routes.go 29.1 KB
Newer Older
Jeffrey Morgan's avatar
Jeffrey Morgan committed
1
2
3
package server

import (
4
	"context"
Michael Yang's avatar
Michael Yang committed
5
	"encoding/json"
6
	"errors"
7
	"fmt"
Jeffrey Morgan's avatar
Jeffrey Morgan committed
8
	"io"
9
	"io/fs"
10
	"log/slog"
Jeffrey Morgan's avatar
Jeffrey Morgan committed
11
12
	"net"
	"net/http"
13
	"os"
14
	"os/signal"
Michael Yang's avatar
Michael Yang committed
15
	"path/filepath"
Jeffrey Morgan's avatar
Jeffrey Morgan committed
16
	"reflect"
17
	"runtime"
Michael Yang's avatar
Michael Yang committed
18
	"strings"
Michael Yang's avatar
Michael Yang committed
19
	"sync"
20
	"syscall"
21
	"time"
Jeffrey Morgan's avatar
Jeffrey Morgan committed
22

Michael Yang's avatar
Michael Yang committed
23
	"github.com/gin-contrib/cors"
Jeffrey Morgan's avatar
Jeffrey Morgan committed
24
25
	"github.com/gin-gonic/gin"

Jeffrey Morgan's avatar
Jeffrey Morgan committed
26
	"github.com/jmorganca/ollama/api"
27
	"github.com/jmorganca/ollama/gpu"
28
	"github.com/jmorganca/ollama/llm"
Michael Yang's avatar
Michael Yang committed
29
	"github.com/jmorganca/ollama/parser"
Michael Yang's avatar
Michael Yang committed
30
	"github.com/jmorganca/ollama/version"
Jeffrey Morgan's avatar
Jeffrey Morgan committed
31
32
)

Michael Yang's avatar
Michael Yang committed
33
34
var mode string = gin.DebugMode

35
36
37
38
type Server struct {
	WorkDir string
}

Michael Yang's avatar
Michael Yang committed
39
40
41
42
43
44
45
46
47
48
49
50
func init() {
	switch mode {
	case gin.DebugMode:
	case gin.ReleaseMode:
	case gin.TestMode:
	default:
		mode = gin.DebugMode
	}

	gin.SetMode(mode)
}

Jeffrey Morgan's avatar
Jeffrey Morgan committed
51
var loaded struct {
Michael Yang's avatar
Michael Yang committed
52
53
	mu sync.Mutex

54
	runner llm.LLM
Michael Yang's avatar
Michael Yang committed
55
56
57

	expireAt    time.Time
	expireTimer *time.Timer
Jeffrey Morgan's avatar
Jeffrey Morgan committed
58

59
60
	*Model
	*api.Options
Michael Yang's avatar
Michael Yang committed
61
62
}

63
64
var defaultSessionDuration = 5 * time.Minute

Bruce MacDonald's avatar
Bruce MacDonald committed
65
// load a model into memory if it is not already loaded, it is up to the caller to lock loaded.mu before calling this function
66
func load(c *gin.Context, model *Model, opts api.Options, sessionDuration time.Duration) error {
Bruce MacDonald's avatar
Bruce MacDonald committed
67
68
	workDir := c.GetString("workDir")

69
70
71
72
73
74
75
	needLoad := loaded.runner == nil || // is there a model loaded?
		loaded.ModelPath != model.ModelPath || // has the base model changed?
		!reflect.DeepEqual(loaded.AdapterPaths, model.AdapterPaths) || // have the adapters changed?
		!reflect.DeepEqual(loaded.Options.Runner, opts.Runner) // have the runner options changed?

	if needLoad {
		if loaded.runner != nil {
76
			slog.Info("changing loaded model")
77
78
79
80
			loaded.runner.Close()
			loaded.runner = nil
			loaded.Model = nil
			loaded.Options = nil
Michael Yang's avatar
Michael Yang committed
81
		}
Michael Yang's avatar
Michael Yang committed
82

Michael Yang's avatar
Michael Yang committed
83
		llmRunner, err := llm.New(workDir, model.ModelPath, model.AdapterPaths, model.ProjectorPaths, opts)
Michael Yang's avatar
Michael Yang committed
84
		if err != nil {
85
86
87
			// some older models are not compatible with newer versions of llama.cpp
			// show a generalized compatibility error until there is a better way to
			// check for model compatibility
Bruce MacDonald's avatar
Bruce MacDonald committed
88
			if errors.Is(llm.ErrUnsupportedFormat, err) || strings.Contains(err.Error(), "failed to load model") {
89
90
91
				err = fmt.Errorf("%v: this model may be incompatible with your version of Ollama. If you previously pulled this model, try updating it by running `ollama pull %s`", err, model.ShortName)
			}

92
			return err
Michael Yang's avatar
Michael Yang committed
93
94
		}

95
96
97
		loaded.Model = model
		loaded.runner = llmRunner
		loaded.Options = &opts
Michael Yang's avatar
Michael Yang committed
98
	}
99

Jeffrey Morgan's avatar
Jeffrey Morgan committed
100
	loaded.expireAt = time.Now().Add(sessionDuration)
Bruce MacDonald's avatar
Bruce MacDonald committed
101

Jeffrey Morgan's avatar
Jeffrey Morgan committed
102
103
104
105
	if loaded.expireTimer == nil {
		loaded.expireTimer = time.AfterFunc(sessionDuration, func() {
			loaded.mu.Lock()
			defer loaded.mu.Unlock()
Michael Yang's avatar
Michael Yang committed
106

Jeffrey Morgan's avatar
Jeffrey Morgan committed
107
			if time.Now().Before(loaded.expireAt) {
Michael Yang's avatar
Michael Yang committed
108
109
110
				return
			}

111
112
			if loaded.runner != nil {
				loaded.runner.Close()
Michael Yang's avatar
Michael Yang committed
113
114
			}

115
116
117
			loaded.runner = nil
			loaded.Model = nil
			loaded.Options = nil
Michael Yang's avatar
Michael Yang committed
118
		})
Michael Yang's avatar
Michael Yang committed
119
	}
120

Jeffrey Morgan's avatar
Jeffrey Morgan committed
121
	loaded.expireTimer.Reset(sessionDuration)
122
123
124
125
126
127
128
129
130
131
132
133
134
135
	return nil
}

func modelOptions(model *Model, requestOpts map[string]interface{}) (api.Options, error) {
	opts := api.DefaultOptions()
	if err := opts.FromMap(model.Options); err != nil {
		return api.Options{}, err
	}

	if err := opts.FromMap(requestOpts); err != nil {
		return api.Options{}, err
	}

	return opts, nil
Bruce MacDonald's avatar
Bruce MacDonald committed
136
137
138
139
140
141
142
143
}

func GenerateHandler(c *gin.Context) {
	loaded.mu.Lock()
	defer loaded.mu.Unlock()

	checkpointStart := time.Now()
	var req api.GenerateRequest
Michael Yang's avatar
Michael Yang committed
144
	err := c.ShouldBindJSON(&req)
Patrick Devine's avatar
Patrick Devine committed
145

Michael Yang's avatar
Michael Yang committed
146
147
148
149
150
151
	switch {
	case errors.Is(err, io.EOF):
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
		return
	case err != nil:
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
Bruce MacDonald's avatar
Bruce MacDonald committed
152
153
154
		return
	}

155
156
157
	// validate the request
	switch {
	case req.Model == "":
158
159
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "model is required"})
		return
160
161
162
	case len(req.Format) > 0 && req.Format != "json":
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "format must be json"})
		return
163
164
165
	case req.Raw && (req.Template != "" || req.System != "" || len(req.Context) > 0):
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "raw mode does not support template, system, or context"})
		return
166
167
	}

168
	model, err := GetModel(req.Model)
Bruce MacDonald's avatar
Bruce MacDonald committed
169
	if err != nil {
170
		var pErr *fs.PathError
171
		if errors.As(err, &pErr) {
172
			c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found, try pulling it first", req.Model)})
173
174
175
176
177
178
179
180
181
			return
		}
		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
		return
	}

	opts, err := modelOptions(model, req.Options)
	if err != nil {
		if errors.Is(err, api.ErrInvalidOpts) {
Bruce MacDonald's avatar
Bruce MacDonald committed
182
			c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
183
			return
184
		}
185
186
187
188
		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
		return
	}

189
190
191
192
193
194
195
	var sessionDuration time.Duration
	if req.KeepAlive == nil {
		sessionDuration = defaultSessionDuration
	} else {
		sessionDuration = req.KeepAlive.Duration
	}

196
197
	if err := load(c, model, opts, sessionDuration); err != nil {
		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
Bruce MacDonald's avatar
Bruce MacDonald committed
198
199
200
		return
	}

Bruce MacDonald's avatar
Bruce MacDonald committed
201
202
	// an empty request loads the model
	if req.Prompt == "" && req.Template == "" && req.System == "" {
203
		c.JSON(http.StatusOK, api.GenerateResponse{
204
205
			CreatedAt: time.Now().UTC(),
			Model:     req.Model,
Michael Yang's avatar
Michael Yang committed
206
207
			Done:      true,
		})
Bruce MacDonald's avatar
Bruce MacDonald committed
208
209
210
211
212
		return
	}

	checkpointLoaded := time.Now()

Bruce MacDonald's avatar
Bruce MacDonald committed
213
	var prompt string
214
	var promptVars PromptVars
Bruce MacDonald's avatar
Bruce MacDonald committed
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
	switch {
	case req.Raw:
		prompt = req.Prompt
	case req.Prompt != "":
		if req.Template != "" {
			// override the default model template
			model.Template = req.Template
		}

		var rebuild strings.Builder
		if req.Context != nil {
			// TODO: context is deprecated, at some point the context logic within this conditional should be removed
			prevCtx, err := loaded.runner.Decode(c.Request.Context(), req.Context)
			if err != nil {
				c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
				return
			}

			// Remove leading spaces from prevCtx if present
			prevCtx = strings.TrimPrefix(prevCtx, " ")
			rebuild.WriteString(prevCtx)
		}
237
		promptVars = PromptVars{
Bruce MacDonald's avatar
Bruce MacDonald committed
238
239
240
			System: req.System,
			Prompt: req.Prompt,
			First:  len(req.Context) == 0,
241
		}
242
243
244
245
246

		if promptVars.System == "" {
			promptVars.System = model.System
		}

247
		p, err := model.PreResponsePrompt(promptVars)
248
249
250
251
		if err != nil {
			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
			return
		}
Bruce MacDonald's avatar
Bruce MacDonald committed
252
253
		rebuild.WriteString(p)
		prompt = rebuild.String()
Bruce MacDonald's avatar
Bruce MacDonald committed
254
255
	}

Bruce MacDonald's avatar
Bruce MacDonald committed
256
	ch := make(chan any)
Bruce MacDonald's avatar
Bruce MacDonald committed
257
	var generated strings.Builder
Bruce MacDonald's avatar
Bruce MacDonald committed
258
259
260
	go func() {
		defer close(ch)

Bruce MacDonald's avatar
Bruce MacDonald committed
261
262
		fn := func(r llm.PredictResult) {
			// Update model expiration
Bruce MacDonald's avatar
Bruce MacDonald committed
263
264
265
			loaded.expireAt = time.Now().Add(sessionDuration)
			loaded.expireTimer.Reset(sessionDuration)

Bruce MacDonald's avatar
Bruce MacDonald committed
266
267
268
269
			// Build up the full response
			if _, err := generated.WriteString(r.Content); err != nil {
				ch <- gin.H{"error": err.Error()}
				return
Bruce MacDonald's avatar
Bruce MacDonald committed
270
271
			}

Bruce MacDonald's avatar
Bruce MacDonald committed
272
			resp := api.GenerateResponse{
273
				Model:     req.Model,
274
				CreatedAt: time.Now().UTC(),
275
276
				Done:      r.Done,
				Response:  r.Content,
Bruce MacDonald's avatar
Bruce MacDonald committed
277
278
279
280
281
282
				Metrics: api.Metrics{
					PromptEvalCount:    r.PromptEvalCount,
					PromptEvalDuration: r.PromptEvalDuration,
					EvalCount:          r.EvalCount,
					EvalDuration:       r.EvalDuration,
				},
Bruce MacDonald's avatar
Bruce MacDonald committed
283
284
			}

285
286
287
288
289
			if r.Done {
				resp.TotalDuration = time.Since(checkpointStart)
				resp.LoadDuration = checkpointLoaded.Sub(checkpointStart)

				if !req.Raw {
290
291
292
293
294
295
296
297
					// append the generated text to the history and template it if needed
					promptVars.Response = generated.String()
					result, err := model.PostResponseTemplate(promptVars)
					if err != nil {
						ch <- gin.H{"error": err.Error()}
						return
					}
					embd, err := loaded.runner.Encode(c.Request.Context(), prompt+result)
298
299
300
301
302
					if err != nil {
						ch <- gin.H{"error": err.Error()}
						return
					}
					resp.Context = embd
Bruce MacDonald's avatar
Bruce MacDonald committed
303
304
305
306
				}
			}

			ch <- resp
Bruce MacDonald's avatar
Bruce MacDonald committed
307
308
		}

Bruce MacDonald's avatar
Bruce MacDonald committed
309
310
		// Start prediction
		predictReq := llm.PredictOpts{
311
312
313
314
			Prompt:  prompt,
			Format:  req.Format,
			Images:  req.Images,
			Options: opts,
Bruce MacDonald's avatar
Bruce MacDonald committed
315
316
		}
		if err := loaded.runner.Predict(c.Request.Context(), predictReq, fn); err != nil {
Bruce MacDonald's avatar
Bruce MacDonald committed
317
318
319
320
321
			ch <- gin.H{"error": err.Error()}
		}
	}()

	if req.Stream != nil && !*req.Stream {
322
323
		// Accumulate responses into the final response
		var final api.GenerateResponse
Bruce MacDonald's avatar
Bruce MacDonald committed
324
		var sb strings.Builder
Bruce MacDonald's avatar
Bruce MacDonald committed
325
		for resp := range ch {
326
327
328
329
330
331
332
333
334
335
336
337
338
339
			switch r := resp.(type) {
			case api.GenerateResponse:
				sb.WriteString(r.Response)
				final = r
			case gin.H:
				if errorMsg, ok := r["error"].(string); ok {
					c.JSON(http.StatusInternalServerError, gin.H{"error": errorMsg})
					return
				} else {
					c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error format in response"})
					return
				}
			default:
				c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error"})
Bruce MacDonald's avatar
Bruce MacDonald committed
340
341
342
				return
			}
		}
343
344
345

		final.Response = sb.String()
		c.JSON(http.StatusOK, final)
Bruce MacDonald's avatar
Bruce MacDonald committed
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
		return
	}

	streamResponse(c, ch)
}

func EmbeddingHandler(c *gin.Context) {
	loaded.mu.Lock()
	defer loaded.mu.Unlock()

	var req api.EmbeddingRequest
	err := c.ShouldBindJSON(&req)
	switch {
	case errors.Is(err, io.EOF):
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
		return
	case err != nil:
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
		return
	}

	if req.Model == "" {
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "model is required"})
		return
	}

372
	model, err := GetModel(req.Model)
Bruce MacDonald's avatar
Bruce MacDonald committed
373
	if err != nil {
Bruce MacDonald's avatar
Bruce MacDonald committed
374
		var pErr *fs.PathError
375
		if errors.As(err, &pErr) {
Bruce MacDonald's avatar
Bruce MacDonald committed
376
			c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found, try pulling it first", req.Model)})
377
378
379
380
381
382
383
384
385
			return
		}
		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
		return
	}

	opts, err := modelOptions(model, req.Options)
	if err != nil {
		if errors.Is(err, api.ErrInvalidOpts) {
Bruce MacDonald's avatar
Bruce MacDonald committed
386
			c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
387
			return
Bruce MacDonald's avatar
Bruce MacDonald committed
388
		}
389
390
391
		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
		return
	}
392
393
394
395
396
397
398
399

	var sessionDuration time.Duration
	if req.KeepAlive == nil {
		sessionDuration = defaultSessionDuration
	} else {
		sessionDuration = req.KeepAlive.Duration
	}

400
401
	if err := load(c, model, opts, sessionDuration); err != nil {
		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
Bruce MacDonald's avatar
Bruce MacDonald committed
402
403
404
		return
	}

405
	if !loaded.Options.EmbeddingOnly {
Bruce MacDonald's avatar
Bruce MacDonald committed
406
407
408
409
		c.JSON(http.StatusBadRequest, gin.H{"error": "embedding option must be set to true"})
		return
	}

410
	embedding, err := loaded.runner.Embedding(c.Request.Context(), req.Prompt)
Bruce MacDonald's avatar
Bruce MacDonald committed
411
	if err != nil {
412
		slog.Info(fmt.Sprintf("embedding generation failed: %v", err))
Bruce MacDonald's avatar
Bruce MacDonald committed
413
414
415
416
417
418
419
420
421
422
		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to generate embedding"})
		return
	}

	resp := api.EmbeddingResponse{
		Embedding: embedding,
	}
	c.JSON(http.StatusOK, resp)
}

423
func PullModelHandler(c *gin.Context) {
Michael Yang's avatar
Michael Yang committed
424
	var req api.PullRequest
Michael Yang's avatar
Michael Yang committed
425
426
427
428
429
430
431
	err := c.ShouldBindJSON(&req)
	switch {
	case errors.Is(err, io.EOF):
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
		return
	case err != nil:
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
Michael Yang's avatar
Michael Yang committed
432
433
434
		return
	}

Michael Yang's avatar
Michael Yang committed
435
436
437
438
439
440
441
	var model string
	if req.Model != "" {
		model = req.Model
	} else if req.Name != "" {
		model = req.Name
	} else {
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "model is required"})
442
443
444
		return
	}

445
446
447
	ch := make(chan any)
	go func() {
		defer close(ch)
448
449
		fn := func(r api.ProgressResponse) {
			ch <- r
450
		}
451

452
453
454
455
		regOpts := &RegistryOptions{
			Insecure: req.Insecure,
		}

456
457
458
		ctx, cancel := context.WithCancel(c.Request.Context())
		defer cancel()

Michael Yang's avatar
Michael Yang committed
459
		if err := PullModel(ctx, model, regOpts, fn); err != nil {
Michael Yang's avatar
Michael Yang committed
460
			ch <- gin.H{"error": err.Error()}
461
462
463
		}
	}()

464
465
466
467
468
	if req.Stream != nil && !*req.Stream {
		waitForStream(c, ch)
		return
	}

469
470
471
	streamResponse(c, ch)
}

472
func PushModelHandler(c *gin.Context) {
473
	var req api.PushRequest
Michael Yang's avatar
Michael Yang committed
474
475
476
477
478
479
480
	err := c.ShouldBindJSON(&req)
	switch {
	case errors.Is(err, io.EOF):
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
		return
	case err != nil:
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
Michael Yang's avatar
Michael Yang committed
481
482
		return
	}
Michael Yang's avatar
Michael Yang committed
483

Michael Yang's avatar
Michael Yang committed
484
485
486
487
488
489
490
	var model string
	if req.Model != "" {
		model = req.Model
	} else if req.Name != "" {
		model = req.Name
	} else {
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "model is required"})
491
492
493
		return
	}

494
495
496
	ch := make(chan any)
	go func() {
		defer close(ch)
497
498
		fn := func(r api.ProgressResponse) {
			ch <- r
499
		}
500

501
502
503
504
		regOpts := &RegistryOptions{
			Insecure: req.Insecure,
		}

Michael Yang's avatar
Michael Yang committed
505
506
507
		ctx, cancel := context.WithCancel(c.Request.Context())
		defer cancel()

Michael Yang's avatar
Michael Yang committed
508
		if err := PushModel(ctx, model, regOpts, fn); err != nil {
Michael Yang's avatar
Michael Yang committed
509
			ch <- gin.H{"error": err.Error()}
510
511
512
		}
	}()

513
514
515
516
517
	if req.Stream != nil && !*req.Stream {
		waitForStream(c, ch)
		return
	}

518
519
520
	streamResponse(c, ch)
}

521
func CreateModelHandler(c *gin.Context) {
522
	var req api.CreateRequest
Michael Yang's avatar
Michael Yang committed
523
524
525
526
527
528
529
	err := c.ShouldBindJSON(&req)
	switch {
	case errors.Is(err, io.EOF):
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
		return
	case err != nil:
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
Michael Yang's avatar
Michael Yang committed
530
		return
531
532
	}

Michael Yang's avatar
Michael Yang committed
533
534
535
536
537
538
539
	var model string
	if req.Model != "" {
		model = req.Model
	} else if req.Name != "" {
		model = req.Name
	} else {
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "model is required"})
540
541
542
		return
	}

Michael Yang's avatar
Michael Yang committed
543
	if err := ParseModelPath(model).Validate(); err != nil {
544
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
545
546
547
		return
	}

Michael Yang's avatar
Michael Yang committed
548
549
	if req.Path == "" && req.Modelfile == "" {
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "path or modelfile are required"})
Michael Yang's avatar
Michael Yang committed
550
551
		return
	}
Michael Yang's avatar
Michael Yang committed
552
553
554

	var modelfile io.Reader = strings.NewReader(req.Modelfile)
	if req.Path != "" && req.Modelfile == "" {
555
		mf, err := os.Open(req.Path)
Michael Yang's avatar
Michael Yang committed
556
557
558
559
		if err != nil {
			c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("error reading modelfile: %s", err)})
			return
		}
560
		defer mf.Close()
Michael Yang's avatar
Michael Yang committed
561

562
		modelfile = mf
Michael Yang's avatar
Michael Yang committed
563
	}
Michael Yang's avatar
Michael Yang committed
564
565
566
567
568
569
570

	commands, err := parser.Parse(modelfile)
	if err != nil {
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
		return
	}

Michael Yang's avatar
Michael Yang committed
571
	ch := make(chan any)
Michael Yang's avatar
Michael Yang committed
572
573
	go func() {
		defer close(ch)
574
575
		fn := func(resp api.ProgressResponse) {
			ch <- resp
576
577
		}

578
579
580
		ctx, cancel := context.WithCancel(c.Request.Context())
		defer cancel()

Michael Yang's avatar
Michael Yang committed
581
		if err := CreateModel(ctx, model, filepath.Dir(req.Path), commands, fn); err != nil {
Michael Yang's avatar
Michael Yang committed
582
			ch <- gin.H{"error": err.Error()}
583
		}
Michael Yang's avatar
Michael Yang committed
584
	}()
Michael Yang's avatar
Michael Yang committed
585

586
587
588
589
590
	if req.Stream != nil && !*req.Stream {
		waitForStream(c, ch)
		return
	}

Michael Yang's avatar
Michael Yang committed
591
	streamResponse(c, ch)
Bruce MacDonald's avatar
Bruce MacDonald committed
592
593
}

594
595
func DeleteModelHandler(c *gin.Context) {
	var req api.DeleteRequest
Michael Yang's avatar
Michael Yang committed
596
597
598
599
600
601
602
	err := c.ShouldBindJSON(&req)
	switch {
	case errors.Is(err, io.EOF):
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
		return
	case err != nil:
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
603
604
605
		return
	}

Michael Yang's avatar
Michael Yang committed
606
607
608
609
610
611
612
	var model string
	if req.Model != "" {
		model = req.Model
	} else if req.Name != "" {
		model = req.Name
	} else {
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "model is required"})
613
614
615
		return
	}

Michael Yang's avatar
Michael Yang committed
616
	if err := DeleteModel(model); err != nil {
617
		if os.IsNotExist(err) {
Michael Yang's avatar
Michael Yang committed
618
			c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found", model)})
619
		} else {
620
621
			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
		}
622
623
		return
	}
Michael Yang's avatar
Michael Yang committed
624
625
626
627
628
629
630
631
632
633
634
635

	manifestsPath, err := GetManifestPath()
	if err != nil {
		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
		return
	}

	if err := PruneDirectory(manifestsPath); err != nil {
		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
		return
	}

636
	c.JSON(http.StatusOK, nil)
637
638
}

Patrick Devine's avatar
Patrick Devine committed
639
640
func ShowModelHandler(c *gin.Context) {
	var req api.ShowRequest
Michael Yang's avatar
Michael Yang committed
641
642
643
644
645
646
647
	err := c.ShouldBindJSON(&req)
	switch {
	case errors.Is(err, io.EOF):
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
		return
	case err != nil:
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
Patrick Devine's avatar
Patrick Devine committed
648
649
650
		return
	}

Michael Yang's avatar
Michael Yang committed
651
	if req.Model != "" {
Michael Yang's avatar
Michael Yang committed
652
		// noop
Michael Yang's avatar
Michael Yang committed
653
	} else if req.Name != "" {
Michael Yang's avatar
Michael Yang committed
654
		req.Model = req.Name
Michael Yang's avatar
Michael Yang committed
655
	} else {
656
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "model is required"})
657
658
659
		return
	}

660
	resp, err := GetModelInfo(req)
Patrick Devine's avatar
Patrick Devine committed
661
662
	if err != nil {
		if os.IsNotExist(err) {
Michael Yang's avatar
Michael Yang committed
663
			c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found", req.Model)})
Patrick Devine's avatar
Patrick Devine committed
664
665
666
667
668
669
670
671
672
		} else {
			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
		}
		return
	}

	c.JSON(http.StatusOK, resp)
}

673
674
func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
	model, err := GetModel(req.Model)
Patrick Devine's avatar
Patrick Devine committed
675
676
677
678
	if err != nil {
		return nil, err
	}

Patrick Devine's avatar
Patrick Devine committed
679
	modelDetails := api.ModelDetails{
680
		ParentModel:       model.ParentModel,
Patrick Devine's avatar
Patrick Devine committed
681
682
683
684
685
686
687
		Format:            model.Config.ModelFormat,
		Family:            model.Config.ModelFamily,
		Families:          model.Config.ModelFamilies,
		ParameterSize:     model.Config.ModelType,
		QuantizationLevel: model.Config.FileType,
	}

688
689
690
691
692
693
694
695
	if req.System != "" {
		model.System = req.System
	}

	if req.Template != "" {
		model.Template = req.Template
	}

696
697
698
699
700
	msgs := make([]api.Message, 0)
	for _, msg := range model.Messages {
		msgs = append(msgs, api.Message{Role: msg.Role, Content: msg.Content})
	}

Patrick Devine's avatar
Patrick Devine committed
701
702
703
704
	resp := &api.ShowResponse{
		License:  strings.Join(model.License, "\n"),
		System:   model.System,
		Template: model.Template,
Patrick Devine's avatar
Patrick Devine committed
705
		Details:  modelDetails,
706
		Messages: msgs,
Patrick Devine's avatar
Patrick Devine committed
707
708
709
710
711
712
713
714
	}

	var params []string
	cs := 30
	for k, v := range model.Options {
		switch val := v.(type) {
		case []interface{}:
			for _, nv := range val {
Patrick Devine's avatar
Patrick Devine committed
715
				params = append(params, fmt.Sprintf("%-*s %#v", cs, k, nv))
Patrick Devine's avatar
Patrick Devine committed
716
			}
Patrick Devine's avatar
Patrick Devine committed
717
718
		default:
			params = append(params, fmt.Sprintf("%-*s %#v", cs, k, v))
Patrick Devine's avatar
Patrick Devine committed
719
720
721
722
		}
	}
	resp.Parameters = strings.Join(params, "\n")

723
724
725
726
727
728
729
730
731
732
733
734
735
	for k, v := range req.Options {
		if _, ok := req.Options[k]; ok {
			model.Options[k] = v
		}
	}

	mf, err := ShowModelfile(model)
	if err != nil {
		return nil, err
	}

	resp.Modelfile = mf

Patrick Devine's avatar
Patrick Devine committed
736
737
738
	return resp, nil
}

739
func ListModelsHandler(c *gin.Context) {
740
	models := make([]api.ModelResponse, 0)
741
	manifestsPath, err := GetManifestPath()
Patrick Devine's avatar
Patrick Devine committed
742
743
744
745
	if err != nil {
		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
		return
	}
Michael Yang's avatar
Michael Yang committed
746

Patrick Devine's avatar
Patrick Devine committed
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
	modelResponse := func(modelName string) (api.ModelResponse, error) {
		model, err := GetModel(modelName)
		if err != nil {
			return api.ModelResponse{}, err
		}

		modelDetails := api.ModelDetails{
			Format:            model.Config.ModelFormat,
			Family:            model.Config.ModelFamily,
			Families:          model.Config.ModelFamilies,
			ParameterSize:     model.Config.ModelType,
			QuantizationLevel: model.Config.FileType,
		}

		return api.ModelResponse{
Michael Yang's avatar
Michael Yang committed
762
			Model:   model.ShortName,
Patrick Devine's avatar
Patrick Devine committed
763
764
765
766
767
768
769
			Name:    model.ShortName,
			Size:    model.Size,
			Digest:  model.Digest,
			Details: modelDetails,
		}, nil
	}

Michael Yang's avatar
Michael Yang committed
770
	walkFunc := func(path string, info os.FileInfo, _ error) error {
Patrick Devine's avatar
Patrick Devine committed
771
		if !info.IsDir() {
772
773
774
775
			path, tag := filepath.Split(path)
			model := strings.Trim(strings.TrimPrefix(path, manifestsPath), string(os.PathSeparator))
			modelPath := strings.Join([]string{model, tag}, ":")
			canonicalModelPath := strings.ReplaceAll(modelPath, string(os.PathSeparator), "/")
776

777
			resp, err := modelResponse(canonicalModelPath)
Patrick Devine's avatar
Patrick Devine committed
778
			if err != nil {
779
				slog.Info(fmt.Sprintf("skipping file: %s", canonicalModelPath))
Michael Yang's avatar
Michael Yang committed
780
				// nolint: nilerr
781
				return nil
Patrick Devine's avatar
Patrick Devine committed
782
			}
Michael Yang's avatar
Michael Yang committed
783

Patrick Devine's avatar
Patrick Devine committed
784
785
			resp.ModifiedAt = info.ModTime()
			models = append(models, resp)
Patrick Devine's avatar
Patrick Devine committed
786
		}
Michael Yang's avatar
Michael Yang committed
787

Patrick Devine's avatar
Patrick Devine committed
788
		return nil
Michael Yang's avatar
Michael Yang committed
789
790
	}

791
	if err := filepath.Walk(manifestsPath, walkFunc); err != nil {
Patrick Devine's avatar
Patrick Devine committed
792
793
794
795
		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
		return
	}

Michael Yang's avatar
Michael Yang committed
796
	c.JSON(http.StatusOK, api.ListResponse{Models: models})
Patrick Devine's avatar
Patrick Devine committed
797
798
}

Patrick Devine's avatar
Patrick Devine committed
799
800
func CopyModelHandler(c *gin.Context) {
	var req api.CopyRequest
Michael Yang's avatar
Michael Yang committed
801
802
803
804
805
806
807
	err := c.ShouldBindJSON(&req)
	switch {
	case errors.Is(err, io.EOF):
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
		return
	case err != nil:
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
Patrick Devine's avatar
Patrick Devine committed
808
809
810
		return
	}

811
812
813
814
815
	if req.Source == "" || req.Destination == "" {
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "source add destination are required"})
		return
	}

816
817
818
819
820
	if err := ParseModelPath(req.Destination).Validate(); err != nil {
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
		return
	}

Patrick Devine's avatar
Patrick Devine committed
821
822
823
824
825
826
827
828
829
830
	if err := CopyModel(req.Source, req.Destination); err != nil {
		if os.IsNotExist(err) {
			c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found", req.Source)})
		} else {
			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
		}
		return
	}
}

Michael Yang's avatar
Michael Yang committed
831
func HeadBlobHandler(c *gin.Context) {
Michael Yang's avatar
Michael Yang committed
832
833
834
835
836
837
838
839
840
841
842
	path, err := GetBlobsPath(c.Param("digest"))
	if err != nil {
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
		return
	}

	if _, err := os.Stat(path); err != nil {
		c.AbortWithStatusJSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("blob %q not found", c.Param("digest"))})
		return
	}

Michael Yang's avatar
Michael Yang committed
843
	c.Status(http.StatusOK)
Michael Yang's avatar
Michael Yang committed
844
845
846
}

func CreateBlobHandler(c *gin.Context) {
847
	layer, err := NewLayer(c.Request.Body, "")
Michael Yang's avatar
Michael Yang committed
848
849
850
851
852
	if err != nil {
		c.AbortWithStatusJSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
		return
	}

853
854
	if layer.Digest != c.Param("digest") {
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("digest mismatch, expected %q, got %q", c.Param("digest"), layer.Digest)})
Michael Yang's avatar
Michael Yang committed
855
856
857
		return
	}

858
	if _, err := layer.Commit(); err != nil {
Michael Yang's avatar
Michael Yang committed
859
860
861
862
		c.AbortWithStatusJSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
		return
	}

Michael Yang's avatar
Michael Yang committed
863
	c.Status(http.StatusCreated)
Michael Yang's avatar
Michael Yang committed
864
865
}

Michael Yang's avatar
Michael Yang committed
866
867
868
869
870
871
var defaultAllowOrigins = []string{
	"localhost",
	"127.0.0.1",
	"0.0.0.0",
}

872
873
874
875
876
func NewServer() (*Server, error) {
	workDir, err := os.MkdirTemp("", "ollama")
	if err != nil {
		return nil, err
	}
877

878
879
880
881
	return &Server{
		WorkDir: workDir,
	}, nil
}
882

883
884
885
886
func (s *Server) GenerateRoutes() http.Handler {
	var origins []string
	if o := os.Getenv("OLLAMA_ORIGINS"); o != "" {
		origins = strings.Split(o, ",")
887
888
	}

Michael Yang's avatar
Michael Yang committed
889
890
	config := cors.DefaultConfig()
	config.AllowWildcard = true
891
	config.AllowBrowserExtensions = true
Michael Yang's avatar
Michael Yang committed
892

893
	config.AllowOrigins = origins
Michael Yang's avatar
Michael Yang committed
894
895
896
897
898
899
900
901
	for _, allowOrigin := range defaultAllowOrigins {
		config.AllowOrigins = append(config.AllowOrigins,
			fmt.Sprintf("http://%s", allowOrigin),
			fmt.Sprintf("https://%s", allowOrigin),
			fmt.Sprintf("http://%s:*", allowOrigin),
			fmt.Sprintf("https://%s:*", allowOrigin),
		)
	}
Michael Yang's avatar
Michael Yang committed
902

Bruce MacDonald's avatar
Bruce MacDonald committed
903
	r := gin.Default()
904
905
906
	r.Use(
		cors.New(config),
		func(c *gin.Context) {
907
			c.Set("workDir", s.WorkDir)
908
909
910
			c.Next()
		},
	)
Bruce MacDonald's avatar
Bruce MacDonald committed
911

912
913
	r.POST("/api/pull", PullModelHandler)
	r.POST("/api/generate", GenerateHandler)
Bruce MacDonald's avatar
Bruce MacDonald committed
914
	r.POST("/api/chat", ChatHandler)
Bruce MacDonald's avatar
Bruce MacDonald committed
915
	r.POST("/api/embeddings", EmbeddingHandler)
916
917
	r.POST("/api/create", CreateModelHandler)
	r.POST("/api/push", PushModelHandler)
Patrick Devine's avatar
Patrick Devine committed
918
	r.POST("/api/copy", CopyModelHandler)
919
	r.DELETE("/api/delete", DeleteModelHandler)
Patrick Devine's avatar
Patrick Devine committed
920
	r.POST("/api/show", ShowModelHandler)
Michael Yang's avatar
Michael Yang committed
921
	r.POST("/api/blobs/:digest", CreateBlobHandler)
Michael Yang's avatar
Michael Yang committed
922
	r.HEAD("/api/blobs/:digest", HeadBlobHandler)
Jeffrey Morgan's avatar
Jeffrey Morgan committed
923

Michael Yang's avatar
Michael Yang committed
924
925
926
927
928
929
	for _, method := range []string{http.MethodGet, http.MethodHead} {
		r.Handle(method, "/", func(c *gin.Context) {
			c.String(http.StatusOK, "Ollama is running")
		})

		r.Handle(method, "/api/tags", ListModelsHandler)
Michael Yang's avatar
Michael Yang committed
930
931
932
		r.Handle(method, "/api/version", func(c *gin.Context) {
			c.JSON(http.StatusOK, gin.H{"version": version.Version})
		})
Michael Yang's avatar
Michael Yang committed
933
934
	}

935
936
937
938
	return r
}

func Serve(ln net.Listener) error {
939
940
941
942
943
944
945
	if debug := os.Getenv("OLLAMA_DEBUG"); debug != "" {
		var programLevel = new(slog.LevelVar)
		h := slog.NewTextHandler(os.Stderr, &slog.HandlerOptions{Level: programLevel, AddSource: true})
		slog.SetDefault(slog.New(h))
		programLevel.Set(slog.LevelDebug)
		slog.Debug("Debug logging enabled")
	}
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
	if noprune := os.Getenv("OLLAMA_NOPRUNE"); noprune == "" {
		// clean up unused layers and manifests
		if err := PruneLayers(); err != nil {
			return err
		}

		manifestsPath, err := GetManifestPath()
		if err != nil {
			return err
		}

		if err := PruneDirectory(manifestsPath); err != nil {
			return err
		}
	}

	s, err := NewServer()
	if err != nil {
		return err
	}
	r := s.GenerateRoutes()

968
	slog.Info(fmt.Sprintf("Listening on %s (version %s)", ln.Addr(), version.Version))
969
	srvr := &http.Server{
Jeffrey Morgan's avatar
Jeffrey Morgan committed
970
971
972
		Handler: r,
	}

973
974
	// listen for a ctrl+c and stop any loaded llm
	signals := make(chan os.Signal, 1)
975
	signal.Notify(signals, syscall.SIGINT, syscall.SIGTERM)
976
977
	go func() {
		<-signals
978
979
		if loaded.runner != nil {
			loaded.runner.Close()
980
		}
981
		os.RemoveAll(s.WorkDir)
982
983
984
		os.Exit(0)
	}()

985
986
987
988
	if err := llm.Init(s.WorkDir); err != nil {
		return fmt.Errorf("unable to initialize llm library %w", err)
	}
	if runtime.GOOS == "linux" { // TODO - windows too
989
		// check compatibility to log warnings
990
		if _, err := gpu.CheckVRAM(); err != nil {
991
			slog.Info(err.Error())
992
993
994
		}
	}

995
	return srvr.Serve(ln)
Jeffrey Morgan's avatar
Jeffrey Morgan committed
996
}
Michael Yang's avatar
Michael Yang committed
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
func waitForStream(c *gin.Context, ch chan interface{}) {
	c.Header("Content-Type", "application/json")
	for resp := range ch {
		switch r := resp.(type) {
		case api.ProgressResponse:
			if r.Status == "success" {
				c.JSON(http.StatusOK, r)
				return
			}
		case gin.H:
			if errorMsg, ok := r["error"].(string); ok {
				c.JSON(http.StatusInternalServerError, gin.H{"error": errorMsg})
				return
			} else {
				c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error format in progress response"})
				return
			}
		default:
			c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected progress response"})
			return
		}
	}
	c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected end of progress response"})
}

Michael Yang's avatar
Michael Yang committed
1023
func streamResponse(c *gin.Context, ch chan any) {
1024
	c.Header("Content-Type", "application/x-ndjson")
Michael Yang's avatar
Michael Yang committed
1025
1026
1027
1028
1029
1030
1031
1032
	c.Stream(func(w io.Writer) bool {
		val, ok := <-ch
		if !ok {
			return false
		}

		bts, err := json.Marshal(val)
		if err != nil {
1033
			slog.Info(fmt.Sprintf("streamResponse: json.Marshal failed with %s", err))
Michael Yang's avatar
Michael Yang committed
1034
1035
1036
			return false
		}

1037
		// Delineate chunks with new-line delimiter
Michael Yang's avatar
Michael Yang committed
1038
1039
		bts = append(bts, '\n')
		if _, err := w.Write(bts); err != nil {
1040
			slog.Info(fmt.Sprintf("streamResponse: w.Write failed with %s", err))
Michael Yang's avatar
Michael Yang committed
1041
1042
1043
1044
1045
1046
			return false
		}

		return true
	})
}
Bruce MacDonald's avatar
Bruce MacDonald committed
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074

func ChatHandler(c *gin.Context) {
	loaded.mu.Lock()
	defer loaded.mu.Unlock()

	checkpointStart := time.Now()

	var req api.ChatRequest
	err := c.ShouldBindJSON(&req)
	switch {
	case errors.Is(err, io.EOF):
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
		return
	case err != nil:
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
		return
	}

	// validate the request
	switch {
	case req.Model == "":
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "model is required"})
		return
	case len(req.Format) > 0 && req.Format != "json":
		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "format must be json"})
		return
	}

1075
	model, err := GetModel(req.Model)
Bruce MacDonald's avatar
Bruce MacDonald committed
1076
1077
	if err != nil {
		var pErr *fs.PathError
1078
		if errors.As(err, &pErr) {
Bruce MacDonald's avatar
Bruce MacDonald committed
1079
			c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found, try pulling it first", req.Model)})
1080
1081
1082
1083
1084
1085
1086
1087
1088
			return
		}
		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
		return
	}

	opts, err := modelOptions(model, req.Options)
	if err != nil {
		if errors.Is(err, api.ErrInvalidOpts) {
Bruce MacDonald's avatar
Bruce MacDonald committed
1089
			c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
1090
			return
Bruce MacDonald's avatar
Bruce MacDonald committed
1091
		}
1092
1093
1094
		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
		return
	}
1095
1096
1097
1098
1099
1100
1101
1102

	var sessionDuration time.Duration
	if req.KeepAlive == nil {
		sessionDuration = defaultSessionDuration
	} else {
		sessionDuration = req.KeepAlive.Duration
	}

1103
1104
	if err := load(c, model, opts, sessionDuration); err != nil {
		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
Bruce MacDonald's avatar
Bruce MacDonald committed
1105
1106
1107
1108
1109
		return
	}

	// an empty request loads the model
	if len(req.Messages) == 0 {
1110
1111
1112
1113
1114
1115
1116
		resp := api.ChatResponse{
			CreatedAt: time.Now().UTC(),
			Model:     req.Model,
			Done:      true,
			Message:   api.Message{Role: "assistant"},
		}
		c.JSON(http.StatusOK, resp)
Bruce MacDonald's avatar
Bruce MacDonald committed
1117
1118
1119
1120
1121
		return
	}

	checkpointLoaded := time.Now()

1122
	prompt, images, err := model.ChatPrompt(req.Messages)
Bruce MacDonald's avatar
Bruce MacDonald committed
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
	if err != nil {
		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
		return
	}

	ch := make(chan any)

	go func() {
		defer close(ch)

		fn := func(r llm.PredictResult) {
			// Update model expiration
			loaded.expireAt = time.Now().Add(sessionDuration)
			loaded.expireTimer.Reset(sessionDuration)

			resp := api.ChatResponse{
1139
				Model:     req.Model,
1140
				CreatedAt: time.Now().UTC(),
1141
				Message:   api.Message{Role: "assistant", Content: r.Content},
Bruce MacDonald's avatar
Bruce MacDonald committed
1142
1143
1144
1145
1146
1147
1148
1149
1150
				Done:      r.Done,
				Metrics: api.Metrics{
					PromptEvalCount:    r.PromptEvalCount,
					PromptEvalDuration: r.PromptEvalDuration,
					EvalCount:          r.EvalCount,
					EvalDuration:       r.EvalDuration,
				},
			}

1151
1152
1153
			if r.Done {
				resp.TotalDuration = time.Since(checkpointStart)
				resp.LoadDuration = checkpointLoaded.Sub(checkpointStart)
Bruce MacDonald's avatar
Bruce MacDonald committed
1154
1155
1156
1157
1158
1159
1160
			}

			ch <- resp
		}

		// Start prediction
		predictReq := llm.PredictOpts{
1161
1162
1163
1164
			Prompt:  prompt,
			Format:  req.Format,
			Images:  images,
			Options: opts,
Bruce MacDonald's avatar
Bruce MacDonald committed
1165
1166
1167
1168
1169
1170
1171
		}
		if err := loaded.runner.Predict(c.Request.Context(), predictReq, fn); err != nil {
			ch <- gin.H{"error": err.Error()}
		}
	}()

	if req.Stream != nil && !*req.Stream {
1172
1173
		// Accumulate responses into the final response
		var final api.ChatResponse
Bruce MacDonald's avatar
Bruce MacDonald committed
1174
1175
		var sb strings.Builder
		for resp := range ch {
1176
1177
			switch r := resp.(type) {
			case api.ChatResponse:
1178
				sb.WriteString(r.Message.Content)
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
				final = r
			case gin.H:
				if errorMsg, ok := r["error"].(string); ok {
					c.JSON(http.StatusInternalServerError, gin.H{"error": errorMsg})
					return
				} else {
					c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error format in response"})
					return
				}
			default:
				c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error"})
				return
Bruce MacDonald's avatar
Bruce MacDonald committed
1191
1192
			}
		}
1193

1194
		final.Message = api.Message{Role: "assistant", Content: sb.String()}
1195
		c.JSON(http.StatusOK, final)
Bruce MacDonald's avatar
Bruce MacDonald committed
1196
1197
1198
1199
1200
		return
	}

	streamResponse(c, ch)
}