server.go 53.8 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/envconfig"
31
	"github.com/ollama/ollama/format"
Michael Yang's avatar
Michael Yang committed
32
	"github.com/ollama/ollama/fs/ggml"
33
	"github.com/ollama/ollama/llama"
34
	"github.com/ollama/ollama/logutil"
Jesse Gross's avatar
Jesse Gross committed
35
	"github.com/ollama/ollama/ml"
36
	"github.com/ollama/ollama/model"
37
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
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
66
type LlamaServer interface {
Jesse Gross's avatar
Jesse Gross committed
67
	ModelPath() string
68
	Load(ctx context.Context, systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, requireFull bool) ([]ml.DeviceID, error)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
69
70
71
	Ping(ctx context.Context) error
	WaitUntilRunning(ctx context.Context) error
	Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error
72
	Embedding(ctx context.Context, input string) ([]float32, error)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
73
74
75
	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
76
77
	VRAMSize() uint64 // Total VRAM across all GPUs
	TotalSize() uint64
78
	VRAMByGPU(id ml.DeviceID) uint64
79
	Pid() int
80
81
82
	GetPort() int
	GetDeviceInfos(ctx context.Context) []ml.DeviceInfo
	HasExited() bool
Daniel Hiltgen's avatar
Daniel Hiltgen committed
83
84
}

Jesse Gross's avatar
Jesse Gross committed
85
// llmServer is an instance of a runner hosting a single model
Daniel Hiltgen's avatar
Daniel Hiltgen committed
86
type llmServer struct {
87
88
89
90
91
92
	port        int
	cmd         *exec.Cmd
	done        chan error // Channel to signal when the process exits
	status      *StatusWriter
	options     api.Options
	numParallel int
93
	modelPath   string
94

Jesse Gross's avatar
Jesse Gross committed
95
96
	loadRequest LoadRequest // Parameters used to initialize the runner

97
98
99
	// 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
100
	llamaModelLock *sync.Mutex
101
102
103
104

	// 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
105

Jesse Gross's avatar
Jesse Gross committed
106
107
	totalLayers  uint64
	loadStart    time.Time // Record how long it took the model to load
108
	loadProgress float32
Daniel Hiltgen's avatar
Daniel Hiltgen committed
109
110

	sem *semaphore.Weighted
111
112
}

Jesse Gross's avatar
Jesse Gross committed
113
114
115
116
type llamaServer struct {
	llmServer

	ggml     *ggml.GGML
117
	gpus     []ml.DeviceInfo // The set of GPUs covered by the memory estimate
Jesse Gross's avatar
Jesse Gross committed
118
119
120
121
122
123
124
125
126
	estimate MemoryEstimate
}

type ollamaServer struct {
	llmServer

	mem *ml.BackendMemory
}

127
128
129
130
131
// 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
132
func LoadModel(model string, maxArraySize int) (*ggml.GGML, error) {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
133
134
135
136
	if _, err := os.Stat(model); err != nil {
		return nil, err
	}

137
138
139
140
141
142
	f, err := os.Open(model)
	if err != nil {
		return nil, err
	}
	defer f.Close()

143
	ggml, err := ggml.Decode(f, maxArraySize)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
144
145
	return ggml, err
}
146

Daniel Hiltgen's avatar
Daniel Hiltgen committed
147
// NewLlamaServer will run a server for the given GPUs
148
func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath string, f *ggml.GGML, adapters, projectors []string, opts api.Options, numParallel int) (LlamaServer, error) {
Jesse Gross's avatar
Jesse Gross committed
149
150
151
152
	var llamaModel *llama.Model
	var textProcessor model.TextProcessor
	var err error
	if envconfig.NewEngine() || f.KV().OllamaEngineRequired() {
153
154
155
156
157
		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
158
159
160
161
		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)
		}
162
	}
Jesse Gross's avatar
Jesse Gross committed
163
164
165
166
167
	if textProcessor == nil {
		llamaModel, err = llama.LoadModelFromFile(modelPath, llama.ModelParams{VocabOnly: true})
		if err != nil {
			return nil, err
		}
168
169
	}

Jesse Gross's avatar
Jesse Gross committed
170
171
172
173
174
	// 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)
175
176
	}

177
178
	opts.NumBatch = min(opts.NumBatch, opts.NumCtx)

Jesse Gross's avatar
Jesse Gross committed
179
	loadRequest := LoadRequest{LoraPath: adapters, KvSize: opts.NumCtx * numParallel, BatchSize: opts.NumBatch, Parallel: numParallel, MultiUserCache: envconfig.MultiUserCache()}
180

181
	defaultThreads := systemInfo.ThreadCount
Jesse Gross's avatar
Jesse Gross committed
182
183
184
185
	if opts.NumThread > 0 {
		loadRequest.NumThreads = opts.NumThread
	} else if defaultThreads > 0 {
		loadRequest.NumThreads = defaultThreads
186
	}
Michael Yang's avatar
Michael Yang committed
187

Jesse Gross's avatar
Jesse Gross committed
188
	// TODO - NUMA support currently doesn't work properly
189
190

	if opts.MainGPU > 0 {
Jesse Gross's avatar
Jesse Gross committed
191
		loadRequest.MainGPU = opts.MainGPU
192
193
	}

Jesse Gross's avatar
Jesse Gross committed
194
195
	if len(projectors) > 0 && llamaModel != nil {
		loadRequest.ProjectorPath = projectors[0]
196
197
	}

198
199
	fa := envconfig.FlashAttention(f.FlashAttention())

Jesse Gross's avatar
Jesse Gross committed
200
201
	// 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.
202
	if fa && !ml.FlashAttentionSupported(gpus) {
203
204
205
		slog.Warn("flash attention enabled but not supported by gpu")
		fa = false
	}
Sam's avatar
Sam committed
206

Michael Yang's avatar
Michael Yang committed
207
	if fa && !f.SupportsFlashAttention() {
208
209
210
211
		slog.Warn("flash attention enabled but not supported by model")
		fa = false
	}

212
	kvct := strings.ToLower(envconfig.KvCacheType())
213
214
215

	if fa {
		slog.Info("enabling flash attention")
Jesse Gross's avatar
Jesse Gross committed
216
		loadRequest.FlashAttention = true
217
218
219

		// Flash Attention also supports kv cache quantization
		// Enable if the requested and kv cache type is supported by the model
220
		if f.SupportsKVCacheType(kvct) {
Jesse Gross's avatar
Jesse Gross committed
221
			loadRequest.KvCacheType = kvct
222
223
		} else {
			slog.Warn("kv cache type not supported by model", "type", kvct)
Sam's avatar
Sam committed
224
		}
225
226
227
	} else if kvct != "" && kvct != "f16" {
		slog.Warn("quantized kv cache requested but flash attention disabled", "type", kvct)
	}
228

229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
	gpuLibs := ml.LibraryPaths(gpus)
	status := NewStatusWriter(os.Stderr)
	cmd, port, err := StartRunner(
		textProcessor != nil,
		modelPath,
		gpuLibs,
		status,
		ml.GetVisibleDevicesEnv(gpus),
	)

	s := llmServer{
		port:           port,
		cmd:            cmd,
		status:         status,
		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),
Jesse Gross's avatar
Jesse Gross committed
254
255
	}

256
257
258
259
260
261
262
263
264
265
	if err != nil {
		var msg string
		if s.status != nil && s.status.LastErrMsg != "" {
			msg = s.status.LastErrMsg
		}
		err := fmt.Errorf("error starting runner: %v %s", err, msg)
		if llamaModel != nil {
			llama.FreeModel(llamaModel)
		}
		return nil, err
Michael Yang's avatar
Michael Yang committed
266
267
	}

268
269
270
271
272
273
274
275
	// reap subprocess when it exits
	go func() {
		err := s.cmd.Wait()
		// Favor a more detailed message over the process exit status
		if err != nil && s.status != nil && s.status.LastErrMsg != "" {
			slog.Error("llama runner terminated", "error", err)
			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"
Jesse Gross's avatar
Jesse Gross committed
276
			}
277
278
279
			s.done <- errors.New(s.status.LastErrMsg)
		} else {
			s.done <- err
280
		}
281
	}()
282

283
284
285
286
	if textProcessor != nil {
		return &ollamaServer{llmServer: s}, nil
	} else {
		return &llamaServer{llmServer: s, ggml: f}, nil
Michael Yang's avatar
Michael Yang committed
287
	}
288
}
Jesse Gross's avatar
Jesse Gross committed
289

290
291
292
func StartRunner(ollamaEngine bool, modelPath string, gpuLibs []string, out io.Writer, extraEnvs map[string]string) (cmd *exec.Cmd, port int, err error) {
	var exe string
	exe, err = os.Executable()
293
	if err != nil {
294
		return nil, 0, fmt.Errorf("unable to lookup executable path: %w", err)
295
296
297
298
299
300
	}

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

301
302
303
304
305
306
	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()
Jesse Gross's avatar
Jesse Gross committed
307
		}
308
309
310
311
312
313
314
315
316
317
	}
	if port == 0 {
		slog.Debug("ResolveTCPAddr failed, using random port")
		port = rand.Intn(65535-49152) + 49152 // get a random port in the ephemeral range
	}
	params := []string{"runner"}
	if ollamaEngine {
		params = append(params, "--ollama-engine")
	}
	if modelPath != "" {
Jesse Gross's avatar
Jesse Gross committed
318
		params = append(params, "--model", modelPath)
319
320
	}
	params = append(params, "--port", strconv.Itoa(port))
Daniel Hiltgen's avatar
Daniel Hiltgen committed
321

322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
	var pathEnv string
	switch runtime.GOOS {
	case "windows":
		pathEnv = "PATH"
	case "darwin":
		pathEnv = "DYLD_LIBRARY_PATH"
	default:
		pathEnv = "LD_LIBRARY_PATH"
	}

	// Note: we always put our dependency paths first
	// since these are the exact version we compiled/linked against
	libraryPaths := append([]string{}, gpuLibs...)
	if libraryPath, ok := os.LookupEnv(pathEnv); ok {
		libraryPaths = append(libraryPaths, filepath.SplitList(libraryPath)...)
	}

	cmd = exec.Command(exe, params...)

	cmd.Env = os.Environ()
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358

	if out != nil {
		stdout, err := cmd.StdoutPipe()
		if err != nil {
			return nil, 0, fmt.Errorf("failed to spawn server stdout pipe: %w", err)
		}
		stderr, err := cmd.StderrPipe()
		if err != nil {
			return nil, 0, fmt.Errorf("failed to spawn server stderr pipe: %w", err)
		}
		go func() {
			io.Copy(out, stdout) //nolint:errcheck
		}()
		go func() {
			io.Copy(out, stderr) //nolint:errcheck
		}()
	}
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
	cmd.SysProcAttr = LlamaServerSysProcAttr

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

	// Update or add the path variable with our adjusted version
	pathNeeded := true
	ollamaPathNeeded := true
	extraEnvsDone := map[string]bool{}
	for k := range extraEnvs {
		extraEnvsDone[k] = false
	}
	for i := range cmd.Env {
		cmp := strings.SplitN(cmd.Env[i], "=", 2)
		if strings.EqualFold(cmp[0], pathEnv) {
			cmd.Env[i] = pathEnv + "=" + pathEnvVal
			pathNeeded = false
		} else if strings.EqualFold(cmp[0], "OLLAMA_LIBRARY_PATH") {
			cmd.Env[i] = "OLLAMA_LIBRARY_PATH=" + strings.Join(gpuLibs, string(filepath.ListSeparator))
			ollamaPathNeeded = false
		} else if len(extraEnvs) != 0 {
			for k, v := range extraEnvs {
				if strings.EqualFold(cmp[0], k) {
					cmd.Env[i] = k + "=" + v
					extraEnvsDone[k] = true
Daniel Hiltgen's avatar
Daniel Hiltgen committed
384
				}
385
			}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
386
		}
387
388
389
390
391
392
393
394
395
396
	}
	if pathNeeded {
		cmd.Env = append(cmd.Env, pathEnv+"="+pathEnvVal)
	}
	if ollamaPathNeeded {
		cmd.Env = append(cmd.Env, "OLLAMA_LIBRARY_PATH="+strings.Join(gpuLibs, string(filepath.ListSeparator)))
	}
	for k, done := range extraEnvsDone {
		if !done {
			cmd.Env = append(cmd.Env, k+"="+extraEnvs[k])
397
		}
398
	}
399

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

403
404
	if err = cmd.Start(); err != nil {
		return nil, 0, err
Jesse Gross's avatar
Jesse Gross committed
405
	}
406
407
	err = nil
	return
Jesse Gross's avatar
Jesse Gross committed
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
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
}

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")

466
467
468
469
func (s *llamaServer) Load(ctx context.Context, systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, requireFull bool) ([]ml.DeviceID, error) {
	systemTotalMemory := systemInfo.TotalMemory
	systemFreeMemory := systemInfo.FreeMemory
	systemSwapFreeMemory := systemInfo.FreeSwap
Jesse Gross's avatar
Jesse Gross committed
470
471
	slog.Info("system memory", "total", format.HumanBytes2(systemTotalMemory), "free", format.HumanBytes2(systemFreeMemory), "free_swap", format.HumanBytes2(systemSwapFreeMemory))

472
473
	if len(gpus) == 0 || s.options.NumGPU == 0 {
		if !verifyCPUFit(s.ggml, s.modelPath, []string{s.loadRequest.ProjectorPath}, s.loadRequest.LoraPath, s.options, systemInfo, s.numParallel) {
474
			slog.Info("model requires more memory than is currently available, evicting a model to make space", "estimate", s.estimate)
475
476
477
478
479
480
481
482
483
484
485
			return nil, fmt.Errorf("model requires more system memory than is currently available %w", ErrLoadRequiredFull)
		}
	} else {
		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 {
				slog.Info("model requires more memory than is currently available, evicting a model to make space", "estimate", s.estimate)
				return nil, ErrLoadRequiredFull
			}
Jesse Gross's avatar
Jesse Gross committed
486
		}
487
		gpus = g
Jesse Gross's avatar
Jesse Gross committed
488
489
490
491
	}

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

492
	if len(gpus) >= 1 {
Jesse Gross's avatar
Jesse Gross committed
493
		switch {
494
495
496
		case s.options.NumGPU == 0:
			gpus = []ml.DeviceInfo{}
		case gpus[0].Library == "Metal" && s.estimate.VRAMSize > systemInfo.TotalMemory:
Jesse Gross's avatar
Jesse Gross committed
497
498
499
			// disable partial offloading when model is greater than total system memory as this
			// can lead to locking up the system
			s.options.NumGPU = 0
500
			gpus = []ml.DeviceInfo{}
501
		case gpus[0].Library != "Metal" && s.estimate.Layers == 0:
Jesse Gross's avatar
Jesse Gross committed
502
			// Don't bother loading into the GPU if no layers can fit
503
504
			gpus = []ml.DeviceInfo{}
		case s.options.NumGPU < 0 && s.estimate.Layers > 0:
Jesse Gross's avatar
Jesse Gross committed
505
506
			s.options.NumGPU = s.estimate.Layers
		}
507
508
	} else {
		s.options.NumGPU = 0
Jesse Gross's avatar
Jesse Gross committed
509
510
511
512
513
514
	}

	// 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
515
		available := systemInfo.FreeMemory + systemInfo.FreeSwap
Jesse Gross's avatar
Jesse Gross committed
516
		if systemMemoryRequired > available {
517
			slog.Warn("model request too large for system", "requested", format.HumanBytes2(systemMemoryRequired), "available", format.HumanBytes2(available), "total", format.HumanBytes2(systemInfo.TotalMemory), "free", format.HumanBytes2(systemInfo.FreeMemory), "swap", format.HumanBytes2(systemInfo.FreeSwap))
518
			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
519
520
521
522
523
524
525
526
527
528
529
530
531
532
		}
	}

	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 {
533
			if g.Library == "Metal" &&
Jesse Gross's avatar
Jesse Gross committed
534
535
536
537
538
539
540
541
542
543
				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
544
545
546
547
		if (runtime.GOOS == "windows" && len(gpus) > 0 && gpus[0].Library == "CUDA" && s.options.UseMMap == nil) ||
			(runtime.GOOS == "linux" && systemInfo.FreeMemory < s.estimate.TotalSize && s.options.UseMMap == nil) ||
			(len(gpus) == 0 && s.options.UseMMap == nil) ||
			(len(gpus) > 0 && gpus[0].Library == "Vulkan" && s.options.UseMMap == nil) ||
Jesse Gross's avatar
Jesse Gross committed
548
549
550
551
552
553
			(s.options.UseMMap != nil && !*s.options.UseMMap) {
			s.loadRequest.UseMmap = false
		}
	}

	if err := s.waitUntilRunnerLaunched(ctx); err != nil {
554
		return nil, err
Jesse Gross's avatar
Jesse Gross committed
555
556
557
558
	}

	resp, err := s.initModel(ctx, s.loadRequest, LoadOperationCommit)
	if err != nil {
559
		return nil, err
Jesse Gross's avatar
Jesse Gross committed
560
561
562
563
564
565
566
567
568
569
	}

	// 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)
570
		return nil, errors.New("failed to allocate memory for model")
Jesse Gross's avatar
Jesse Gross committed
571
572
573
574
575
576
	}

	// 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 {
577
		return uniqueDeviceIDs(s.loadRequest.GPULayers), s.WaitUntilRunning(ctx)
Jesse Gross's avatar
Jesse Gross committed
578
	} else {
579
		return uniqueDeviceIDs(s.loadRequest.GPULayers), nil
Jesse Gross's avatar
Jesse Gross committed
580
581
582
583
584
	}
}

// createGPULayers maps from the tensor splits assigned by the memory estimates to explicit assignment
// of particular layers onto GPUs
585
586
func createGPULayers(estimate MemoryEstimate, ggml *ggml.GGML, gpus []ml.DeviceInfo, numGPU int) ml.GPULayersList {
	if numGPU <= 0 || len(gpus) == 0 {
Jesse Gross's avatar
Jesse Gross committed
587
		return nil
588
	}
Jesse Gross's avatar
Jesse Gross committed
589
590
591

	gpuLayers := make(ml.GPULayersList, len(gpus))
	for i := range gpuLayers {
592
		gpuLayers[i].DeviceID = gpus[i].DeviceID
Jesse Gross's avatar
Jesse Gross committed
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
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
	}

	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.
640
641
//
// Returns the list of GPU IDs that were used in the final allocation on success
642
func (s *ollamaServer) Load(ctx context.Context, systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, requireFull bool) ([]ml.DeviceID, error) {
Jesse Gross's avatar
Jesse Gross committed
643
644
645
646
647
	var success bool
	defer func() {
		if !success {
			s.initModel(ctx, LoadRequest{}, LoadOperationClose)
		}
648
649
650
		if s.mem != nil {
			s.mem.Log(slog.LevelInfo)
		}
Jesse Gross's avatar
Jesse Gross committed
651
652
653
654
	}()

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

655
656
657
	systemTotalMemory := systemInfo.TotalMemory
	systemFreeMemory := systemInfo.FreeMemory
	systemSwapFreeMemory := systemInfo.FreeSwap
Jesse Gross's avatar
Jesse Gross committed
658
659
	slog.Info("system memory", "total", format.HumanBytes2(systemTotalMemory), "free", format.HumanBytes2(systemFreeMemory), "free_swap", format.HumanBytes2(systemSwapFreeMemory))

660
661
662
663
	for _, gpu := range gpus {
		available := gpu.FreeMemory - envconfig.GpuOverhead() - gpu.MinimumMemory()
		if gpu.FreeMemory < envconfig.GpuOverhead()+gpu.MinimumMemory() {
			available = 0
Jesse Gross's avatar
Jesse Gross committed
664
		}
665
666
667
668
669
		slog.Info("gpu memory", "id", gpu.ID, "library", gpu.Library,
			"available", format.HumanBytes2(available),
			"free", format.HumanBytes2(gpu.FreeMemory),
			"minimum", format.HumanBytes2(gpu.MinimumMemory()),
			"overhead", format.HumanBytes2(envconfig.GpuOverhead()))
Jesse Gross's avatar
Jesse Gross committed
670
671
672
673
674
675
676
	}

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

	gpuLayers, err := s.createLayout(systemInfo, gpus, s.mem, requireFull, backoff)
	if err != nil {
677
		return nil, err
Jesse Gross's avatar
Jesse Gross committed
678
679
680
	}

	if err := s.waitUntilRunnerLaunched(ctx); err != nil {
681
		return nil, err
Jesse Gross's avatar
Jesse Gross committed
682
683
684
685
686
687
688
689
690
	}

nextOperation:
	for operation := LoadOperationFit; operation < LoadOperationCommit; operation++ {
	nextLoad:
		for {
			s.loadRequest.GPULayers = gpuLayers
			resp, err := s.initModel(ctx, s.loadRequest, operation)
			if err != nil {
691
				return nil, err
Jesse Gross's avatar
Jesse Gross committed
692
693
694
695
696
697
698
699
700
701
702
			}

			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 {
703
					return nil, err
Jesse Gross's avatar
Jesse Gross committed
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
				}

				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 {
737
							return nil, err
Jesse Gross's avatar
Jesse Gross committed
738
739
740
741
742
743
						}
						slog.Debug("new layout created", "layers", newGPULayers)

						s.loadRequest.GPULayers = newGPULayers
						resp, err = s.initModel(ctx, s.loadRequest, operation)
						if err != nil {
744
							return nil, err
Jesse Gross's avatar
Jesse Gross committed
745
746
747
748
749
750
751
752
						}

						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 {
753
								return nil, err
Jesse Gross's avatar
Jesse Gross committed
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
							}

							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 {
778
					return nil, fmt.Errorf("memory layout cannot be allocated with num_gpu = %v", s.options.NumGPU)
Jesse Gross's avatar
Jesse Gross committed
779
780
781
782
783
				}

				// 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
784
				// time. Apply a backoff to try to find the real amount of available space.
Jesse Gross's avatar
Jesse Gross committed
785
786
				if backoff > 1 {
					slog.Warn("memory layout cannot be allocated", "memory", resp.Memory)
787
					return nil, errors.New("memory layout cannot be allocated")
Jesse Gross's avatar
Jesse Gross committed
788
				} else {
789
					backoff += 0.1
Jesse Gross's avatar
Jesse Gross committed
790
791
792
793
794
795
796
797
798
799
				}

				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 {
800
		return nil, err
Jesse Gross's avatar
Jesse Gross committed
801
802
803
804
805
806
807
	}

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

	if !success {
		slog.Warn("failed to commit memory for model", "memory", resp.Memory)
808
		return nil, errors.New("failed to commit memory for model")
Jesse Gross's avatar
Jesse Gross committed
809
810
	}

811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
	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
829
830
831
832
833
834
835
836
}

// 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
837
func (s *ollamaServer) createLayout(systemInfo ml.SystemInfo, systemGPUs []ml.DeviceInfo, memory *ml.BackendMemory, requireFull bool, backoff float32) (ml.GPULayersList, error) {
Jesse Gross's avatar
Jesse Gross committed
838
839
	if memory == nil {
		memory = &ml.BackendMemory{CPU: ml.DeviceMemory{
840
841
			Weights: make([]uint64, s.totalLayers),
			Cache:   make([]uint64, s.totalLayers),
Jesse Gross's avatar
Jesse Gross committed
842
843
		}}
	}
844
845
846
847
848
849
850
851
852
853
854
855
856
857
	gpuLayers, layers, err := s.buildLayout(systemGPUs, memory, requireFull, backoff)
	if err != nil {
		return nil, err
	}
	err = s.verifyLayout(systemInfo, memory, requireFull, gpuLayers, layers)
	if err != nil {
		return nil, err
	}
	return gpuLayers, nil
}

func (s *ollamaServer) buildLayout(systemGPUs []ml.DeviceInfo, memory *ml.BackendMemory, requireFull bool, backoff float32) (ml.GPULayersList, []uint64, error) {
	gpus := append(make([]ml.DeviceInfo, 0, len(systemGPUs)), systemGPUs...)
	sort.Sort(sort.Reverse(ml.ByFreeMemory(gpus)))
Jesse Gross's avatar
Jesse Gross committed
858
859
860
861

	layers := make([]uint64, len(memory.CPU.Weights))
	for i := range layers {
		for j := range memory.GPUs {
862
863
			layers[i] += memory.GPUs[j].Weights[i]
			layers[i] += memory.GPUs[j].Cache[i]
Jesse Gross's avatar
Jesse Gross committed
864
		}
865
866
		layers[i] += memory.CPU.Weights[i]
		layers[i] += memory.CPU.Cache[i]
867
		logutil.Trace("layer to assign", "layer", i, "size", format.HumanBytes2(layers[i]))
Jesse Gross's avatar
Jesse Gross committed
868
869
870
	}

	gpuLayers := ml.GPULayersList{}
871
	for _, gl := range ml.ByLibrary(gpus) {
Jesse Gross's avatar
Jesse Gross committed
872
873
874
875
876
877
878
879
		// 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 {
880
				if gl[i].DeviceID == memory.GPUs[j].DeviceID {
881
					if memory.GPUs[j].Graph != 0 {
Jesse Gross's avatar
Jesse Gross committed
882
883
884
						lastUsedGPU = i
					}

885
					reserved := uint64(float32(gl[i].FreeMemory)*backoff) + gl[i].MinimumMemory() + envconfig.GpuOverhead() + memory.GPUs[j].Graph
Jesse Gross's avatar
Jesse Gross committed
886
887
888
889
890
891
					if gl[i].FreeMemory > reserved {
						gl[i].FreeMemory -= reserved
					} else {
						gl[i].FreeMemory = 0
					}

892
					slog.Debug("available gpu", "id", gl[i].ID, "library", gl[i].Library,
Jesse Gross's avatar
Jesse Gross committed
893
						"available layer vram", format.HumanBytes2(gl[i].FreeMemory),
894
						"backoff", fmt.Sprintf("%.2f", backoff), "minimum", format.HumanBytes2(gl[i].MinimumMemory()),
Jesse Gross's avatar
Jesse Gross committed
895
						"overhead", format.HumanBytes2(envconfig.GpuOverhead()),
896
						"graph", format.HumanBytes2(memory.GPUs[j].Graph))
Jesse Gross's avatar
Jesse Gross committed
897
898
899
900
901
902
903
904
905
906
907

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

908
		libraryGpuLayers := assignLayers(layers, gl, requireFull, s.options.NumGPU, lastUsedGPU)
Jesse Gross's avatar
Jesse Gross committed
909
910
911
912
		if libraryGpuLayers.Sum() > gpuLayers.Sum() {
			gpuLayers = libraryGpuLayers
		}
	}
913
914
	return gpuLayers, layers, nil
}
Jesse Gross's avatar
Jesse Gross committed
915

916
917
// verifyLayout ensures that we don't exceed limits, such as requirements about partial offloading or system memory
func (s *ollamaServer) verifyLayout(systemInfo ml.SystemInfo, memory *ml.BackendMemory, requireFull bool, gpuLayers ml.GPULayersList, layers []uint64) error {
Jesse Gross's avatar
Jesse Gross committed
918
	// These sizes will only increase as we go through additional iterations and get additional information.
919
	cpuSize := memory.InputWeights + memory.CPU.Graph
Jesse Gross's avatar
Jesse Gross committed
920
921
922
	var vramSize uint64
	for _, gl := range gpuLayers {
		for _, gpu := range memory.GPUs {
923
			if gl.DeviceID == gpu.DeviceID {
924
				vramSize += gpu.Graph
Jesse Gross's avatar
Jesse Gross committed
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
				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) {
945
			return ErrLoadRequiredFull
Jesse Gross's avatar
Jesse Gross committed
946
947
		}

948
949
		if cpuSize > systemInfo.FreeMemory {
			return ErrLoadRequiredFull
Jesse Gross's avatar
Jesse Gross committed
950
951
952
953
954
955
		}
	}

	// 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" {
956
		available := systemInfo.FreeMemory + systemInfo.FreeSwap
Jesse Gross's avatar
Jesse Gross committed
957
		if cpuSize > available {
958
959
			slog.Warn("model request too large for system", "requested", format.HumanBytes2(cpuSize), "available", format.HumanBytes2(available), "total", format.HumanBytes2(systemInfo.TotalMemory), "free", format.HumanBytes2(systemInfo.FreeMemory), "swap", format.HumanBytes2(systemInfo.FreeSwap))
			return fmt.Errorf("model requires more system memory (%s) than is available (%s)", format.HumanBytes2(cpuSize), format.HumanBytes2(available))
Jesse Gross's avatar
Jesse Gross committed
960
961
		}
	} else {
962
		if vramSize > systemInfo.TotalMemory {
Jesse Gross's avatar
Jesse Gross committed
963
964
965
966
967
968
969
970
971
972
973
			// 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")
	}

974
	return nil
Jesse Gross's avatar
Jesse Gross committed
975
976
977
}

// assignLayers packs the maximum number of layers onto the smallest set of GPUs and comes up with a layer assignment
978
func assignLayers(layers []uint64, gpus []ml.DeviceInfo, requireFull bool, requestedLayers int, lastUsedGPU int) (gpuLayers ml.GPULayersList) {
Jesse Gross's avatar
Jesse Gross committed
979
980
981
982
983
984
985
986
	// 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
987
				forceRequest := i == len(gpus)-1 && !requireFull
Jesse Gross's avatar
Jesse Gross committed
988
989
990
991
992
993
				gpuLayers = findBestFit(layers, gpus[:i+1], requestedLayers, forceRequest)
				if gpuLayers.Sum() == len(layers) || gpuLayers.Sum() == requestedLayers {
					break
				}
			}
		} else {
994
			gpuLayers = findBestFit(layers, gpus, requestedLayers, !requireFull)
Jesse Gross's avatar
Jesse Gross committed
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
		}

		// 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.
1012
func findBestFit(layers []uint64, gpus []ml.DeviceInfo, requestedLayers int, forceRequest bool) (gpuLayers ml.GPULayersList) {
Jesse Gross's avatar
Jesse Gross committed
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
	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
1041
func greedyFit(layers []uint64, gpus []ml.DeviceInfo, capacity float32, requestedLayers int) (gpuLayers ml.GPULayersList) {
Jesse Gross's avatar
Jesse Gross committed
1042
	device := len(gpus) - 1
1043
	gpuLayers = ml.GPULayersList{{DeviceID: gpus[device].DeviceID}}
Jesse Gross's avatar
Jesse Gross committed
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
	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
			}
1061
			gpuLayers = append(ml.GPULayersList{{DeviceID: gpus[device].DeviceID}}, gpuLayers...)
Jesse Gross's avatar
Jesse Gross committed
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
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
			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
1127
1128
1129
1130
1131
1132
}

type ServerStatus int

const ( // iota is reset to 0
	ServerStatusReady ServerStatus = iota
1133
	ServerStatusNoSlotsAvailable
Jesse Gross's avatar
Jesse Gross committed
1134
	ServerStatusLaunched
1135
1136
1137
1138
1139
	ServerStatusLoadingModel
	ServerStatusNotResponding
	ServerStatusError
)

1140
func (s ServerStatus) String() string {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1141
1142
1143
	switch s {
	case ServerStatusReady:
		return "llm server ready"
1144
	case ServerStatusNoSlotsAvailable:
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1145
		return "llm busy - no slots available"
Jesse Gross's avatar
Jesse Gross committed
1146
1147
	case ServerStatusLaunched:
		return "llm server launched"
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1148
1149
1150
1151
1152
1153
1154
1155
1156
	case ServerStatusLoadingModel:
		return "llm server loading model"
	case ServerStatusNotResponding:
		return "llm server not responding"
	default:
		return "llm server error"
	}
}

1157
1158
1159
type ServerStatusResponse struct {
	Status   ServerStatus `json:"status"`
	Progress float32      `json:"progress"`
1160
1161
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
1162
func (s *llmServer) getServerStatus(ctx context.Context) (ServerStatus, error) {
1163
1164
1165
1166
1167
1168
	// Fail fast if its exited
	if s.cmd.ProcessState != nil {
		msg := ""
		if s.status != nil && s.status.LastErrMsg != "" {
			msg = s.status.LastErrMsg
		}
1169
1170
		if s.cmd.ProcessState.ExitCode() == -1 {
			// Most likely a signal killed it, log some more details to try to help troubleshoot
1171
			slog.Warn("llama runner process no longer running", "sys", s.cmd.ProcessState.Sys(), "string", s.cmd.ProcessState)
1172
		}
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
		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
1185
			return ServerStatusNotResponding, errors.New("server not responding")
1186
		}
1187
1188
1189
		if strings.Contains(err.Error(), "connection refused") {
			return ServerStatusNotResponding, errors.New("connection refused")
		}
1190
1191
1192
1193
1194
1195
1196
1197
1198
		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)
	}

1199
1200
	var ssr ServerStatusResponse
	if err := json.Unmarshal(body, &ssr); err != nil {
1201
1202
1203
		return ServerStatusError, fmt.Errorf("health unmarshal encode response: %w", err)
	}

1204
1205
1206
1207
	switch ssr.Status {
	case ServerStatusLoadingModel:
		s.loadProgress = ssr.Progress
		return ssr.Status, nil
Jesse Gross's avatar
Jesse Gross committed
1208
	case ServerStatusLaunched, ServerStatusReady, ServerStatusNoSlotsAvailable:
1209
		return ssr.Status, nil
1210
	default:
1211
		return ssr.Status, fmt.Errorf("server error: %+v", ssr)
1212
1213
1214
	}
}

1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
// 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
1238
func (s *llmServer) Ping(ctx context.Context) error {
1239
1240
1241
1242
1243
1244
1245
1246
	_, err := s.getServerStatus(ctx)
	if err != nil {
		slog.Debug("server unhealthy", "error", err)
		return err
	}
	return nil
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
1247
func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
1248
	stallDuration := envconfig.LoadTimeout()    // If no progress happens
1249
	stallTimer := time.Now().Add(stallDuration) // give up if we stall
1250
1251
1252

	slog.Info("waiting for llama runner to start responding")
	var lastStatus ServerStatus = -1
1253
	fullyLoaded := false
ManniX-ITA's avatar
ManniX-ITA committed
1254

1255
1256
	for {
		select {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1257
		case <-ctx.Done():
1258
			slog.Warn("client connection closed before server finished loading, aborting load")
1259
			return fmt.Errorf("timed out waiting for llama runner to start: %w", ctx.Err())
1260
		case err := <-s.done:
1261
			return fmt.Errorf("llama runner process has terminated: %w", err)
1262
1263
		default:
		}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1264
		if time.Now().After(stallTimer) {
ManniX-ITA's avatar
ManniX-ITA committed
1265
			// timeout
1266
1267
1268
1269
			msg := ""
			if s.status != nil && s.status.LastErrMsg != "" {
				msg = s.status.LastErrMsg
			}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1270
			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
1271
1272
1273
1274
1275
		}
		if s.cmd.ProcessState != nil {
			msg := ""
			if s.status != nil && s.status.LastErrMsg != "" {
				msg = s.status.LastErrMsg
1276
			}
ManniX-ITA's avatar
ManniX-ITA committed
1277
1278
			return fmt.Errorf("llama runner process no longer running: %d %s", s.cmd.ProcessState.ExitCode(), msg)
		}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1279
1280
		ctx, cancel := context.WithTimeout(ctx, 200*time.Millisecond)
		defer cancel()
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1281
		priorProgress := s.loadProgress
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1282
1283
1284
		status, _ := s.getServerStatus(ctx)
		if lastStatus != status && status != ServerStatusReady {
			// Only log on status changes
1285
			slog.Info("waiting for server to become available", "status", status)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1286
		}
ManniX-ITA's avatar
ManniX-ITA committed
1287
1288
		switch status {
		case ServerStatusReady:
Jesse Gross's avatar
Jesse Gross committed
1289
			slog.Info(fmt.Sprintf("llama runner started in %0.2f seconds", time.Since(s.loadStart).Seconds()))
ManniX-ITA's avatar
ManniX-ITA committed
1290
1291
			return nil
		default:
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1292
			lastStatus = status
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1293
1294
1295
1296
			// 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)
1297
			} else if !fullyLoaded && int(s.loadProgress*100.0) >= 100 {
1298
				slog.Debug("model load completed, waiting for server to become available", "status", status)
1299
				stallTimer = time.Now().Add(stallDuration)
1300
				fullyLoaded = true
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1301
			}
ManniX-ITA's avatar
ManniX-ITA committed
1302
1303
			time.Sleep(time.Millisecond * 250)
			continue
1304
1305
1306
1307
		}
	}
}

1308
1309
1310
1311
1312
1313
1314
func (s *llmServer) Pid() int {
	if s.cmd != nil && s.cmd.Process != nil {
		return s.cmd.Process.Pid
	}
	return -1
}

1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
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
}

1326
var grammarJSON = `
1327
1328
1329
1330
root   ::= object
value  ::= object | array | string | number | ("true" | "false" | "null") ws
object ::=
  "{" ws (
1331
         string ":" ws value
1332
    ("," ws string ":" ws value)*
1333
  )? ws "}" 
1334
1335
1336
1337
array  ::=
  "[" ws (
            value
    ("," ws value)*
1338
  )? ws "]" 
1339
1340
string ::=
  "\"" (
1341
    [^"\\\x7F\x00-\x1F] |
1342
    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) # escapes
1343
1344
  )* "\"" 
number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? 
1345
1346
1347
1348
1349
1350
1351
# 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 {
1352
1353
	Data []byte `json:"data"`
	ID   int    `json:"id"`
1354
1355
1356
1357
}

type CompletionRequest struct {
	Prompt  string
1358
	Format  json.RawMessage
1359
	Images  []ImageData
Michael Yang's avatar
Michael Yang committed
1360
	Options *api.Options
1361

1362
1363
1364
	Grammar  string // set before sending the request to the subprocess
	Shift    bool
	Truncate bool
1365
1366
1367
1368
1369
1370

	// Logprobs specifies whether to include log probabilities in the response
	Logprobs bool

	// TopLogprobs specifies the number of most likely alternative tokens to return (0-20)
	TopLogprobs int
1371
1372
}

1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
// 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
	}
}

1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
// TokenLogprob represents log probability information for a single token alternative.
type TokenLogprob struct {
	Token   string  `json:"token"`
	Logprob float64 `json:"logprob"`
}

// Logprob contains log probability information for a generated token.
type Logprob struct {
	TokenLogprob
	TopLogprobs []TokenLogprob `json:"top_logprobs,omitempty"`
}

1408
type CompletionResponse struct {
1409
1410
1411
1412
1413
1414
1415
	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"`
1416
1417
1418

	// Logprobs contains log probability information if requested
	Logprobs []Logprob `json:"logprobs,omitempty"`
1419
1420
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
1421
func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error {
1422
	slog.Debug("completion request", "images", len(req.Images), "prompt", len(req.Prompt), "format", string(req.Format))
1423
	logutil.Trace("completion request", "prompt", req.Prompt)
1424

1425
	if len(req.Format) > 0 {
1426
1427
1428
1429
1430
1431
		switch string(req.Format) {
		case `null`, `""`:
			// Field was set, but "missing" a value. We accept
			// these as "not set".
			break
		case `"json"`:
1432
			req.Grammar = grammarJSON
1433
1434
1435
1436
		default:
			if req.Format[0] != '{' {
				return fmt.Errorf("invalid format: %q; expected \"json\" or a valid JSON Schema object", req.Format)
			}
1437

1438
1439
1440
1441
			// User provided a JSON schema
			g := llama.SchemaToGrammar(req.Format)
			if g == nil {
				return fmt.Errorf("invalid JSON schema in format")
1442
			}
1443
			req.Grammar = string(g)
1444
1445
1446
		}
	}

1447
1448
1449
1450
1451
	if req.Options == nil {
		opts := api.DefaultOptions()
		req.Options = &opts
	}

1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
	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
	}

1467
	// Make sure the server is ready
1468
	status, err := s.getServerStatusRetry(ctx)
1469
1470
1471
	if err != nil {
		return err
	} else if status != ServerStatusReady {
1472
		return fmt.Errorf("unexpected server status: %s", status)
1473
1474
	}

1475
1476
1477
1478
	// Handling JSON marshaling with special characters unescaped.
	buffer := &bytes.Buffer{}
	enc := json.NewEncoder(buffer)
	enc.SetEscapeHTML(false)
1479

1480
	if err := enc.Encode(req); err != nil {
1481
1482
		return fmt.Errorf("failed to marshal data: %v", err)
	}
1483

1484
1485
1486
1487
1488
1489
	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")
1490

1491
	res, err := http.DefaultClient.Do(serverReq)
1492
1493
1494
1495
	if err != nil && errors.Is(err, context.Canceled) {
		// client closed connection
		return err
	} else if err != nil {
1496
1497
		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")
1498
1499
	}
	defer res.Body.Close()
1500

1501
1502
	if res.StatusCode >= 400 {
		bodyBytes, err := io.ReadAll(res.Body)
1503
		if err != nil {
1504
			return fmt.Errorf("failed reading llm error response: %w", err)
1505
		}
1506
		log.Printf("llm predict error: %s", bodyBytes)
1507
		return api.StatusError{StatusCode: res.StatusCode, ErrorMessage: strings.TrimSpace(string(bodyBytes))}
1508
	}
1509

1510
1511
1512
	scanner := bufio.NewScanner(res.Body)
	buf := make([]byte, 0, maxBufferSize)
	scanner.Buffer(buf, maxBufferSize)
1513

1514
1515
1516
	// keep track of the last token generated, this is used to abort if the model starts looping
	var lastToken string
	var tokenRepeat int
1517

1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
	for scanner.Scan() {
		select {
		case <-ctx.Done():
			// This handles the request cancellation
			return ctx.Err()
		default:
			line := scanner.Bytes()
			if len(line) == 0 {
				continue
			}
1528

1529
1530
			evt, ok := bytes.CutPrefix(line, []byte("data: "))
			if !ok {
1531
				evt = line
1532
			}
1533

1534
			var c CompletionResponse
1535
			if err := json.Unmarshal(evt, &c); err != nil {
1536
				return fmt.Errorf("error unmarshalling llm prediction response: %v", err)
1537
1538
			}
			switch {
1539
			case strings.TrimSpace(c.Content) == lastToken:
1540
1541
1542
1543
1544
				tokenRepeat++
			default:
				lastToken = strings.TrimSpace(c.Content)
				tokenRepeat = 0
			}
1545

1546
1547
1548
1549
1550
			// 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()
			}
1551

1552
1553
			if c.Content != "" {
				fn(CompletionResponse{
1554
1555
					Content:  c.Content,
					Logprobs: c.Logprobs,
1556
				})
1557
			}
1558

1559
			if c.Done {
1560
				fn(c)
1561
				return nil
1562
			}
1563
		}
1564
	}
1565

1566
	if err := scanner.Err(); err != nil {
1567
		if strings.Contains(err.Error(), "unexpected EOF") || strings.Contains(err.Error(), "forcibly closed") {
1568
			s.Close()
1569
			var msg string
1570
1571
			if s.status != nil && s.status.LastErrMsg != "" {
				msg = s.status.LastErrMsg
1572
1573
			} else {
				msg = err.Error()
1574
			}
1575
			return fmt.Errorf("an error was encountered while running the model: %s", msg)
1576
1577
		}

1578
		return fmt.Errorf("error reading llm response: %v", err)
1579
1580
	}

1581
	return nil
1582
1583
}

1584
type EmbeddingRequest struct {
1585
	Content string `json:"content"`
1586
1587
}

1588
type EmbeddingResponse struct {
1589
	Embedding []float32 `json:"embedding"`
1590
1591
}

1592
func (s *llmServer) Embedding(ctx context.Context, input string) ([]float32, error) {
1593
	logutil.Trace("embedding request", "input", input)
1594

1595
	if err := s.sem.Acquire(ctx, 1); err != nil {
1596
1597
1598
1599
1600
		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)
		}
1601
		return nil, err
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1602
	}
1603
	defer s.sem.Release(1)
1604

1605
	// Make sure the server is ready
1606
	status, err := s.getServerStatusRetry(ctx)
1607
	if err != nil {
1608
		return nil, err
1609
	} else if status != ServerStatusReady {
1610
		return nil, fmt.Errorf("unexpected server status: %s", status)
1611
1612
	}

1613
	data, err := json.Marshal(EmbeddingRequest{Content: input})
Michael Yang's avatar
Michael Yang committed
1614
	if err != nil {
1615
		return nil, fmt.Errorf("error marshaling embed data: %w", err)
1616
1617
	}

1618
	r, err := http.NewRequestWithContext(ctx, http.MethodPost, fmt.Sprintf("http://127.0.0.1:%d/embedding", s.port), bytes.NewBuffer(data))
1619
	if err != nil {
1620
		return nil, fmt.Errorf("error creating embed request: %w", err)
1621
	}
1622
	r.Header.Set("Content-Type", "application/json")
1623

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

	body, err := io.ReadAll(resp.Body)
	if err != nil {
1632
		return nil, fmt.Errorf("error reading embed response: %w", err)
1633
1634
1635
	}

	if resp.StatusCode >= 400 {
1636
		log.Printf("llm embedding error: %s", body)
1637
		return nil, fmt.Errorf("%s", body)
1638
1639
	}

1640
	var e EmbeddingResponse
1641
	if err := json.Unmarshal(body, &e); err != nil {
1642
		return nil, fmt.Errorf("unmarshal tokenize response: %w", err)
1643
1644
	}

1645
	return e.Embedding, nil
1646
1647
}

Michael Yang's avatar
Michael Yang committed
1648
1649
1650
1651
1652
1653
1654
1655
type TokenizeRequest struct {
	Content string `json:"content"`
}

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

Daniel Hiltgen's avatar
Daniel Hiltgen committed
1656
func (s *llmServer) Tokenize(ctx context.Context, content string) ([]int, error) {
1657
1658
	s.llamaModelLock.Lock()
	defer s.llamaModelLock.Unlock()
1659

1660
1661
	if s.llamaModel != nil {
		return s.llamaModel.Tokenize(content, false, true)
Michael Yang's avatar
Michael Yang committed
1662
	}
1663
	if s.textProcessor != nil {
1664
		tokens, err := s.textProcessor.Encode(content, false)
1665
1666
		if err != nil {
			return nil, err
1667
		}
1668
1669
1670
1671
1672
		toks := make([]int, len(tokens))
		for i, t := range tokens {
			toks[i] = int(t)
		}
		return toks, nil
Michael Yang's avatar
Michael Yang committed
1673
	}
1674
1675
	// not reached
	return nil, fmt.Errorf("no tokenizer configured")
Michael Yang's avatar
Michael Yang committed
1676
1677
1678
1679
1680
1681
1682
1683
}

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

type DetokenizeResponse struct {
	Content string `json:"content"`
1684
1685
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
1686
func (s *llmServer) Detokenize(ctx context.Context, tokens []int) (string, error) {
1687
1688
1689
1690
	s.llamaModelLock.Lock()
	defer s.llamaModelLock.Unlock()

	if s.llamaModel != nil {
1691
1692
		var resp string
		for _, token := range tokens {
1693
			resp += s.llamaModel.TokenToPiece(token)
1694
1695
1696
		}
		return resp, nil
	}
1697
1698
1699
1700
	if s.textProcessor != nil {
		toks := make([]int32, len(tokens))
		for i, t := range tokens {
			toks[i] = int32(t)
1701
		}
1702
1703
1704
		content, err := s.textProcessor.Decode(toks)
		if err != nil {
			return "", err
1705
		}
1706
		return content, nil
Michael Yang's avatar
Michael Yang committed
1707
	}
1708
1709
	// not reached
	return "", fmt.Errorf("no tokenizer configured")
1710
1711
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
1712
func (s *llmServer) Close() error {
1713
1714
1715
1716
	s.llamaModelLock.Lock()
	if s.llamaModel != nil {
		llama.FreeModel(s.llamaModel)
		s.llamaModel = nil
1717
	}
1718
	s.llamaModelLock.Unlock()
1719

1720
	if s.cmd != nil {
1721
		slog.Debug("stopping llama server", "pid", s.Pid())
1722
1723
1724
		if err := s.cmd.Process.Kill(); err != nil {
			return err
		}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1725
1726
		// if ProcessState is already populated, Wait already completed, no need to wait again
		if s.cmd.ProcessState == nil {
1727
			slog.Debug("waiting for llama server to exit", "pid", s.Pid())
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1728
1729
			<-s.done
		}
1730

1731
		slog.Debug("llama server stopped", "pid", s.Pid())
1732
1733
1734
1735
1736
	}

	return nil
}

Jesse Gross's avatar
Jesse Gross committed
1737
func (s *llamaServer) VRAMSize() uint64 {
1738
	return s.estimate.VRAMSize
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1739
1740
}

Jesse Gross's avatar
Jesse Gross committed
1741
func (s *llamaServer) TotalSize() uint64 {
1742
	return s.estimate.TotalSize
1743
1744
}

1745
func (s *llamaServer) VRAMByGPU(id ml.DeviceID) uint64 {
1746
	for i, gpu := range s.gpus {
1747
		if gpu.DeviceID == id {
1748
1749
1750
			if i < len(s.estimate.GPUSizes) {
				return s.estimate.GPUSizes[i]
			}
1751
1752
1753
1754
		}
	}
	return 0
}
Jesse Gross's avatar
Jesse Gross committed
1755

1756
1757
1758
1759
1760
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
1761
1762
1763
1764
1765
1766
1767
1768
func (s *ollamaServer) VRAMSize() uint64 {
	if s.mem == nil {
		return 0
	}

	var mem uint64

	for _, g := range s.mem.GPUs {
1769
		mem += g.Size()
Jesse Gross's avatar
Jesse Gross committed
1770
1771
1772
1773
1774
1775
	}

	// 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 {
1776
		if s.mem.CPU.Weights[i] != 0 || s.mem.CPU.Cache[i] != 0 {
Jesse Gross's avatar
Jesse Gross committed
1777
1778
1779
1780
1781
			noCPULayers = false
			break
		}
	}
	if noCPULayers {
1782
1783
		mem += s.mem.InputWeights
		mem += s.mem.CPU.Graph
Jesse Gross's avatar
Jesse Gross committed
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
	}

	return mem
}

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

1794
1795
	mem := s.mem.InputWeights
	mem += s.mem.CPU.Size()
Jesse Gross's avatar
Jesse Gross committed
1796
	for _, g := range s.mem.GPUs {
1797
		mem += g.Size()
Jesse Gross's avatar
Jesse Gross committed
1798
1799
1800
1801
1802
	}

	return mem
}

1803
func (s *ollamaServer) VRAMByGPU(id ml.DeviceID) uint64 {
Jesse Gross's avatar
Jesse Gross committed
1804
1805
1806
1807
1808
	if s.mem == nil {
		return 0
	}

	for _, g := range s.mem.GPUs {
1809
		if g.DeviceID == id {
1810
			return g.Size()
Jesse Gross's avatar
Jesse Gross committed
1811
1812
1813
1814
1815
		}
	}

	return 0
}
1816
1817

func (s *ollamaServer) GetDeviceInfos(ctx context.Context) []ml.DeviceInfo {
1818
	devices, err := ml.GetDevicesFromRunner(ctx, s)
1819
1820
1821
1822
1823
1824
1825
1826
1827
	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
}