forcing offload yields the same or greater number of layers offloaded, then
				// the trial is successful.
				//
				// This alternate strategy does not introduce the possibility of loops with the overall
				// state machine, as it exits this code block either with a successful result, moving
				// to the next operation or the original number of layers offloaded.
				if s.options.NumGPU < 0 && newGPULayers.Sum()-gpuLayers.Sum() > 1 {
					for i := newGPULayers.Sum() - 1; i >= gpuLayers.Sum(); i-- {
						slog.Debug("exploring intermediate layers", "layer", i)

						s.options.NumGPU = i
						newGPULayers, err = s.createLayout(systemInfo, gpus, s.mem, requireFull, backoff)
						s.options.NumGPU = -1
						if err != nil {
762
							return nil, err
Jesse Gross's avatar
Jesse Gross committed
763
764
765
766
767
768
769
						}

						slog.Debug("new layout created", "layers", newGPULayers)

						s.loadRequest.GPULayers = newGPULayers
						resp, err = s.initModel(ctx, s.loadRequest, operation)
						if err != nil {
770
							return nil, err
Jesse Gross's avatar
Jesse Gross committed
771
772
773
774
775
776
777
778
						}

						resp.Memory.Log(slog.LevelDebug)
						slog.Debug("memory", "success", resp.Success, "required", resp.Memory)

						if resp.Success {
							verifyGPULayers, err := s.createLayout(systemInfo, gpus, &resp.Memory, requireFull, backoff)
							if err != nil {
779
								return nil, err
Jesse Gross's avatar
Jesse Gross committed
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
							}

							slog.Debug("verifying layout", "layers", verifyGPULayers)

							if newGPULayers.Sum() <= verifyGPULayers.Sum() {
								gpuLayers = newGPULayers

								// Since we are going backwards (increasing the number of layers), ensure that
								// we can come back down if needed
								clear(pastAllocations)

								continue nextOperation
							}
						}
					}
				}

				// If we generated a layout a second time or go backwards, then we've converged. Use the last
				// layout before the repeat, which is already allocated.
				if resp.Success {
					continue nextOperation
				}

				if s.options.NumGPU >= 0 {
804
					return nil, fmt.Errorf("memory layout cannot be allocated with num_gpu = %v", s.options.NumGPU)
Jesse Gross's avatar
Jesse Gross committed
805
806
807
808
809
810
811
812
813
				}

				// Memory allocation failed even though we created a layout that we thought should
				// fit in available memory. This could happen if either our free memory reports
				// are incorrect or if available memory is changing between layout and allocation
				// time. Apply an exponential backoff to try to find the real amount of available
				// space.
				if backoff > 1 {
					slog.Warn("memory layout cannot be allocated", "memory", resp.Memory)
814
					return nil, errors.New("memory layout cannot be allocated")
Jesse Gross's avatar
Jesse Gross committed
815
816
817
818
819
820
821
822
823
824
825
826
827
828
				} else if backoff == 0 {
					backoff = 0.01
				} else {
					backoff *= 2
				}

				slog.Info("model layout did not fit, applying backoff", "backoff", fmt.Sprintf("%.2f", backoff))
			}
		}
	}

	s.loadRequest.GPULayers = gpuLayers
	resp, err := s.initModel(ctx, s.loadRequest, LoadOperationCommit)
	if err != nil {
829
		return nil, err
Jesse Gross's avatar
Jesse Gross committed
830
831
832
833
834
835
836
	}

	success = resp.Success
	s.mem = &resp.Memory

	if !success {
		slog.Warn("failed to commit memory for model", "memory", resp.Memory)
837
		return nil, errors.New("failed to commit memory for model")
Jesse Gross's avatar
Jesse Gross committed
838
839
	}

840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
	return uniqueDeviceIDs(gpuLayers), nil
}

func uniqueDeviceIDs(gpuLayers ml.GPULayersList) []ml.DeviceID {
	devices := []ml.DeviceID{}
	for _, layer := range gpuLayers {
		new := true
		for _, ID := range devices {
			if layer.DeviceID == ID {
				new = false
				break
			}
		}
		if new {
			devices = append(devices, layer.DeviceID)
		}
	}
	return devices
Jesse Gross's avatar
Jesse Gross committed
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
}

// createLayout uses the current best view of memory requirements and creates a layout of model layers on GPUs.
// It does this by:
// - Calculating how much space each layer requires
// - Calculating how much space each GPU has available for layers, based on free memory and space occupied by the graph
// - Assigning layers
// - Ensuring that we don't exceed limits, such as requirements about partial offloading or system memory
func (s *ollamaServer) createLayout(systemInfo discover.SystemInfo, systemGPUs discover.GpuInfoList, memory *ml.BackendMemory, requireFull bool, backoff float32) (ml.GPULayersList, error) {
	if s.totalLayers == 0 || s.options.NumGPU == 0 || len(systemGPUs) == 0 || (len(systemGPUs) == 1 && systemGPUs[0].Library == "cpu") {
		return ml.GPULayersList{}, nil
	}

	gpus := append(make(discover.GpuInfoList, 0, len(systemGPUs)), systemGPUs...)
	sort.Sort(sort.Reverse(discover.ByFreeMemory(gpus)))

	if memory == nil {
		memory = &ml.BackendMemory{CPU: ml.DeviceMemory{
876
877
			Weights: make([]uint64, s.totalLayers),
			Cache:   make([]uint64, s.totalLayers),
Jesse Gross's avatar
Jesse Gross committed
878
879
880
881
882
883
		}}
	}

	layers := make([]uint64, len(memory.CPU.Weights))
	for i := range layers {
		for j := range memory.GPUs {
884
885
			layers[i] += memory.GPUs[j].Weights[i]
			layers[i] += memory.GPUs[j].Cache[i]
Jesse Gross's avatar
Jesse Gross committed
886
		}
887
888
		layers[i] += memory.CPU.Weights[i]
		layers[i] += memory.CPU.Cache[i]
889
		logutil.Trace("layer to assign", "layer", i, "size", format.HumanBytes2(layers[i]))
Jesse Gross's avatar
Jesse Gross committed
890
891
892
893
894
895
896
897
898
899
900
901
	}

	gpuLayers := ml.GPULayersList{}
	for _, gl := range gpus.ByLibrary() {
		// If a GPU already has a graph allocated on it, then we should continue to use it.
		// Otherwise, we lose information that we got from previous allocations, which can
		// cause cycling. Plus, we get more information about required allocation from each
		// iteration, so it doesn't make sense that a later iteration would use fewer GPUs.
		lastUsedGPU := 0
		for i := range gl {
			found := false
			for j := range memory.GPUs {
902
				if gl[i].DeviceID == memory.GPUs[j].DeviceID {
903
					if memory.GPUs[j].Graph != 0 {
Jesse Gross's avatar
Jesse Gross committed
904
905
906
						lastUsedGPU = i
					}

907
					reserved := uint64(float32(gl[i].FreeMemory)*backoff) + gl[i].MinimumMemory + envconfig.GpuOverhead() + memory.GPUs[j].Graph
Jesse Gross's avatar
Jesse Gross committed
908
909
910
911
912
913
					if gl[i].FreeMemory > reserved {
						gl[i].FreeMemory -= reserved
					} else {
						gl[i].FreeMemory = 0
					}

914
					slog.Debug("available gpu", "id", gl[i].ID, "library", gl[i].Library,
Jesse Gross's avatar
Jesse Gross committed
915
916
917
						"available layer vram", format.HumanBytes2(gl[i].FreeMemory),
						"backoff", fmt.Sprintf("%.2f", backoff), "minimum", format.HumanBytes2(gl[i].MinimumMemory),
						"overhead", format.HumanBytes2(envconfig.GpuOverhead()),
918
						"graph", format.HumanBytes2(memory.GPUs[j].Graph))
Jesse Gross's avatar
Jesse Gross committed
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936

					found = true
					break
				}
			}
			if !found {
				// The runner doesn't report seeing this GPU
				gl[i].FreeMemory = 0
			}
		}

		libraryGpuLayers := assignLayers(layers, gl, s.options.NumGPU, lastUsedGPU)
		if libraryGpuLayers.Sum() > gpuLayers.Sum() {
			gpuLayers = libraryGpuLayers
		}
	}

	// These sizes will only increase as we go through additional iterations and get additional information.
937
	cpuSize := memory.InputWeights + memory.CPU.Graph
Jesse Gross's avatar
Jesse Gross committed
938
939
940
	var vramSize uint64
	for _, gl := range gpuLayers {
		for _, gpu := range memory.GPUs {
941
			if gl.DeviceID == gpu.DeviceID {
942
				vramSize += gpu.Graph
Jesse Gross's avatar
Jesse Gross committed
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
				break
			}
		}
	}

nextLayer:
	for i := range layers {
		for _, g := range gpuLayers {
			for _, gl := range g.Layers {
				if i == gl {
					vramSize += layers[i]
					continue nextLayer
				}
			}
		}
		cpuSize += layers[i]
	}

	if requireFull {
		if gpuLayers.Sum() < len(layers) && (s.options.NumGPU < 0 || gpuLayers.Sum() < s.options.NumGPU) {
			return nil, ErrLoadRequiredFull
		}

		if cpuSize > systemInfo.System.FreeMemory {
			return nil, ErrLoadRequiredFull
		}
	}

	// On linux and windows, over-allocating CPU memory will almost always result in an error
	// Darwin has fully dynamic swap so has no direct concept of free swap space
	if runtime.GOOS != "darwin" {
		available := systemInfo.System.FreeMemory + systemInfo.System.FreeSwap
		if cpuSize > available {
			slog.Warn("model request too large for system", "requested", format.HumanBytes2(cpuSize), "available", format.HumanBytes2(available), "total", format.HumanBytes2(systemInfo.System.TotalMemory), "free", format.HumanBytes2(systemInfo.System.FreeMemory), "swap", format.HumanBytes2(systemInfo.System.FreeSwap))
			return nil, fmt.Errorf("model requires more system memory (%s) than is available (%s)", format.HumanBytes2(cpuSize), format.HumanBytes2(available))
		}
	} else {
		if vramSize > systemInfo.System.TotalMemory {
			// disable partial offloading when model is greater than total system memory as this
			// can lead to locking up the system
			s.options.NumGPU = 0
			gpuLayers = ml.GPULayersList{}
		}
	}

	if gpuLayers.Sum() == 0 {
		slog.Debug("insufficient VRAM to load any model layers")
	}

	return gpuLayers, nil
}

// assignLayers packs the maximum number of layers onto the smallest set of GPUs and comes up with a layer assignment
func assignLayers(layers []uint64, gpus discover.GpuInfoList, requestedLayers int, lastUsedGPU int) (gpuLayers ml.GPULayersList) {
	// If we can't fit everything then prefer offloading layers other than the output layer
	for range 2 {
		// requestedLayers may be -1 if nothing was requested
		requestedLayers = min(len(layers), requestedLayers)

		if !envconfig.SchedSpread() {
			for i := lastUsedGPU; i < len(gpus); i++ {
				// Try to pack things into as few GPUs as possible
				forceRequest := i == len(gpus)-1
				gpuLayers = findBestFit(layers, gpus[:i+1], requestedLayers, forceRequest)
				if gpuLayers.Sum() == len(layers) || gpuLayers.Sum() == requestedLayers {
					break
				}
			}
		} else {
			gpuLayers = findBestFit(layers, gpus, requestedLayers, true)
		}

		// We only stop if we've gotten all of the layers - even if we got requestedLayers, we still
		// might want to try dropping the output layer.
		if gpuLayers.Sum() == len(layers) {
			return gpuLayers
		}

		layers = layers[:len(layers)-1]
	}

	return gpuLayers
}

// findBestFit binary searches to find the smallest capacity factor that can fit
// the max number of layers. The capacity factor is multiplied by the free space on
// each GPU and a small one will force even balancing.
func findBestFit(layers []uint64, gpus discover.GpuInfoList, requestedLayers int, forceRequest bool) (gpuLayers ml.GPULayersList) {
	var high float32 = 1
	var low float32 = 0

	// If we need to fulfill the requested number of layers, pretend we have almost infinite VRAM
	if requestedLayers >= 0 && forceRequest {
		high = 1000
	}

	bestAssignments := greedyFit(layers, gpus, high, requestedLayers)
	maxNumGPU := bestAssignments.Sum()
	if maxNumGPU == 0 {
		return bestAssignments
	}

	for high-low > 1e-6 {
		mid := (low + high) / 2
		assignments := greedyFit(layers, gpus, mid, requestedLayers)
		if assignments.Sum() == maxNumGPU {
			high = mid
			bestAssignments = assignments
		} else {
			low = mid
		}
	}

	return bestAssignments
}

// greedyFit assigns layers incrementally to GPUs, spilling over as each runs out of free space
func greedyFit(layers []uint64, gpus discover.GpuInfoList, capacity float32, requestedLayers int) (gpuLayers ml.GPULayersList) {
	device := len(gpus) - 1
1062
	gpuLayers = ml.GPULayersList{{DeviceID: gpus[device].DeviceID}}
Jesse Gross's avatar
Jesse Gross committed
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
	freeSpace := uint64(float32(gpus[device].FreeMemory) * capacity)
	for i := len(layers) - 1; i >= 0; i-- {
		if requestedLayers >= 0 && len(layers)-1-i >= requestedLayers {
			break
		}

		for {
			if layers[i] <= freeSpace {
				gpuLayers[0].Layers = append([]int{i}, gpuLayers[0].Layers...)
				freeSpace -= layers[i]
				break
			}

			device--
			if device < 0 {
				return gpuLayers
			}
1080
			gpuLayers = append(ml.GPULayersList{{DeviceID: gpus[device].DeviceID}}, gpuLayers...)
Jesse Gross's avatar
Jesse Gross committed
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
			freeSpace = uint64(float32(gpus[device].FreeMemory) * capacity)
		}
	}

	return gpuLayers
}

// waitUntilRunnerLaunched sleeps until the runner subprocess is alive enough
// to respond to status requests
func (s *llmServer) waitUntilRunnerLaunched(ctx context.Context) error {
	for {
		_, err := s.getServerStatus(ctx)
		if err == nil {
			break
		}

		t := time.NewTimer(10 * time.Millisecond)
		select {
		case <-t.C:
			continue
		case <-ctx.Done():
			return ctx.Err()
		}
	}

	return nil
}

// initModel sends a load request to the runner based on the request operation (fit, alloc, commit)
// and parameters
func (s *llmServer) initModel(ctx context.Context, req LoadRequest, operation LoadOperation) (*LoadResponse, error) {
	req.Operation = operation

	data, err := json.Marshal(req)
	if err != nil {
		return nil, fmt.Errorf("error marshaling load data: %w", err)
	}

	r, err := http.NewRequestWithContext(ctx, http.MethodPost, fmt.Sprintf("http://127.0.0.1:%d/load", s.port), bytes.NewBuffer(data))
	if err != nil {
		return nil, fmt.Errorf("error creating load request: %w", err)
	}
	r.Header.Set("Content-Type", "application/json")

	resp, err := http.DefaultClient.Do(r)
	if err != nil {
		return nil, fmt.Errorf("do load request: %w", err)
	}
	defer resp.Body.Close()

	body, err := io.ReadAll(resp.Body)
	if err != nil {
		return nil, fmt.Errorf("read load request: %w", err)
	}

	if resp.StatusCode >= 400 {
		log.Printf("llm load error: %s", body)
		return nil, fmt.Errorf("%s", body)
	}

	var llmResp LoadResponse
	if err := json.Unmarshal(body, &llmResp); err != nil {
		return nil, fmt.Errorf("load unmarshal encode response: %w", err)
	}

	return &llmResp, nil
1147
1148
1149
1150
1151
1152
}

type ServerStatus int

const ( // iota is reset to 0
	ServerStatusReady ServerStatus = iota
1153
	ServerStatusNoSlotsAvailable
Jesse Gross's avatar
Jesse Gross committed
1154
	ServerStatusLaunched
1155
1156
1157
1158
1159
	ServerStatusLoadingModel
	ServerStatusNotResponding
	ServerStatusError
)

1160
func (s ServerStatus) String() string {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1161
1162
1163
	switch s {
	case ServerStatusReady:
		return "llm server ready"
1164
	case ServerStatusNoSlotsAvailable:
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1165
		return "llm busy - no slots available"
Jesse Gross's avatar
Jesse Gross committed
1166
1167
	case ServerStatusLaunched:
		return "llm server launched"
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1168
1169
1170
1171
1172
1173
1174
1175
1176
	case ServerStatusLoadingModel:
		return "llm server loading model"
	case ServerStatusNotResponding:
		return "llm server not responding"
	default:
		return "llm server error"
	}
}

1177
1178
1179
type ServerStatusResponse struct {
	Status   ServerStatus `json:"status"`
	Progress float32      `json:"progress"`
1180
1181
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
1182
func (s *llmServer) getServerStatus(ctx context.Context) (ServerStatus, error) {
1183
1184
1185
1186
1187
1188
	// Fail fast if its exited
	if s.cmd.ProcessState != nil {
		msg := ""
		if s.status != nil && s.status.LastErrMsg != "" {
			msg = s.status.LastErrMsg
		}
1189
1190
		if s.cmd.ProcessState.ExitCode() == -1 {
			// Most likely a signal killed it, log some more details to try to help troubleshoot
1191
			slog.Warn("llama runner process no longer running", "sys", s.cmd.ProcessState.Sys(), "string", s.cmd.ProcessState)
1192
		}
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
		return ServerStatusError, fmt.Errorf("llama runner process no longer running: %d %s", s.cmd.ProcessState.ExitCode(), msg)
	}

	req, err := http.NewRequestWithContext(ctx, http.MethodGet, fmt.Sprintf("http://127.0.0.1:%d/health", s.port), nil)
	if err != nil {
		return ServerStatusError, fmt.Errorf("error creating GET request: %v", err)
	}
	req.Header.Set("Content-Type", "application/json")

	resp, err := http.DefaultClient.Do(req)
	if err != nil {
		if errors.Is(err, context.DeadlineExceeded) {
Michael Yang's avatar
Michael Yang committed
1205
			return ServerStatusNotResponding, errors.New("server not responding")
1206
		}
1207
1208
1209
		if strings.Contains(err.Error(), "connection refused") {
			return ServerStatusNotResponding, errors.New("connection refused")
		}
1210
1211
1212
1213
1214
1215
1216
1217
1218
		return ServerStatusError, fmt.Errorf("health resp: %w", err)
	}
	defer resp.Body.Close()

	body, err := io.ReadAll(resp.Body)
	if err != nil {
		return ServerStatusError, fmt.Errorf("read health request: %w", err)
	}

1219
1220
	var ssr ServerStatusResponse
	if err := json.Unmarshal(body, &ssr); err != nil {
1221
1222
1223
		return ServerStatusError, fmt.Errorf("health unmarshal encode response: %w", err)
	}

1224
1225
1226
1227
	switch ssr.Status {
	case ServerStatusLoadingModel:
		s.loadProgress = ssr.Progress
		return ssr.Status, nil
Jesse Gross's avatar
Jesse Gross committed
1228
	case ServerStatusLaunched, ServerStatusReady, ServerStatusNoSlotsAvailable:
1229
		return ssr.Status, nil
1230
	default:
1231
		return ssr.Status, fmt.Errorf("server error: %+v", ssr)
1232
1233
1234
	}
}

1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
// getServerStatusRetry will retry if ServerStatusNoSlotsAvailable is received
func (s *llmServer) getServerStatusRetry(ctx context.Context) (ServerStatus, error) {
	var retries int
	for {
		status, err := s.getServerStatus(ctx)
		if err != nil {
			return status, err
		}

		if status == ServerStatusNoSlotsAvailable {
			if retries >= 10 {
				return status, fmt.Errorf("no slots available after %d retries", retries)
			}

			time.Sleep(5 * time.Millisecond)
			retries++
			continue
		}

		return status, nil
	}
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
1258
func (s *llmServer) Ping(ctx context.Context) error {
1259
1260
1261
1262
1263
1264
1265
1266
	_, err := s.getServerStatus(ctx)
	if err != nil {
		slog.Debug("server unhealthy", "error", err)
		return err
	}
	return nil
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
1267
func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
1268
	stallDuration := envconfig.LoadTimeout()    // If no progress happens
1269
	stallTimer := time.Now().Add(stallDuration) // give up if we stall
1270
1271
1272

	slog.Info("waiting for llama runner to start responding")
	var lastStatus ServerStatus = -1
1273
	fullyLoaded := false
ManniX-ITA's avatar
ManniX-ITA committed
1274

1275
1276
	for {
		select {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1277
		case <-ctx.Done():
1278
			slog.Warn("client connection closed before server finished loading, aborting load")
1279
			return fmt.Errorf("timed out waiting for llama runner to start: %w", ctx.Err())
1280
		case err := <-s.done:
1281
			return fmt.Errorf("llama runner process has terminated: %w", err)
1282
1283
		default:
		}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1284
		if time.Now().After(stallTimer) {
ManniX-ITA's avatar
ManniX-ITA committed
1285
			// timeout
1286
1287
1288
1289
			msg := ""
			if s.status != nil && s.status.LastErrMsg != "" {
				msg = s.status.LastErrMsg
			}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1290
			return fmt.Errorf("timed out waiting for llama runner to start - progress %0.2f - %s", s.loadProgress, msg)
ManniX-ITA's avatar
ManniX-ITA committed
1291
1292
1293
1294
1295
		}
		if s.cmd.ProcessState != nil {
			msg := ""
			if s.status != nil && s.status.LastErrMsg != "" {
				msg = s.status.LastErrMsg
1296
			}
ManniX-ITA's avatar
ManniX-ITA committed
1297
1298
			return fmt.Errorf("llama runner process no longer running: %d %s", s.cmd.ProcessState.ExitCode(), msg)
		}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1299
1300
		ctx, cancel := context.WithTimeout(ctx, 200*time.Millisecond)
		defer cancel()
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1301
		priorProgress := s.loadProgress
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1302
1303
1304
		status, _ := s.getServerStatus(ctx)
		if lastStatus != status && status != ServerStatusReady {
			// Only log on status changes
1305
			slog.Info("waiting for server to become available", "status", status)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1306
		}
ManniX-ITA's avatar
ManniX-ITA committed
1307
1308
		switch status {
		case ServerStatusReady:
Jesse Gross's avatar
Jesse Gross committed
1309
			slog.Info(fmt.Sprintf("llama runner started in %0.2f seconds", time.Since(s.loadStart).Seconds()))
ManniX-ITA's avatar
ManniX-ITA committed
1310
1311
			return nil
		default:
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1312
			lastStatus = status
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1313
1314
1315
1316
			// Reset the timer as long as we're making forward progress on the load
			if priorProgress != s.loadProgress {
				slog.Debug(fmt.Sprintf("model load progress %0.2f", s.loadProgress))
				stallTimer = time.Now().Add(stallDuration)
1317
			} else if !fullyLoaded && int(s.loadProgress*100.0) >= 100 {
1318
				slog.Debug("model load completed, waiting for server to become available", "status", status)
1319
				stallTimer = time.Now().Add(stallDuration)
1320
				fullyLoaded = true
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1321
			}
ManniX-ITA's avatar
ManniX-ITA committed
1322
1323
			time.Sleep(time.Millisecond * 250)
			continue
1324
1325
1326
1327
		}
	}
}

1328
1329
1330
1331
1332
1333
1334
func (s *llmServer) Pid() int {
	if s.cmd != nil && s.cmd.Process != nil {
		return s.cmd.Process.Pid
	}
	return -1
}

1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
func (s *llmServer) GetPort() int {
	return s.port
}

func (s *llmServer) HasExited() bool {
	if s.cmd != nil && s.cmd.ProcessState != nil && s.cmd.ProcessState.ExitCode() >= 0 {
		return true
	}
	return false
}

1346
var grammarJSON = `
1347
1348
1349
1350
root   ::= object
value  ::= object | array | string | number | ("true" | "false" | "null") ws
object ::=
  "{" ws (
1351
         string ":" ws value
1352
    ("," ws string ":" ws value)*
1353
  )? ws "}" 
1354
1355
1356
1357
array  ::=
  "[" ws (
            value
    ("," ws value)*
1358
  )? ws "]" 
1359
1360
string ::=
  "\"" (
1361
    [^"\\\x7F\x00-\x1F] |
1362
    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) # escapes
1363
1364
  )* "\"" 
number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? 
1365
1366
1367
1368
1369
1370
1371
# Optional space: by convention, applied in this grammar after literal chars when allowed
ws ::= ([ \t\n] ws)?
`

const maxBufferSize = 512 * format.KiloByte

type ImageData struct {
1372
1373
	Data []byte `json:"data"`
	ID   int    `json:"id"`
1374
1375
1376
1377
}

type CompletionRequest struct {
	Prompt  string
1378
	Format  json.RawMessage
1379
	Images  []ImageData
Michael Yang's avatar
Michael Yang committed
1380
	Options *api.Options
1381

1382
1383
1384
	Grammar  string // set before sending the request to the subprocess
	Shift    bool
	Truncate bool
1385
1386
}

1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
// DoneReason represents the reason why a completion response is done
type DoneReason int

const (
	// DoneReasonStop indicates the completion stopped naturally
	DoneReasonStop DoneReason = iota
	// DoneReasonLength indicates the completion stopped due to length limits
	DoneReasonLength
	// DoneReasonConnectionClosed indicates the completion stopped due to the connection being closed
	DoneReasonConnectionClosed
)

func (d DoneReason) String() string {
	switch d {
	case DoneReasonLength:
		return "length"
	case DoneReasonStop:
		return "stop"
	default:
		return "" // closed
	}
}

1410
type CompletionResponse struct {
1411
1412
1413
1414
1415
1416
1417
	Content            string        `json:"content"`
	DoneReason         DoneReason    `json:"done_reason"`
	Done               bool          `json:"done"`
	PromptEvalCount    int           `json:"prompt_eval_count"`
	PromptEvalDuration time.Duration `json:"prompt_eval_duration"`
	EvalCount          int           `json:"eval_count"`
	EvalDuration       time.Duration `json:"eval_duration"`
1418
1419
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
1420
func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error {
1421
	slog.Debug("completion request", "images", len(req.Images), "prompt", len(req.Prompt), "format", string(req.Format))
1422
	logutil.Trace("completion request", "prompt", req.Prompt)
1423

1424
	if len(req.Format) > 0 {
1425
1426
1427
1428
1429
1430
		switch string(req.Format) {
		case `null`, `""`:
			// Field was set, but "missing" a value. We accept
			// these as "not set".
			break
		case `"json"`:
1431
			req.Grammar = grammarJSON
1432
1433
1434
1435
		default:
			if req.Format[0] != '{' {
				return fmt.Errorf("invalid format: %q; expected \"json\" or a valid JSON Schema object", req.Format)
			}
1436

1437
1438
1439
1440
			// User provided a JSON schema
			g := llama.SchemaToGrammar(req.Format)
			if g == nil {
				return fmt.Errorf("invalid JSON schema in format")
1441
			}
1442
			req.Grammar = string(g)
1443
1444
1445
		}
	}

1446
1447
1448
1449
1450
	if req.Options == nil {
		opts := api.DefaultOptions()
		req.Options = &opts
	}

1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
	if err := s.sem.Acquire(ctx, 1); err != nil {
		if errors.Is(err, context.Canceled) {
			slog.Info("aborting completion request due to client closing the connection")
		} else {
			slog.Error("Failed to acquire semaphore", "error", err)
		}
		return err
	}
	defer s.sem.Release(1)

	// put an upper limit on num_predict to avoid the model running on forever
	if req.Options.NumPredict < 0 || req.Options.NumPredict > 10*s.options.NumCtx {
		req.Options.NumPredict = 10 * s.options.NumCtx
	}

1466
	// Make sure the server is ready
1467
	status, err := s.getServerStatusRetry(ctx)
1468
1469
1470
	if err != nil {
		return err
	} else if status != ServerStatusReady {
1471
		return fmt.Errorf("unexpected server status: %s", status)
1472
1473
	}

1474
1475
1476
1477
	// Handling JSON marshaling with special characters unescaped.
	buffer := &bytes.Buffer{}
	enc := json.NewEncoder(buffer)
	enc.SetEscapeHTML(false)
1478

1479
	if err := enc.Encode(req); err != nil {
1480
1481
		return fmt.Errorf("failed to marshal data: %v", err)
	}
1482

1483
1484
1485
1486
1487
1488
	endpoint := fmt.Sprintf("http://127.0.0.1:%d/completion", s.port)
	serverReq, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, buffer)
	if err != nil {
		return fmt.Errorf("error creating POST request: %v", err)
	}
	serverReq.Header.Set("Content-Type", "application/json")
1489

1490
	res, err := http.DefaultClient.Do(serverReq)
1491
1492
1493
1494
	if err != nil && errors.Is(err, context.Canceled) {
		// client closed connection
		return err
	} else if err != nil {
1495
1496
		slog.Error("post predict", "error", err)
		return errors.New("model runner has unexpectedly stopped, this may be due to resource limitations or an internal error, check ollama server logs for details")
1497
1498
	}
	defer res.Body.Close()
1499

1500
1501
	if res.StatusCode >= 400 {
		bodyBytes, err := io.ReadAll(res.Body)
1502
		if err != nil {
1503
			return fmt.Errorf("failed reading llm error response: %w", err)
1504
		}
1505
		log.Printf("llm predict error: %s", bodyBytes)
1506
		return api.StatusError{StatusCode: res.StatusCode, ErrorMessage: strings.TrimSpace(string(bodyBytes))}
1507
	}
1508

1509
1510
1511
	scanner := bufio.NewScanner(res.Body)
	buf := make([]byte, 0, maxBufferSize)
	scanner.Buffer(buf, maxBufferSize)
1512

1513
1514
1515
	// keep track of the last token generated, this is used to abort if the model starts looping
	var lastToken string
	var tokenRepeat int
1516

1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
	for scanner.Scan() {
		select {
		case <-ctx.Done():
			// This handles the request cancellation
			return ctx.Err()
		default:
			line := scanner.Bytes()
			if len(line) == 0 {
				continue
			}
1527

1528
1529
			evt, ok := bytes.CutPrefix(line, []byte("data: "))
			if !ok {
1530
				evt = line
1531
			}
1532

1533
			var c CompletionResponse
1534
			if err := json.Unmarshal(evt, &c); err != nil {
1535
				return fmt.Errorf("error unmarshalling llm prediction response: %v", err)
1536
1537
			}
			switch {
1538
			case strings.TrimSpace(c.Content) == lastToken:
1539
1540
1541
1542
1543
				tokenRepeat++
			default:
				lastToken = strings.TrimSpace(c.Content)
				tokenRepeat = 0
			}
1544

1545
1546
1547
1548
1549
			// 30 picked as an arbitrary max token repeat limit, modify as needed
			if tokenRepeat > 30 {
				slog.Debug("prediction aborted, token repeat limit reached")
				return ctx.Err()
			}
1550

1551
1552
1553
1554
			if c.Content != "" {
				fn(CompletionResponse{
					Content: c.Content,
				})
1555
			}
1556

1557
			if c.Done {
1558
				fn(c)
1559
				return nil
1560
			}
1561
		}
1562
	}
1563

1564
	if err := scanner.Err(); err != nil {
1565
		if strings.Contains(err.Error(), "unexpected EOF") || strings.Contains(err.Error(), "forcibly closed") {
1566
			s.Close()
1567
			var msg string
1568
1569
			if s.status != nil && s.status.LastErrMsg != "" {
				msg = s.status.LastErrMsg
1570
1571
			} else {
				msg = err.Error()
1572
			}
1573
			return fmt.Errorf("an error was encountered while running the model: %s", msg)
1574
1575
		}

1576
		return fmt.Errorf("error reading llm response: %v", err)
1577
1578
	}

1579
	return nil
1580
1581
}

1582
1583
type EmbeddingRequest struct {
	Content string `json:"content"`
1584
1585
}

1586
1587
type EmbeddingResponse struct {
	Embedding []float32 `json:"embedding"`
1588
1589
}

1590
func (s *llmServer) Embedding(ctx context.Context, input string) ([]float32, error) {
1591
	logutil.Trace("embedding request", "input", input)
1592

1593
	if err := s.sem.Acquire(ctx, 1); err != nil {
1594
1595
1596
1597
1598
		if errors.Is(err, context.Canceled) {
			slog.Info("aborting embedding request due to client closing the connection")
		} else {
			slog.Error("Failed to acquire semaphore", "error", err)
		}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1599
1600
		return nil, err
	}
1601
	defer s.sem.Release(1)
1602

1603
	// Make sure the server is ready
1604
	status, err := s.getServerStatusRetry(ctx)
1605
1606
1607
	if err != nil {
		return nil, err
	} else if status != ServerStatusReady {
1608
		return nil, fmt.Errorf("unexpected server status: %s", status)
1609
1610
	}

1611
	data, err := json.Marshal(EmbeddingRequest{Content: input})
Michael Yang's avatar
Michael Yang committed
1612
	if err != nil {
1613
1614
1615
		return nil, fmt.Errorf("error marshaling embed data: %w", err)
	}

1616
	r, err := http.NewRequestWithContext(ctx, http.MethodPost, fmt.Sprintf("http://127.0.0.1:%d/embedding", s.port), bytes.NewBuffer(data))
1617
1618
1619
	if err != nil {
		return nil, fmt.Errorf("error creating embed request: %w", err)
	}
1620
	r.Header.Set("Content-Type", "application/json")
1621

1622
	resp, err := http.DefaultClient.Do(r)
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
	if err != nil {
		return nil, fmt.Errorf("do embedding request: %w", err)
	}
	defer resp.Body.Close()

	body, err := io.ReadAll(resp.Body)
	if err != nil {
		return nil, fmt.Errorf("error reading embed response: %w", err)
	}

	if resp.StatusCode >= 400 {
1634
		log.Printf("llm embedding error: %s", body)
1635
1636
1637
		return nil, fmt.Errorf("%s", body)
	}

1638
	var e EmbeddingResponse
1639
	if err := json.Unmarshal(body, &e); err != nil {
1640
1641
1642
		return nil, fmt.Errorf("unmarshal tokenize response: %w", err)
	}

1643
	return e.Embedding, nil
1644
1645
}

Michael Yang's avatar
Michael Yang committed
1646
1647
1648
1649
1650
1651
1652
1653
type TokenizeRequest struct {
	Content string `json:"content"`
}

type TokenizeResponse struct {
	Tokens []int `json:"tokens"`
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
1654
func (s *llmServer) Tokenize(ctx context.Context, content string) ([]int, error) {
1655
1656
	s.llamaModelLock.Lock()
	defer s.llamaModelLock.Unlock()
1657

1658
1659
	if s.llamaModel != nil {
		return s.llamaModel.Tokenize(content, false, true)
Michael Yang's avatar
Michael Yang committed
1660
	}
1661
	if s.textProcessor != nil {
1662
		tokens, err := s.textProcessor.Encode(content, false)
1663
1664
		if err != nil {
			return nil, err
1665
		}
1666
1667
1668
1669
1670
		toks := make([]int, len(tokens))
		for i, t := range tokens {
			toks[i] = int(t)
		}
		return toks, nil
Michael Yang's avatar
Michael Yang committed
1671
	}
1672
1673
	// not reached
	return nil, fmt.Errorf("no tokenizer configured")
Michael Yang's avatar
Michael Yang committed
1674
1675
1676
1677
1678
1679
1680
1681
}

type DetokenizeRequest struct {
	Tokens []int `json:"tokens"`
}

type DetokenizeResponse struct {
	Content string `json:"content"`
1682
1683
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
1684
func (s *llmServer) Detokenize(ctx context.Context, tokens []int) (string, error) {
1685
1686
1687
1688
	s.llamaModelLock.Lock()
	defer s.llamaModelLock.Unlock()

	if s.llamaModel != nil {
1689
1690
		var resp string
		for _, token := range tokens {
1691
			resp += s.llamaModel.TokenToPiece(token)
1692
1693
1694
		}
		return resp, nil
	}
1695
1696
1697
1698
	if s.textProcessor != nil {
		toks := make([]int32, len(tokens))
		for i, t := range tokens {
			toks[i] = int32(t)
1699
		}
1700
1701
1702
		content, err := s.textProcessor.Decode(toks)
		if err != nil {
			return "", err
1703
		}
1704
		return content, nil
Michael Yang's avatar
Michael Yang committed
1705
	}
1706
1707
	// not reached
	return "", fmt.Errorf("no tokenizer configured")
1708
1709
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
1710
func (s *llmServer) Close() error {
1711
1712
1713
1714
	s.llamaModelLock.Lock()
	if s.llamaModel != nil {
		llama.FreeModel(s.llamaModel)
		s.llamaModel = nil
1715
	}
1716
	s.llamaModelLock.Unlock()
1717

1718
	if s.cmd != nil {
1719
		slog.Debug("stopping llama server", "pid", s.Pid())
1720
1721
1722
		if err := s.cmd.Process.Kill(); err != nil {
			return err
		}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1723
1724
		// if ProcessState is already populated, Wait already completed, no need to wait again
		if s.cmd.ProcessState == nil {
1725
			slog.Debug("waiting for llama server to exit", "pid", s.Pid())
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1726
1727
			<-s.done
		}
1728

1729
		slog.Debug("llama server stopped", "pid", s.Pid())
1730
1731
1732
1733
1734
	}

	return nil
}

Jesse Gross's avatar
Jesse Gross committed
1735
func (s *llamaServer) VRAMSize() uint64 {
1736
	return s.estimate.VRAMSize
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1737
1738
}

Jesse Gross's avatar
Jesse Gross committed
1739
func (s *llamaServer) TotalSize() uint64 {
1740
	return s.estimate.TotalSize
1741
1742
}

1743
func (s *llamaServer) VRAMByGPU(id ml.DeviceID) uint64 {
1744
	for i, gpu := range s.gpus {
1745
		if gpu.DeviceID == id {
1746
1747
1748
			if i < len(s.estimate.GPUSizes) {
				return s.estimate.GPUSizes[i]
			}
1749
1750
1751
1752
		}
	}
	return 0
}
Jesse Gross's avatar
Jesse Gross committed
1753

1754
1755
1756
1757
1758
func (s *llamaServer) GetDeviceInfos(ctx context.Context) []ml.DeviceInfo {
	slog.Debug("llamarunner free vram reporting not supported")
	return nil
}

Jesse Gross's avatar
Jesse Gross committed
1759
1760
1761
1762
1763
1764
1765
1766
func (s *ollamaServer) VRAMSize() uint64 {
	if s.mem == nil {
		return 0
	}

	var mem uint64

	for _, g := range s.mem.GPUs {
1767
		mem += g.Size()
Jesse Gross's avatar
Jesse Gross committed
1768
1769
1770
1771
1772
1773
	}

	// Some elements are always on CPU. However, if we have allocated all layers
	// on the GPU then include the CPU components as well, to represent complete offloading.
	noCPULayers := true
	for i := range s.mem.CPU.Weights {
1774
		if s.mem.CPU.Weights[i] != 0 || s.mem.CPU.Cache[i] != 0 {
Jesse Gross's avatar
Jesse Gross committed
1775
1776
1777
1778
1779
			noCPULayers = false
			break
		}
	}
	if noCPULayers {
1780
1781
		mem += s.mem.InputWeights
		mem += s.mem.CPU.Graph
Jesse Gross's avatar
Jesse Gross committed
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
	}

	return mem
}

func (s *ollamaServer) TotalSize() uint64 {
	if s.mem == nil {
		return 0
	}

1792
1793
	mem := s.mem.InputWeights
	mem += s.mem.CPU.Size()
Jesse Gross's avatar
Jesse Gross committed
1794
	for _, g := range s.mem.GPUs {
1795
		mem += g.Size()
Jesse Gross's avatar
Jesse Gross committed
1796
1797
1798
1799
1800
	}

	return mem
}

1801
func (s *ollamaServer) VRAMByGPU(id ml.DeviceID) uint64 {
Jesse Gross's avatar
Jesse Gross committed
1802
1803
1804
1805
1806
	if s.mem == nil {
		return 0
	}

	for _, g := range s.mem.GPUs {
1807
		if g.DeviceID == id {
1808
			return g.Size()
Jesse Gross's avatar
Jesse Gross committed
1809
1810
1811
1812
1813
		}
	}

	return 0
}
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825

func (s *ollamaServer) GetDeviceInfos(ctx context.Context) []ml.DeviceInfo {
	devices, err := discover.GetDevicesFromRunner(ctx, s)
	if err != nil {
		if s.cmd != nil && s.cmd.ProcessState == nil {
			// Still running but hit an error, log
			slog.Debug("failure refreshing GPU information", "error", err)
		}
		// else no longer running so suppress logging as a failure is expected
	}
	return devices
}