types.go 27.8 KB
Newer Older
Jeffrey Morgan's avatar
Jeffrey Morgan committed
1
2
package api

3
import (
Michael Yang's avatar
Michael Yang committed
4
	"encoding/json"
5
	"fmt"
6
	"log/slog"
Michael Yang's avatar
Michael Yang committed
7
	"math"
8
	"os"
9
	"reflect"
10
	"strconv"
11
	"strings"
12
	"time"
13
14

	"github.com/ollama/ollama/envconfig"
15
	"github.com/ollama/ollama/types/model"
16
)
Michael Yang's avatar
Michael Yang committed
17

18
// StatusError is an error with an HTTP status code and message.
Patrick Devine's avatar
Patrick Devine committed
19
type StatusError struct {
20
21
22
	StatusCode   int
	Status       string
	ErrorMessage string `json:"error"`
Patrick Devine's avatar
Patrick Devine committed
23
24
25
}

func (e StatusError) Error() string {
26
27
28
29
30
31
32
33
34
35
	switch {
	case e.Status != "" && e.ErrorMessage != "":
		return fmt.Sprintf("%s: %s", e.Status, e.ErrorMessage)
	case e.Status != "":
		return e.Status
	case e.ErrorMessage != "":
		return e.ErrorMessage
	default:
		// this should not happen
		return "something went wrong, please see the ollama server logs for details"
Patrick Devine's avatar
Patrick Devine committed
36
37
38
	}
}

39
// ImageData represents the raw binary data of an image file.
Patrick Devine's avatar
Patrick Devine committed
40
41
type ImageData []byte

42
43
44
// GenerateRequest describes a request sent by [Client.Generate]. While you
// have to specify the Model and Prompt fields, all the other fields have
// reasonable defaults for basic uses.
45
type GenerateRequest struct {
46
47
48
	// Model is the model name; it should be a name familiar to Ollama from
	// the library at https://ollama.com/library
	Model string `json:"model"`
49

50
51
52
	// Prompt is the textual prompt to send to the model.
	Prompt string `json:"prompt"`

53
54
55
	// Suffix is the text that comes after the inserted text.
	Suffix string `json:"suffix"`

56
57
58
59
60
61
62
	// System overrides the model's default system message/prompt.
	System string `json:"system"`

	// Template overrides the model's default prompt template.
	Template string `json:"template"`

	// Context is the context parameter returned from a previous call to
63
	// [Client.Generate]. It can be used to keep a short conversational memory.
64
65
66
67
68
69
70
71
72
	Context []int `json:"context,omitempty"`

	// Stream specifies whether the response is streaming; it is true by default.
	Stream *bool `json:"stream,omitempty"`

	// Raw set to true means that no formatting will be applied to the prompt.
	Raw bool `json:"raw,omitempty"`

	// Format specifies the format to return a response in.
73
	Format json.RawMessage `json:"format,omitempty"`
74
75
76
77
78

	// KeepAlive controls how long the model will stay loaded in memory following
	// this request.
	KeepAlive *Duration `json:"keep_alive,omitempty"`

79
	// Images is an optional list of raw image bytes accompanying this
80
81
82
83
84
	// request, for multimodal models.
	Images []ImageData `json:"images,omitempty"`

	// Options lists model-specific options. For example, temperature can be
	// set through this field, if the model supports it.
85
	Options map[string]any `json:"options"`
86
87

	// Think controls whether thinking/reasoning models will think before
Michael Yang's avatar
Michael Yang committed
88
89
	// responding. Can be a boolean (true/false) or a string ("high", "medium", "low")
	// for supported models. Needs to be a pointer so we can distinguish between false
90
91
	// (request that thinking _not_ be used) and unset (use the old behavior
	// before this option was introduced)
Michael Yang's avatar
Michael Yang committed
92
	Think *ThinkValue `json:"think,omitempty"`
93
94
95
96

	// DebugRenderOnly is a debug option that, when set to true, returns the rendered
	// template instead of calling the model.
	DebugRenderOnly bool `json:"_debug_render_only,omitempty"`
97
98
}

99
// ChatRequest describes a request sent by [Client.Chat].
Bruce MacDonald's avatar
Bruce MacDonald committed
100
type ChatRequest struct {
101
102
103
104
105
106
	// Model is the model name, as in [GenerateRequest].
	Model string `json:"model"`

	// Messages is the messages of the chat - can be used to keep a chat memory.
	Messages []Message `json:"messages"`

107
	// Stream enables streaming of returned responses; true by default.
108
109
110
	Stream *bool `json:"stream,omitempty"`

	// Format is the format to return the response in (e.g. "json").
111
	Format json.RawMessage `json:"format,omitempty"`
112
113

	// KeepAlive controls how long the model will stay loaded into memory
114
	// following the request.
115
	KeepAlive *Duration `json:"keep_alive,omitempty"`
Bruce MacDonald's avatar
Bruce MacDonald committed
116

Michael Yang's avatar
tools  
Michael Yang committed
117
	// Tools is an optional list of tools the model has access to.
118
	Tools `json:"tools,omitempty"`
Michael Yang's avatar
tools  
Michael Yang committed
119

120
	// Options lists model-specific options.
121
	Options map[string]any `json:"options"`
122
123

	// Think controls whether thinking/reasoning models will think before
Michael Yang's avatar
Michael Yang committed
124
125
126
	// responding. Can be a boolean (true/false) or a string ("high", "medium", "low")
	// for supported models.
	Think *ThinkValue `json:"think,omitempty"`
127
128
129
130

	// DebugRenderOnly is a debug option that, when set to true, returns the rendered
	// template instead of calling the model.
	DebugRenderOnly bool `json:"_debug_render_only,omitempty"`
Bruce MacDonald's avatar
Bruce MacDonald committed
131
132
}

133
134
135
136
137
138
139
type Tools []Tool

func (t Tools) String() string {
	bts, _ := json.Marshal(t)
	return string(bts)
}

140
141
142
143
144
func (t Tool) String() string {
	bts, _ := json.Marshal(t)
	return string(bts)
}

145
146
147
// Message is a single message in a chat sequence. The message contains the
// role ("system", "user", or "assistant"), the content and an optional list
// of images.
Bruce MacDonald's avatar
Bruce MacDonald committed
148
type Message struct {
149
150
151
152
153
	Role    string `json:"role"`
	Content string `json:"content"`
	// Thinking contains the text that was inside thinking tags in the
	// original model output when ChatRequest.Think is enabled.
	Thinking  string      `json:"thinking,omitempty"`
Michael Yang's avatar
tools  
Michael Yang committed
154
155
	Images    []ImageData `json:"images,omitempty"`
	ToolCalls []ToolCall  `json:"tool_calls,omitempty"`
156
	ToolName  string      `json:"tool_name,omitempty"`
Michael Yang's avatar
tools  
Michael Yang committed
157
158
}

159
160
161
162
163
164
165
166
167
168
169
170
func (m *Message) UnmarshalJSON(b []byte) error {
	type Alias Message
	var a Alias
	if err := json.Unmarshal(b, &a); err != nil {
		return err
	}

	*m = Message(a)
	m.Role = strings.ToLower(m.Role)
	return nil
}

171
172
173
174
175
type ToolCall struct {
	Function ToolCallFunction `json:"function"`
}

type ToolCallFunction struct {
176
	Index     int                       `json:"index,omitempty"`
177
178
179
180
181
182
183
184
185
186
187
188
189
	Name      string                    `json:"name"`
	Arguments ToolCallFunctionArguments `json:"arguments"`
}

type ToolCallFunctionArguments map[string]any

func (t *ToolCallFunctionArguments) String() string {
	bts, _ := json.Marshal(t)
	return string(bts)
}

type Tool struct {
	Type     string       `json:"type"`
190
	Items    any          `json:"items,omitempty"`
191
192
193
	Function ToolFunction `json:"function"`
}

194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
// PropertyType can be either a string or an array of strings
type PropertyType []string

// UnmarshalJSON implements the json.Unmarshaler interface
func (pt *PropertyType) UnmarshalJSON(data []byte) error {
	// Try to unmarshal as a string first
	var s string
	if err := json.Unmarshal(data, &s); err == nil {
		*pt = []string{s}
		return nil
	}

	// If that fails, try to unmarshal as an array of strings
	var a []string
	if err := json.Unmarshal(data, &a); err != nil {
		return err
	}
	*pt = a
	return nil
}

// MarshalJSON implements the json.Marshaler interface
func (pt PropertyType) MarshalJSON() ([]byte, error) {
	if len(pt) == 1 {
		// If there's only one type, marshal as a string
		return json.Marshal(pt[0])
	}
	// Otherwise marshal as an array
	return json.Marshal([]string(pt))
}

// String returns a string representation of the PropertyType
func (pt PropertyType) String() string {
	if len(pt) == 0 {
		return ""
	}
	if len(pt) == 1 {
		return pt[0]
	}
	return fmt.Sprintf("%v", []string(pt))
}

Devon Rifkin's avatar
Devon Rifkin committed
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
type ToolProperty struct {
	AnyOf       []ToolProperty `json:"anyOf,omitempty"`
	Type        PropertyType   `json:"type"`
	Items       any            `json:"items,omitempty"`
	Description string         `json:"description"`
	Enum        []any          `json:"enum,omitempty"`
}

// ToTypeScriptType converts a ToolProperty to a TypeScript type string
func (tp ToolProperty) ToTypeScriptType() string {
	if len(tp.AnyOf) > 0 {
		var types []string
		for _, anyOf := range tp.AnyOf {
			types = append(types, anyOf.ToTypeScriptType())
		}
		return strings.Join(types, " | ")
	}

	if len(tp.Type) == 0 {
		return "any"
	}

	if len(tp.Type) == 1 {
		return mapToTypeScriptType(tp.Type[0])
	}

	var types []string
	for _, t := range tp.Type {
		types = append(types, mapToTypeScriptType(t))
	}
	return strings.Join(types, " | ")
}

// mapToTypeScriptType maps JSON Schema types to TypeScript types
func mapToTypeScriptType(jsonType string) string {
	switch jsonType {
	case "string":
		return "string"
	case "number", "integer":
		return "number"
	case "boolean":
		return "boolean"
	case "array":
		return "any[]"
	case "object":
		return "Record<string, any>"
	case "null":
		return "null"
	default:
		return "any"
	}
}

289
290
291
292
type ToolFunction struct {
	Name        string `json:"name"`
	Description string `json:"description"`
	Parameters  struct {
Devon Rifkin's avatar
Devon Rifkin committed
293
294
295
296
297
		Type       string                  `json:"type"`
		Defs       any                     `json:"$defs,omitempty"`
		Items      any                     `json:"items,omitempty"`
		Required   []string                `json:"required"`
		Properties map[string]ToolProperty `json:"properties"`
298
299
300
301
302
303
304
305
	} `json:"parameters"`
}

func (t *ToolFunction) String() string {
	bts, _ := json.Marshal(t)
	return string(bts)
}

306
307
// ChatResponse is the response returned by [Client.Chat]. Its fields are
// similar to [GenerateResponse].
Bruce MacDonald's avatar
Bruce MacDonald committed
308
type ChatResponse struct {
309
310
311
	Model      string    `json:"model"`
	CreatedAt  time.Time `json:"created_at"`
	Message    Message   `json:"message"`
Bruce MacDonald's avatar
Bruce MacDonald committed
312
	DoneReason string    `json:"done_reason,omitempty"`
Bruce MacDonald's avatar
Bruce MacDonald committed
313
314
315
316
317
318

	Done bool `json:"done"`

	Metrics
}

319
320
321
322
323
324
325
326
327
328
329
330
331
// DebugInfo contains debug information for template rendering
type DebugInfo struct {
	RenderedTemplate string `json:"rendered_template"`
	ImageCount       int    `json:"image_count,omitempty"`
}

// DebugTemplateResponse is returned when _debug_render_only is set to true
type DebugTemplateResponse struct {
	Model     string    `json:"model"`
	CreatedAt time.Time `json:"created_at"`
	DebugInfo DebugInfo `json:"_debug_info"`
}

Bruce MacDonald's avatar
Bruce MacDonald committed
332
333
334
335
336
337
338
339
340
type Metrics struct {
	TotalDuration      time.Duration `json:"total_duration,omitempty"`
	LoadDuration       time.Duration `json:"load_duration,omitempty"`
	PromptEvalCount    int           `json:"prompt_eval_count,omitempty"`
	PromptEvalDuration time.Duration `json:"prompt_eval_duration,omitempty"`
	EvalCount          int           `json:"eval_count,omitempty"`
	EvalDuration       time.Duration `json:"eval_duration,omitempty"`
}

341
342
// Options specified in [GenerateRequest].  If you add a new option here, also
// add it to the API docs.
343
344
345
346
347
348
349
350
351
type Options struct {
	Runner

	// Predict options used at runtime
	NumKeep          int      `json:"num_keep,omitempty"`
	Seed             int      `json:"seed,omitempty"`
	NumPredict       int      `json:"num_predict,omitempty"`
	TopK             int      `json:"top_k,omitempty"`
	TopP             float32  `json:"top_p,omitempty"`
352
	MinP             float32  `json:"min_p,omitempty"`
353
354
355
356
357
358
359
360
361
362
363
	TypicalP         float32  `json:"typical_p,omitempty"`
	RepeatLastN      int      `json:"repeat_last_n,omitempty"`
	Temperature      float32  `json:"temperature,omitempty"`
	RepeatPenalty    float32  `json:"repeat_penalty,omitempty"`
	PresencePenalty  float32  `json:"presence_penalty,omitempty"`
	FrequencyPenalty float32  `json:"frequency_penalty,omitempty"`
	Stop             []string `json:"stop,omitempty"`
}

// Runner options which must be set when the model is loaded into memory
type Runner struct {
364
365
366
367
368
369
	NumCtx    int   `json:"num_ctx,omitempty"`
	NumBatch  int   `json:"num_batch,omitempty"`
	NumGPU    int   `json:"num_gpu,omitempty"`
	MainGPU   int   `json:"main_gpu,omitempty"`
	UseMMap   *bool `json:"use_mmap,omitempty"`
	NumThread int   `json:"num_thread,omitempty"`
370
371
}

372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
// EmbedRequest is the request passed to [Client.Embed].
type EmbedRequest struct {
	// Model is the model name.
	Model string `json:"model"`

	// Input is the input to embed.
	Input any `json:"input"`

	// KeepAlive controls how long the model will stay loaded in memory following
	// this request.
	KeepAlive *Duration `json:"keep_alive,omitempty"`

	Truncate *bool `json:"truncate,omitempty"`

	// Options lists model-specific options.
387
	Options map[string]any `json:"options"`
388
389
390
391
392
}

// EmbedResponse is the response from [Client.Embed].
type EmbedResponse struct {
	Model      string      `json:"model"`
393
	Embeddings [][]float32 `json:"embeddings"`
394
395
396
397

	TotalDuration   time.Duration `json:"total_duration,omitempty"`
	LoadDuration    time.Duration `json:"load_duration,omitempty"`
	PromptEvalCount int           `json:"prompt_eval_count,omitempty"`
398
399
}

400
// EmbeddingRequest is the request passed to [Client.Embeddings].
Bruce MacDonald's avatar
Bruce MacDonald committed
401
type EmbeddingRequest struct {
402
403
404
405
406
407
408
409
	// Model is the model name.
	Model string `json:"model"`

	// Prompt is the textual prompt to embed.
	Prompt string `json:"prompt"`

	// KeepAlive controls how long the model will stay loaded in memory following
	// this request.
410
	KeepAlive *Duration `json:"keep_alive,omitempty"`
Bruce MacDonald's avatar
Bruce MacDonald committed
411

412
	// Options lists model-specific options.
413
	Options map[string]any `json:"options"`
Bruce MacDonald's avatar
Bruce MacDonald committed
414
415
}

416
// EmbeddingResponse is the response from [Client.Embeddings].
Bruce MacDonald's avatar
Bruce MacDonald committed
417
418
419
420
type EmbeddingResponse struct {
	Embedding []float64 `json:"embedding"`
}

421
// CreateRequest is the request passed to [Client.Create].
422
type CreateRequest struct {
423
424
425
426
427
428
429
430
431
432
433
434
435
	Model    string `json:"model"`
	Stream   *bool  `json:"stream,omitempty"`
	Quantize string `json:"quantize,omitempty"`

	From       string            `json:"from,omitempty"`
	Files      map[string]string `json:"files,omitempty"`
	Adapters   map[string]string `json:"adapters,omitempty"`
	Template   string            `json:"template,omitempty"`
	License    any               `json:"license,omitempty"`
	System     string            `json:"system,omitempty"`
	Parameters map[string]any    `json:"parameters,omitempty"`
	Messages   []Message         `json:"messages,omitempty"`

Michael Yang's avatar
Michael Yang committed
436
	// Deprecated: set the model name with Model instead
Michael Yang's avatar
Michael Yang committed
437
	Name string `json:"name"`
Michael Yang's avatar
Michael Yang committed
438
	// Deprecated: use Quantize instead
439
	Quantization string `json:"quantization,omitempty"`
440
441
}

442
// DeleteRequest is the request passed to [Client.Delete].
443
type DeleteRequest struct {
Michael Yang's avatar
Michael Yang committed
444
445
	Model string `json:"model"`

Michael Yang's avatar
Michael Yang committed
446
	// Deprecated: set the model name with Model instead
447
448
449
	Name string `json:"name"`
}

450
// ShowRequest is the request passed to [Client.Show].
Patrick Devine's avatar
Patrick Devine committed
451
type ShowRequest struct {
452
453
	Model  string `json:"model"`
	System string `json:"system"`
Patrick Devine's avatar
Patrick Devine committed
454
455

	// Template is deprecated
456
	Template string `json:"template"`
457
	Verbose  bool   `json:"verbose"`
458

459
	Options map[string]any `json:"options"`
Michael Yang's avatar
Michael Yang committed
460

Michael Yang's avatar
Michael Yang committed
461
	// Deprecated: set the model name with Model instead
Michael Yang's avatar
Michael Yang committed
462
	Name string `json:"name"`
Patrick Devine's avatar
Patrick Devine committed
463
464
}

465
// ShowResponse is the response returned from [Client.Show].
Patrick Devine's avatar
Patrick Devine committed
466
type ShowResponse struct {
467
468
469
470
471
472
473
474
475
476
477
478
	License       string             `json:"license,omitempty"`
	Modelfile     string             `json:"modelfile,omitempty"`
	Parameters    string             `json:"parameters,omitempty"`
	Template      string             `json:"template,omitempty"`
	System        string             `json:"system,omitempty"`
	Details       ModelDetails       `json:"details,omitempty"`
	Messages      []Message          `json:"messages,omitempty"`
	ModelInfo     map[string]any     `json:"model_info,omitempty"`
	ProjectorInfo map[string]any     `json:"projector_info,omitempty"`
	Tensors       []Tensor           `json:"tensors,omitempty"`
	Capabilities  []model.Capability `json:"capabilities,omitempty"`
	ModifiedAt    time.Time          `json:"modified_at,omitempty"`
Patrick Devine's avatar
Patrick Devine committed
479
480
}

481
// CopyRequest is the request passed to [Client.Copy].
Patrick Devine's avatar
Patrick Devine committed
482
483
484
485
486
type CopyRequest struct {
	Source      string `json:"source"`
	Destination string `json:"destination"`
}

487
// PullRequest is the request passed to [Client.Pull].
Bruce MacDonald's avatar
Bruce MacDonald committed
488
type PullRequest struct {
Michael Yang's avatar
Michael Yang committed
489
	Model    string `json:"model"`
490
491
492
	Insecure bool   `json:"insecure,omitempty"` // Deprecated: ignored
	Username string `json:"username"`           // Deprecated: ignored
	Password string `json:"password"`           // Deprecated: ignored
493
	Stream   *bool  `json:"stream,omitempty"`
Michael Yang's avatar
Michael Yang committed
494

Michael Yang's avatar
Michael Yang committed
495
	// Deprecated: set the model name with Model instead
Michael Yang's avatar
Michael Yang committed
496
	Name string `json:"name"`
Bruce MacDonald's avatar
Bruce MacDonald committed
497
498
}

499
500
// ProgressResponse is the response passed to progress functions like
// [PullProgressFunc] and [PushProgressFunc].
501
type ProgressResponse struct {
502
503
	Status    string `json:"status"`
	Digest    string `json:"digest,omitempty"`
Michael Yang's avatar
Michael Yang committed
504
505
	Total     int64  `json:"total,omitempty"`
	Completed int64  `json:"completed,omitempty"`
Bruce MacDonald's avatar
Bruce MacDonald committed
506
507
}

508
// PushRequest is the request passed to [Client.Push].
509
type PushRequest struct {
Michael Yang's avatar
Michael Yang committed
510
	Model    string `json:"model"`
511
	Insecure bool   `json:"insecure,omitempty"`
512
513
514
	Username string `json:"username"`
	Password string `json:"password"`
	Stream   *bool  `json:"stream,omitempty"`
Michael Yang's avatar
Michael Yang committed
515

Michael Yang's avatar
Michael Yang committed
516
	// Deprecated: set the model name with Model instead
Michael Yang's avatar
Michael Yang committed
517
	Name string `json:"name"`
518
}
519

520
// ListResponse is the response from [Client.List].
Patrick Devine's avatar
Patrick Devine committed
521
type ListResponse struct {
522
	Models []ListModelResponse `json:"models"`
Patrick Devine's avatar
Patrick Devine committed
523
524
}

525
526
527
528
529
530
531
// ProcessResponse is the response from [Client.Process].
type ProcessResponse struct {
	Models []ProcessModelResponse `json:"models"`
}

// ListModelResponse is a single model description in [ListResponse].
type ListModelResponse struct {
532
533
534
535
536
537
	Name       string       `json:"name"`
	Model      string       `json:"model"`
	ModifiedAt time.Time    `json:"modified_at"`
	Size       int64        `json:"size"`
	Digest     string       `json:"digest"`
	Details    ModelDetails `json:"details,omitempty"`
538
539
540
541
}

// ProcessModelResponse is a single model description in [ProcessResponse].
type ProcessModelResponse struct {
542
543
544
545
546
547
548
549
	Name          string       `json:"name"`
	Model         string       `json:"model"`
	Size          int64        `json:"size"`
	Digest        string       `json:"digest"`
	Details       ModelDetails `json:"details,omitempty"`
	ExpiresAt     time.Time    `json:"expires_at"`
	SizeVRAM      int64        `json:"size_vram"`
	ContextLength int          `json:"context_length"`
Patrick Devine's avatar
Patrick Devine committed
550
551
}

Patrick Devine's avatar
Patrick Devine committed
552
553
554
555
type TokenResponse struct {
	Token string `json:"token"`
}

556
// GenerateResponse is the response passed into [GenerateResponseFunc].
Michael Yang's avatar
Michael Yang committed
557
type GenerateResponse struct {
558
559
560
	// Model is the model name that generated the response.
	Model string `json:"model"`

Michael Yang's avatar
Michael Yang committed
561
	// CreatedAt is the timestamp of the response.
562
563
	CreatedAt time.Time `json:"created_at"`

564
565
566
	// Response is the textual response itself.
	Response string `json:"response"`

567
568
569
570
	// Thinking contains the text that was inside thinking tags in the
	// original model output when ChatRequest.Think is enabled.
	Thinking string `json:"thinking,omitempty"`

571
572
573
	// Done specifies if the response is complete.
	Done bool `json:"done"`

574
	// DoneReason is the reason the model stopped generating text.
Bruce MacDonald's avatar
Bruce MacDonald committed
575
	DoneReason string `json:"done_reason,omitempty"`
576

577
578
	// Context is an encoding of the conversation used in this response; this
	// can be sent in the next request to keep a conversational memory.
Michael Yang's avatar
Michael Yang committed
579
	Context []int `json:"context,omitempty"`
580

Bruce MacDonald's avatar
Bruce MacDonald committed
581
	Metrics
Michael Yang's avatar
Michael Yang committed
582
583

	ToolCalls []ToolCall `json:"tool_calls,omitempty"`
584
585
}

586
// ModelDetails provides details about a model.
Patrick Devine's avatar
Patrick Devine committed
587
type ModelDetails struct {
588
	ParentModel       string   `json:"parent_model"`
Patrick Devine's avatar
Patrick Devine committed
589
590
591
592
593
594
595
	Format            string   `json:"format"`
	Family            string   `json:"family"`
	Families          []string `json:"families"`
	ParameterSize     string   `json:"parameter_size"`
	QuantizationLevel string   `json:"quantization_level"`
}

596
597
598
599
600
601
602
// Tensor describes the metadata for a given tensor.
type Tensor struct {
	Name  string   `json:"name"`
	Type  string   `json:"type"`
	Shape []uint64 `json:"shape"`
}

Bruce MacDonald's avatar
Bruce MacDonald committed
603
604
605
func (m *Metrics) Summary() {
	if m.TotalDuration > 0 {
		fmt.Fprintf(os.Stderr, "total duration:       %v\n", m.TotalDuration)
606
607
	}

Bruce MacDonald's avatar
Bruce MacDonald committed
608
609
	if m.LoadDuration > 0 {
		fmt.Fprintf(os.Stderr, "load duration:        %v\n", m.LoadDuration)
Michael Yang's avatar
Michael Yang committed
610
611
	}

Bruce MacDonald's avatar
Bruce MacDonald committed
612
613
	if m.PromptEvalCount > 0 {
		fmt.Fprintf(os.Stderr, "prompt eval count:    %d token(s)\n", m.PromptEvalCount)
614
615
	}

Bruce MacDonald's avatar
Bruce MacDonald committed
616
617
618
	if m.PromptEvalDuration > 0 {
		fmt.Fprintf(os.Stderr, "prompt eval duration: %s\n", m.PromptEvalDuration)
		fmt.Fprintf(os.Stderr, "prompt eval rate:     %.2f tokens/s\n", float64(m.PromptEvalCount)/m.PromptEvalDuration.Seconds())
619
620
	}

Bruce MacDonald's avatar
Bruce MacDonald committed
621
622
	if m.EvalCount > 0 {
		fmt.Fprintf(os.Stderr, "eval count:           %d token(s)\n", m.EvalCount)
623
624
	}

Bruce MacDonald's avatar
Bruce MacDonald committed
625
626
627
	if m.EvalDuration > 0 {
		fmt.Fprintf(os.Stderr, "eval duration:        %s\n", m.EvalDuration)
		fmt.Fprintf(os.Stderr, "eval rate:            %.2f tokens/s\n", float64(m.EvalCount)/m.EvalDuration.Seconds())
628
	}
629
630
}

631
func (opts *Options) FromMap(m map[string]any) error {
632
633
634
635
636
637
638
639
640
641
642
643
644
	valueOpts := reflect.ValueOf(opts).Elem() // names of the fields in the options struct
	typeOpts := reflect.TypeOf(opts).Elem()   // types of the fields in the options struct

	// build map of json struct tags to their types
	jsonOpts := make(map[string]reflect.StructField)
	for _, field := range reflect.VisibleFields(typeOpts) {
		jsonTag := strings.Split(field.Tag.Get("json"), ",")[0]
		if jsonTag != "" {
			jsonOpts[jsonTag] = field
		}
	}

	for key, val := range m {
645
646
		opt, ok := jsonOpts[key]
		if !ok {
647
			slog.Warn("invalid option provided", "option", key)
648
649
			continue
		}
Michael Yang's avatar
Michael Yang committed
650

651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
		field := valueOpts.FieldByName(opt.Name)
		if field.IsValid() && field.CanSet() {
			if val == nil {
				continue
			}

			switch field.Kind() {
			case reflect.Int:
				switch t := val.(type) {
				case int64:
					field.SetInt(t)
				case float64:
					// when JSON unmarshals numbers, it uses float64, not int
					field.SetInt(int64(t))
				default:
					return fmt.Errorf("option %q must be of type integer", key)
				}
			case reflect.Bool:
				val, ok := val.(bool)
				if !ok {
					return fmt.Errorf("option %q must be of type boolean", key)
				}
				field.SetBool(val)
			case reflect.Float32:
				// JSON unmarshals to float64
				val, ok := val.(float64)
				if !ok {
					return fmt.Errorf("option %q must be of type float32", key)
				}
				field.SetFloat(val)
			case reflect.String:
				val, ok := val.(string)
				if !ok {
					return fmt.Errorf("option %q must be of type string", key)
				}
				field.SetString(val)
			case reflect.Slice:
688
689
				// JSON unmarshals to []any, not []string
				val, ok := val.([]any)
690
691
692
				if !ok {
					return fmt.Errorf("option %q must be of type array", key)
				}
693
				// convert []any to []string
694
695
696
				slice := make([]string, len(val))
				for i, item := range val {
					str, ok := item.(string)
697
					if !ok {
698
						return fmt.Errorf("option %q must be of an array of strings", key)
699
					}
700
					slice[i] = str
701
				}
702
				field.Set(reflect.ValueOf(slice))
703
704
705
706
707
708
709
710
711
712
713
			case reflect.Pointer:
				var b bool
				if field.Type() == reflect.TypeOf(&b) {
					val, ok := val.(bool)
					if !ok {
						return fmt.Errorf("option %q must be of type boolean", key)
					}
					field.Set(reflect.ValueOf(&val))
				} else {
					return fmt.Errorf("unknown type loading config params: %v %v", field.Kind(), field.Type())
				}
714
715
			default:
				return fmt.Errorf("unknown type loading config params: %v", field.Kind())
716
717
718
			}
		}
	}
719

720
721
722
	return nil
}

723
724
// DefaultOptions is the default set of options for [GenerateRequest]; these
// values are used unless the user specifies other values explicitly.
Michael Yang's avatar
Michael Yang committed
725
726
func DefaultOptions() Options {
	return Options{
727
		// options set on request to runner
728
729
730
731
		NumPredict: -1,

		// set a minimal num_keep to avoid issues on context shifts
		NumKeep:          4,
Michael Yang's avatar
Michael Yang committed
732
733
734
735
		Temperature:      0.8,
		TopK:             40,
		TopP:             0.9,
		TypicalP:         1.0,
736
737
738
739
740
		RepeatLastN:      64,
		RepeatPenalty:    1.1,
		PresencePenalty:  0.0,
		FrequencyPenalty: 0.0,
		Seed:             -1,
Michael Yang's avatar
Michael Yang committed
741

742
743
		Runner: Runner{
			// options set when the model is loaded
744
			NumCtx:    int(envconfig.ContextLength()),
Michael Yang's avatar
Michael Yang committed
745
746
			NumBatch:  512,
			NumGPU:    -1, // -1 here indicates that NumGPU should be set dynamically
747
			NumThread: 0,  // let the runtime decide
748
			UseMMap:   nil,
749
		},
Michael Yang's avatar
Michael Yang committed
750
	}
Jeffrey Morgan's avatar
Jeffrey Morgan committed
751
}
Michael Yang's avatar
Michael Yang committed
752

Michael Yang's avatar
Michael Yang committed
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
// ThinkValue represents a value that can be a boolean or a string ("high", "medium", "low")
type ThinkValue struct {
	// Value can be a bool or string
	Value interface{}
}

// IsValid checks if the ThinkValue is valid
func (t *ThinkValue) IsValid() bool {
	if t == nil || t.Value == nil {
		return true // nil is valid (means not set)
	}

	switch v := t.Value.(type) {
	case bool:
		return true
	case string:
		return v == "high" || v == "medium" || v == "low"
	default:
		return false
	}
}

// IsBool returns true if the value is a boolean
func (t *ThinkValue) IsBool() bool {
	if t == nil || t.Value == nil {
		return false
	}
	_, ok := t.Value.(bool)
	return ok
}

// IsString returns true if the value is a string
func (t *ThinkValue) IsString() bool {
	if t == nil || t.Value == nil {
		return false
	}
	_, ok := t.Value.(string)
	return ok
}

793
794
// Bool returns the value as a bool (true if enabled in any way)
func (t *ThinkValue) Bool() bool {
Michael Yang's avatar
Michael Yang committed
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
	if t == nil || t.Value == nil {
		return false
	}

	switch v := t.Value.(type) {
	case bool:
		return v
	case string:
		// Any string value ("high", "medium", "low") means thinking is enabled
		return v == "high" || v == "medium" || v == "low"
	default:
		return false
	}
}

810
811
// String returns the value as a string
func (t *ThinkValue) String() string {
Michael Yang's avatar
Michael Yang committed
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
	if t == nil || t.Value == nil {
		return ""
	}

	switch v := t.Value.(type) {
	case string:
		return v
	case bool:
		if v {
			return "medium" // Default level when just true
		}
		return ""
	default:
		return ""
	}
}

// UnmarshalJSON implements json.Unmarshaler
func (t *ThinkValue) UnmarshalJSON(data []byte) error {
	// Try to unmarshal as bool first
	var b bool
	if err := json.Unmarshal(data, &b); err == nil {
		t.Value = b
		return nil
	}

	// Try to unmarshal as string
	var s string
	if err := json.Unmarshal(data, &s); err == nil {
		// Validate string values
		if s != "high" && s != "medium" && s != "low" {
			return fmt.Errorf("invalid think value: %q (must be \"high\", \"medium\", \"low\", true, or false)", s)
		}
		t.Value = s
		return nil
	}

	return fmt.Errorf("think must be a boolean or string (\"high\", \"medium\", \"low\")")
}

// MarshalJSON implements json.Marshaler
func (t *ThinkValue) MarshalJSON() ([]byte, error) {
	if t == nil || t.Value == nil {
		return []byte("null"), nil
	}
	return json.Marshal(t.Value)
}

Michael Yang's avatar
Michael Yang committed
860
861
862
863
type Duration struct {
	time.Duration
}

864
865
866
867
868
869
870
func (d Duration) MarshalJSON() ([]byte, error) {
	if d.Duration < 0 {
		return []byte("-1"), nil
	}
	return []byte("\"" + d.Duration.String() + "\""), nil
}

Michael Yang's avatar
Michael Yang committed
871
872
873
874
875
876
877
878
879
880
881
func (d *Duration) UnmarshalJSON(b []byte) (err error) {
	var v any
	if err := json.Unmarshal(b, &v); err != nil {
		return err
	}

	d.Duration = 5 * time.Minute

	switch t := v.(type) {
	case float64:
		if t < 0 {
bnorick's avatar
bnorick committed
882
			d.Duration = time.Duration(math.MaxInt64)
883
		} else {
884
			d.Duration = time.Duration(int(t) * int(time.Second))
Michael Yang's avatar
Michael Yang committed
885
886
887
888
889
890
		}
	case string:
		d.Duration, err = time.ParseDuration(t)
		if err != nil {
			return err
		}
891
		if d.Duration < 0 {
bnorick's avatar
bnorick committed
892
			d.Duration = time.Duration(math.MaxInt64)
893
		}
894
895
	default:
		return fmt.Errorf("Unsupported type: '%s'", reflect.TypeOf(v))
Michael Yang's avatar
Michael Yang committed
896
897
898
899
	}

	return nil
}
900
901

// FormatParams converts specified parameter options to their correct types
902
func FormatParams(params map[string][]string) (map[string]any, error) {
903
904
905
906
907
908
909
910
911
912
913
914
915
	opts := Options{}
	valueOpts := reflect.ValueOf(&opts).Elem() // names of the fields in the options struct
	typeOpts := reflect.TypeOf(opts)           // types of the fields in the options struct

	// build map of json struct tags to their types
	jsonOpts := make(map[string]reflect.StructField)
	for _, field := range reflect.VisibleFields(typeOpts) {
		jsonTag := strings.Split(field.Tag.Get("json"), ",")[0]
		if jsonTag != "" {
			jsonOpts[jsonTag] = field
		}
	}

916
	out := make(map[string]any)
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
	// iterate params and set values based on json struct tags
	for key, vals := range params {
		if opt, ok := jsonOpts[key]; !ok {
			return nil, fmt.Errorf("unknown parameter '%s'", key)
		} else {
			field := valueOpts.FieldByName(opt.Name)
			if field.IsValid() && field.CanSet() {
				switch field.Kind() {
				case reflect.Float32:
					floatVal, err := strconv.ParseFloat(vals[0], 32)
					if err != nil {
						return nil, fmt.Errorf("invalid float value %s", vals)
					}

					out[key] = float32(floatVal)
				case reflect.Int:
					intVal, err := strconv.ParseInt(vals[0], 10, 64)
					if err != nil {
						return nil, fmt.Errorf("invalid int value %s", vals)
					}

					out[key] = intVal
				case reflect.Bool:
					boolVal, err := strconv.ParseBool(vals[0])
					if err != nil {
						return nil, fmt.Errorf("invalid bool value %s", vals)
					}

					out[key] = boolVal
				case reflect.String:
					out[key] = vals[0]
				case reflect.Slice:
					// TODO: only string slices are supported right now
					out[key] = vals
951
952
953
954
955
956
957
958
959
960
961
				case reflect.Pointer:
					var b bool
					if field.Type() == reflect.TypeOf(&b) {
						boolVal, err := strconv.ParseBool(vals[0])
						if err != nil {
							return nil, fmt.Errorf("invalid bool value %s", vals)
						}
						out[key] = &boolVal
					} else {
						return nil, fmt.Errorf("unknown type %s for %s", field.Kind(), key)
					}
962
963
964
965
966
967
968
969
970
				default:
					return nil, fmt.Errorf("unknown type %s for %s", field.Kind(), key)
				}
			}
		}
	}

	return out, nil
}