ggml.go 23.4 KB
Newer Older
Michael Yang's avatar
Michael Yang committed
1
2
package ggml

3
4
5
6
7
8
// #cgo CPPFLAGS: -I${SRCDIR}/ggml/include
// #include <stdlib.h>
// #include <stdint.h>
// #include "ggml.h"
// #include "ggml-cpu.h"
// #include "ggml-backend.h"
Michael Yang's avatar
Michael Yang committed
9
10
11
import "C"

import (
12
	"errors"
Michael Yang's avatar
Michael Yang committed
13
14
15
	"fmt"
	"io"
	"log/slog"
16
	"maps"
Michael Yang's avatar
Michael Yang committed
17
	"os"
18
19
20
21
	"slices"
	"strconv"
	"strings"
	"unicode"
Michael Yang's avatar
Michael Yang committed
22
23
24
25
26
	"unsafe"

	"github.com/ollama/ollama/format"
	fs "github.com/ollama/ollama/fs/ggml"
	"github.com/ollama/ollama/ml"
27
	ggml "github.com/ollama/ollama/ml/backend/ggml/ggml/src"
Michael Yang's avatar
Michael Yang committed
28
29
30
	"golang.org/x/sync/errgroup"
)

Michael Yang's avatar
Michael Yang committed
31
32
33
34
35
func devices() []*C.struct_ggml_backend_device {
	ggml.OnceLoad()
	ds := make([]*C.struct_ggml_backend_device, C.ggml_backend_dev_count())
	for i := range ds {
		ds[i] = C.ggml_backend_dev_get(C.size_t(i))
Michael Yang's avatar
Michael Yang committed
36
	}
Michael Yang's avatar
Michael Yang committed
37
38

	return ds
39
}
Michael Yang's avatar
Michael Yang committed
40
41

type Backend struct {
42
43
44
	meta    *fs.GGML
	sched   *C.struct_ggml_backend_sched
	tensors map[string]*C.struct_ggml_tensor
Michael Yang's avatar
Michael Yang committed
45
46

	// input is the backend used for inputs
47
	input *C.struct_ggml_backend_buffer_type
Michael Yang's avatar
Michael Yang committed
48
49

	// output is the backend used for outputs
50
	output *C.struct_ggml_backend_buffer_type
Michael Yang's avatar
Michael Yang committed
51
52

	// layers is the backend used for repeating layers
53
	layers map[int]*C.struct_ggml_backend_buffer_type
54

55
	flashAttention bool
Michael Yang's avatar
Michael Yang committed
56
57
58

	// maxGraphNodes is the maximum allowed number of graph nodes in this scheduler
	maxGraphNodes int
Michael Yang's avatar
Michael Yang committed
59
60
}

61
func New(r *os.File, params ml.BackendParams) (ml.Backend, error) {
Michael Yang's avatar
Michael Yang committed
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
	meta, n, err := fs.Decode(r, -1)
	if err != nil {
		return nil, err
	}

	slog.Info(
		"",
		"architecture", meta.KV().Architecture(),
		"file_type", meta.KV().FileType(),
		"name", meta.KV().String("general.name"),
		"description", meta.KV().String("general.description"),
		"num_tensors", len(meta.Tensors().Items()),
		"num_key_values", len(meta.KV()),
	)

77
	type deviceBufferType struct {
78
79
80
81
82
		d   *C.struct_ggml_backend_device
		bts []*C.struct_ggml_backend_buffer_type
	}

	var cpus, accels, gpus []*C.struct_ggml_backend_device
Michael Yang's avatar
Michael Yang committed
83
	for _, d := range devices() {
84
85
		switch C.ggml_backend_dev_type(d) {
		case C.GGML_BACKEND_DEVICE_TYPE_CPU:
86
87
88
89
			if len(cpus) == 0 {
				// only the first cpu device should be used
				cpus = append(cpus, d)
			}
90
91
		case C.GGML_BACKEND_DEVICE_TYPE_ACCEL:
			accels = append(accels, d)
Michael Yang's avatar
Michael Yang committed
92
		case C.GGML_BACKEND_DEVICE_TYPE_GPU:
93
			gpus = append(gpus, d)
Michael Yang's avatar
Michael Yang committed
94
95
96
		}
	}

Michael Yang's avatar
Michael Yang committed
97
	// create list of buffer types for the cpu
Michael Yang's avatar
Michael Yang committed
98
	cpuDeviceBufferType := deviceBufferType{d: C.ggml_backend_dev_by_type(C.GGML_BACKEND_DEVICE_TYPE_CPU)}
99
100
101
102
	for _, d := range append(accels, append(gpus, cpus...)...) {
		switch C.ggml_backend_dev_type(d) {
		case C.GGML_BACKEND_DEVICE_TYPE_CPU,
			C.GGML_BACKEND_DEVICE_TYPE_ACCEL:
Michael Yang's avatar
Michael Yang committed
103
			cpuDeviceBufferType.bts = append(cpuDeviceBufferType.bts, C.ggml_backend_dev_buffer_type(d))
Michael Yang's avatar
Michael Yang committed
104
		}
105
106
	}

Michael Yang's avatar
Michael Yang committed
107
	// create list of buffer types for each gpu
108
	var gpuDeviceBufferTypes []deviceBufferType
109
110
	for _, d := range gpus {
		bt := C.ggml_backend_dev_buffer_type(d)
111
		gpuDeviceBufferTypes = append(gpuDeviceBufferTypes, deviceBufferType{
112
			d:   d,
Michael Yang's avatar
Michael Yang committed
113
			bts: append([]*C.struct_ggml_backend_buffer_type{bt}, cpuDeviceBufferType.bts...),
114
		})
Michael Yang's avatar
Michael Yang committed
115
116
	}

Michael Yang's avatar
Michael Yang committed
117
118
119
120
121
	useDefaultSplit := true
	for _, s := range params.TensorSplit {
		if s != 0 {
			useDefaultSplit = false
			break
122
		}
Michael Yang's avatar
Michael Yang committed
123
	}
124

Michael Yang's avatar
Michael Yang committed
125
126
127
128
	// calculate splits
	splits := make([]float32, len(gpus))
	if useDefaultSplit {
		// default: split on free memory
129
130
131
132
133
		for i := range splits {
			var free, total C.size_t
			C.ggml_backend_dev_memory(gpus[i], &free, &total)
			splits[i] = float32(free)
		}
Michael Yang's avatar
Michael Yang committed
134
135
	} else {
		splits = params.TensorSplit
136
137
138
	}

	var sum float32
Michael Yang's avatar
Michael Yang committed
139
	// cumulative sum of all splits
140
141
142
143
144
	for i := range splits {
		sum += splits[i]
		splits[i] = sum
	}

Michael Yang's avatar
Michael Yang committed
145
	// normalize splits
146
	for i := range splits {
147
		splits[i] /= sum
148
149
	}

Michael Yang's avatar
Michael Yang committed
150
	// inputs always use cpu
Michael Yang's avatar
Michael Yang committed
151
	input := cpuDeviceBufferType
152

153
	blocks := int(meta.KV().BlockCount())
Michael Yang's avatar
Michael Yang committed
154
155
156
157

	// define a range of gpu layers. anything outside of this range is assigned to the cpu
	gpuRangeStart := max(0, blocks-params.NumGPULayers)
	gpuRangeStop := min(gpuRangeStart+params.NumGPULayers, blocks+1)
Michael Yang's avatar
Michael Yang committed
158
	assignLayer := func(i int) deviceBufferType {
Michael Yang's avatar
Michael Yang committed
159
		if i < gpuRangeStart || i >= gpuRangeStop {
Michael Yang's avatar
Michael Yang committed
160
			return cpuDeviceBufferType
161
		}
162

Michael Yang's avatar
Michael Yang committed
163
		index := slices.IndexFunc(splits, func(f float32) bool { return float32(i-gpuRangeStart)/float32(gpuRangeStop-gpuRangeStart) < f })
164
		if index < 0 || index >= len(gpuDeviceBufferTypes) {
Michael Yang's avatar
Michael Yang committed
165
			return cpuDeviceBufferType
166
167
168
		}

		return gpuDeviceBufferTypes[index]
169
170
	}

Michael Yang's avatar
Michael Yang committed
171
	// repeating layers are assigned based on their index in reverse order, e.g. i / (block_count + 1)
172
	layers := make([]deviceBufferType, blocks)
173
	for i := range layers {
174
		layers[i] = assignLayer(i)
175
176
	}

Michael Yang's avatar
Michael Yang committed
177
	// outputs are assigned iff allowed by splits and configured number of gpu layers
178
	output := assignLayer(blocks)
179
180
181

	maxTensors := len(meta.Tensors().Items())
	maxTensors += 1
Michael Yang's avatar
Michael Yang committed
182
	// each layer has at most 2 extra tensors for rope operations
183
184
	maxTensors += blocks * 2

185
186
187
188
189
	type tensor struct {
		source *fs.Tensor
		target string
	}

Michael Yang's avatar
Michael Yang committed
190
	// some tensors are mapped to different names so keep a list
191
192
	targets := make(map[string][]string)

Michael Yang's avatar
Michael Yang committed
193
	// contexts are shared by tensors of the same buffer type
194
	ctxs := make(map[*C.struct_ggml_backend_buffer_type]*C.struct_ggml_context)
195
	createTensor := func(t tensor, bts []*C.struct_ggml_backend_buffer_type) *C.struct_ggml_tensor {
196
197
198
199
200
201
202
		for _, bt := range bts {
			if _, ok := ctxs[bt]; !ok {
				ctxs[bt] = C.ggml_init(C.struct_ggml_init_params{
					mem_size: C.ggml_tensor_overhead() * C.size_t(maxTensors),
					no_alloc: true,
				})
			}
Michael Yang's avatar
Michael Yang committed
203

204
205
206
207
208
209
210
211
			targets[t.source.Name] = append(targets[t.source.Name], t.target)

			name := t.source.Name
			if t.target != "" {
				name = t.target
			}

			cname := C.CString(name)
Michael Yang's avatar
Michael Yang committed
212
			defer C.free(unsafe.Pointer(cname))
213
214
215
216
			if tt := C.ggml_get_tensor(ctxs[bt], cname); tt != nil {
				return tt
			}

217
			tt := C.ggml_new_tensor(ctxs[bt], t.source.Kind, C.int(len(t.source.Shape)), (*C.int64_t)(unsafe.Pointer(&t.source.Shape[0])))
Michael Yang's avatar
Michael Yang committed
218
219
			C.ggml_set_name(tt, cname)

220
			slog.Debug("created tensor", "name", name, "shape", t.source.Shape, "dtype", t.source.Kind, "buffer_type", C.GoString(C.ggml_backend_buft_name(bt)))
221
222
223
224
225
			//nolint:staticcheck // TODO: check if buffer type supports this tensor
			return tt
		}

		return nil
Michael Yang's avatar
Michael Yang committed
226
227
	}

228
	contains := func(s string, parts ...string) bool {
229
230
231
232
233
234
235
236
		split := strings.Split(s, ".")
		for _, part := range parts {
			if slices.Contains(split, part) {
				return true
			}
		}

		return false
Michael Yang's avatar
Michael Yang committed
237
238
	}

239
240
	for _, t := range meta.Tensors().Items() {
		switch {
241
		case contains(t.Name, "position_embd", "token_embd", "token_norm_embd", "token_types"):
242
			createTensor(tensor{source: t}, input.bts)
243
		case contains(t.Name, "cls", "output", "output_norm"):
244
			createTensor(tensor{source: t}, output.bts)
245
		case strings.HasPrefix(t.Name, "v.") || strings.HasPrefix(t.Name, "mm."):
Michael Yang's avatar
Michael Yang committed
246
			// TODO: assign vision tensors to the gpu if possible
247
			createTensor(tensor{source: t}, input.bts)
248
		default:
Michael Yang's avatar
Michael Yang committed
249
250
251
252
			layerIndex := -1
			if fields := strings.FieldsFunc(t.Name, func(r rune) bool { return !unicode.IsNumber(r) }); len(fields) > 0 {
				if i, err := strconv.Atoi(fields[0]); err == nil {
					layerIndex = i
253
				}
Michael Yang's avatar
Michael Yang committed
254
			}
255

Michael Yang's avatar
Michael Yang committed
256
257
			if layerIndex >= 0 {
				createTensor(tensor{source: t}, layers[layerIndex].bts)
258
			} else {
Michael Yang's avatar
Michael Yang committed
259
260
				// this is a repeating tensor that doesn't explicitly associated with a layer so
				// duplicate it for each layer
261
262
263
264
265
				for i, layer := range layers {
					createTensor(tensor{
						source: t,
						target: "blk." + strconv.Itoa(i) + "." + t.Name,
					}, layer.bts)
266
267
268
269
				}
			}
		}
	}
Michael Yang's avatar
Michael Yang committed
270

Michael Yang's avatar
Michael Yang committed
271
272
	// allocate buffers for each context
	bbs := make(map[*C.struct_ggml_context]*C.struct_ggml_backend_buffer, len(ctxs))
273
274
275
276
277
278
279
	for bt, c := range ctxs {
		if C.ggml_get_first_tensor(c) == nil {
			continue
		}

		b := C.ggml_backend_alloc_ctx_tensors_from_buft(c, bt)
		C.ggml_backend_buffer_set_usage(b, C.GGML_BACKEND_BUFFER_USAGE_WEIGHTS)
Michael Yang's avatar
Michael Yang committed
280
		bbs[c] = b
281
282
283
	}

	for bs := range maps.Values(bbs) {
Michael Yang's avatar
Michael Yang committed
284
		slog.Info("model weights", "buffer", C.GoString(C.ggml_backend_buffer_name(bs)), "size", format.HumanBytes2(uint64(C.ggml_backend_buffer_get_size(bs))))
285
286
	}

Michael Yang's avatar
Michael Yang committed
287
	// map tensor names to tensors for easy lookup later
288
289
290
291
292
293
294
	tensors := make(map[string]*C.struct_ggml_tensor)
	for _, c := range ctxs {
		for t := C.ggml_get_first_tensor(c); t != nil; t = C.ggml_get_next_tensor(c, t) {
			tensors[C.GoString(C.ggml_get_name(t))] = t
		}
	}

Michael Yang's avatar
Michael Yang committed
295
	// concurrently read in tensor data. uses a section reader which is safe for concurrent reads
296
	sr := io.NewSectionReader(r, int64(meta.Tensors().Offset), n-int64(meta.Tensors().Offset))
Michael Yang's avatar
Michael Yang committed
297
	var g errgroup.Group
298
	for _, t := range meta.Tensors().Items() {
299
300
301
302
303
		for _, target := range targets[t.Name] {
			g.Go(func() error {
				if target == "" {
					target = t.Name
				}
304

305
306
307
308
				tt, ok := tensors[target]
				if !ok {
					return fmt.Errorf("unassigned tensor: %s", t.Name)
				}
Michael Yang's avatar
Michael Yang committed
309

310
311
312
313
314
				bts := make([]byte, t.Size())
				n, err := io.ReadFull(io.NewSectionReader(sr, int64(t.Offset), int64(t.Size())), bts)
				if err != nil {
					return err
				}
Michael Yang's avatar
Michael Yang committed
315

316
317
318
				if n != len(bts) {
					return errors.New("short read")
				}
Michael Yang's avatar
Michael Yang committed
319

320
321
322
323
				C.ggml_backend_tensor_set(tt, unsafe.Pointer(&bts[0]), 0, C.size_t(t.Size()))
				return nil
			})
		}
Michael Yang's avatar
Michael Yang committed
324
325
	}

326
	if g.Wait() != nil {
Michael Yang's avatar
Michael Yang committed
327
328
329
		return nil, err
	}

330
331
	// map devices to backend buffer types so new tensors can be assigned to the correct device
	deviceBufferTypes := make(map[*C.struct_ggml_backend_device]*C.struct_ggml_backend_buffer_type)
Michael Yang's avatar
Michael Yang committed
332
333
334
335

	// create backends and buffer types used for the compute graph scheduler
	var schedBackends []*C.struct_ggml_backend
	var schedBufts []*C.struct_ggml_backend_buffer_type
336
337
338
339
	for _, d := range append(gpus, append(accels, cpus...)...) {
		b := C.ggml_backend_dev_init(d, nil)
		bt := C.ggml_backend_get_default_buffer_type(b)
		if d := C.ggml_backend_get_device(b); C.ggml_backend_dev_type(d) == C.GGML_BACKEND_DEVICE_TYPE_CPU && len(gpus) > 0 {
340
341
			// use the first gpu host buffer type for gpu if possible
			if hbt := C.ggml_backend_dev_host_buffer_type(gpus[0]); hbt != nil {
342
343
344
345
				bt = hbt
			}
		}

346
347
348
		deviceBufferTypes[d] = bt

		schedBackends = append(schedBackends, b)
Michael Yang's avatar
Michael Yang committed
349
		schedBufts = append(schedBufts, bt)
350

351
		slog.Info("compute graph", "backend", C.GoString(C.ggml_backend_name(b)), "buffer_type", C.GoString(C.ggml_backend_buft_name(bt)))
352
353

		if C.ggml_backend_is_cpu(b) {
Michael Yang's avatar
Michael Yang committed
354
			// set number of threads for cpu backend
355
356
			C.ggml_backend_cpu_set_n_threads(b, C.int(params.NumThreads))
		}
357
358
	}

Michael Yang's avatar
Michael Yang committed
359
	maxGraphNodes := max(8192, len(meta.Tensors().Items())*5)
Michael Yang's avatar
Michael Yang committed
360
	return &Backend{
361
		flashAttention: params.FlashAttention,
362
363
		meta:           meta,
		tensors:        tensors,
364
		sched: C.ggml_backend_sched_new(
Michael Yang's avatar
Michael Yang committed
365
366
367
368
			(*C.ggml_backend_t)(unsafe.Pointer(&schedBackends[0])),
			(*C.ggml_backend_buffer_type_t)(unsafe.Pointer(&schedBufts[0])),
			C.int(len(schedBackends)),
			C.size_t(maxGraphNodes),
369
370
			true,
		),
371
372
373
374
		input:  deviceBufferTypes[input.d],
		output: deviceBufferTypes[output.d],
		layers: func() map[int]*C.struct_ggml_backend_buffer_type {
			m := make(map[int]*C.struct_ggml_backend_buffer_type)
375
			for i, layer := range layers {
376
				m[i] = deviceBufferTypes[layer.d]
377
378
379
			}
			return m
		}(),
Michael Yang's avatar
Michael Yang committed
380
		maxGraphNodes: maxGraphNodes,
Michael Yang's avatar
Michael Yang committed
381
382
383
384
385
386
387
388
389
390
391
392
	}, nil
}

func init() {
	ml.RegisterBackend("ggml", New)
}

func (b *Backend) Config() ml.Config {
	return b.meta.KV()
}

func (b *Backend) Get(name string) ml.Tensor {
393
394
	if t, ok := b.tensors[name]; ok {
		return &Tensor{b: b, t: t}
Michael Yang's avatar
Michael Yang committed
395
396
397
398
399
400
	}

	return nil
}

func (b *Backend) NewContext() ml.Context {
Michael Yang's avatar
Michael Yang committed
401
	return b.NewContextSize(b.maxGraphNodes)
402
403
404
}

func (b *Backend) NewContextSize(n int) ml.Context {
Jesse Gross's avatar
Jesse Gross committed
405
406
407
408
	if n > b.maxGraphNodes {
		panic(fmt.Errorf("requested number of graph nodes (%v) for new context exceeds maximum (%v)", n, b.maxGraphNodes))
	}

Michael Yang's avatar
Michael Yang committed
409
	return &Context{
410
411
		b:             b,
		maxGraphNodes: n,
412
		ctx: C.ggml_init(C.struct_ggml_init_params{
413
			mem_size: C.size_t(n)*C.ggml_tensor_overhead() + C.ggml_graph_overhead_custom(C.size_t(n), false),
414
415
			no_alloc: true,
		}),
Michael Yang's avatar
Michael Yang committed
416
417
418
	}
}

419
func (b *Backend) CacheConfig() ml.CacheConfig {
420
421
422
423
424
	if b.flashAttention {
		return ml.CacheConfig{CachePadding: 256, MaskDType: ml.DTypeF16, MaskBatchPadding: C.GGML_KQ_MASK_PAD}
	} else {
		return ml.CacheConfig{CachePadding: 32, PermutedV: true}
	}
425
426
}

Michael Yang's avatar
Michael Yang committed
427
type Context struct {
428
	b *Backend
Michael Yang's avatar
Michael Yang committed
429

430
	ctx   *C.struct_ggml_context
Michael Yang's avatar
Michael Yang committed
431
	graph *C.struct_ggml_cgraph
432

433
434
	// buft is the buffer type used for new tensors
	buft *C.struct_ggml_backend_buffer_type
435

Michael Yang's avatar
Michael Yang committed
436
	// maxGraphNodes is the maximum allowed number of graph nodes in this context
437
	maxGraphNodes int
Michael Yang's avatar
Michael Yang committed
438
439
}

Michael Yang's avatar
Michael Yang committed
440
441
func (c Context) Input() ml.Context {
	if c.b.input != nil {
442
443
444
		return &Context{
			b:             c.b,
			ctx:           c.ctx,
445
			buft:          c.b.input,
446
447
448
449
			maxGraphNodes: c.maxGraphNodes,
		}
	}

Michael Yang's avatar
Michael Yang committed
450
	return &c
451
452
}

Michael Yang's avatar
Michael Yang committed
453
454
func (c Context) Output() ml.Context {
	if c.b.output != nil {
455
456
457
		return &Context{
			b:             c.b,
			ctx:           c.ctx,
458
			buft:          c.b.output,
459
460
461
462
			maxGraphNodes: c.maxGraphNodes,
		}
	}

Michael Yang's avatar
Michael Yang committed
463
	return &c
464
465
}

Michael Yang's avatar
Michael Yang committed
466
func (c Context) Layer(i int) ml.Context {
467
	if buft, ok := c.b.layers[i]; ok {
468
469
470
		return &Context{
			b:             c.b,
			ctx:           c.ctx,
471
			buft:          buft,
472
473
474
475
			maxGraphNodes: c.maxGraphNodes,
		}
	}

Michael Yang's avatar
Michael Yang committed
476
	return &c
477
478
}

479
func (c *Context) Forward(tensors ...ml.Tensor) ml.Context {
Michael Yang's avatar
Michael Yang committed
480
	if c.graph == nil {
481
		c.graph = C.ggml_new_graph_custom(c.ctx, C.size_t(c.maxGraphNodes), false)
Michael Yang's avatar
Michael Yang committed
482
483
	}

484
485
486
487
488
	for _, tensor := range tensors {
		C.ggml_build_forward_expand(c.graph, tensor.(*Tensor).t)
	}

	return c
Michael Yang's avatar
Michael Yang committed
489
490
}

Michael Yang's avatar
Michael Yang committed
491
func (c Context) Compute(tensors ...ml.Tensor) {
492
	C.ggml_backend_sched_graph_compute_async(c.b.sched, c.graph)
Michael Yang's avatar
Michael Yang committed
493
	C.ggml_backend_sched_reset(c.b.sched)
Michael Yang's avatar
Michael Yang committed
494

495
496
497
	needSync := true
	sync := func() {
		if needSync {
498
			C.ggml_backend_sched_synchronize(c.b.sched)
499
500
501
			needSync = false
		}
	}
Michael Yang's avatar
Michael Yang committed
502

503
504
505
	for _, t := range tensors {
		if C.ggml_nbytes(t.(*Tensor).t) > 0 {
			t.(*Tensor).sync = sync
506
507
		}
	}
Michael Yang's avatar
Michael Yang committed
508
509
}

Michael Yang's avatar
Michael Yang committed
510
func (c Context) MaxGraphNodes() int {
511
	return c.maxGraphNodes
Jesse Gross's avatar
Jesse Gross committed
512
513
}

514
515
516
func shapeToGGML(shape []int) *C.int64_t {
	sh := make([]C.int64_t, len(shape))
	for i, s := range shape {
517
		sh[i] = C.int64_t(s)
518
519
520
521
522
	}

	return &sh[0]
}

523
func (c Context) newTensor(dtype ml.DType, shape []int) ml.Tensor {
524
525
526
527
	if c.buft == nil {
		panic("set Input, Output, or Layer before creating tensors")
	}

Michael Yang's avatar
Michael Yang committed
528
529
530
531
532
533
534
535
536
537
538
539
	var cdtype uint32
	switch dtype {
	case ml.DTypeF32:
		cdtype = C.GGML_TYPE_F32
	case ml.DTypeF16:
		cdtype = C.GGML_TYPE_F16
	case ml.DTypeI32:
		cdtype = C.GGML_TYPE_I32
	default:
		panic("unsupported dtype")
	}

Jesse Gross's avatar
Jesse Gross committed
540
	if len(shape) < 1 || shape[0] == 0 {
Michael Yang's avatar
Michael Yang committed
541
542
543
		var shape C.int64_t = 0
		return &Tensor{b: c.b, t: C.ggml_new_tensor(c.ctx, cdtype, 1, &shape)}
	} else if len(shape) > 4 {
Michael Yang's avatar
Michael Yang committed
544
545
546
547
548
549
550
551
552
		panic("unsupported number of dimensions")
	}

	for _, dim := range shape {
		if dim < 1 {
			panic("invalid shape")
		}
	}

Michael Yang's avatar
Michael Yang committed
553
	t := C.ggml_new_tensor(c.ctx, cdtype, C.int(len(shape)), shapeToGGML(shape))
554
	b := C.ggml_backend_buft_alloc_buffer(c.buft, C.ggml_nbytes(t))
Michael Yang's avatar
Michael Yang committed
555
	C.ggml_backend_tensor_alloc(b, t, C.ggml_backend_buffer_get_base(b))
556
	return &Tensor{b: c.b, t: t}
557
558
559
}

func (c Context) Empty(dtype ml.DType, shape ...int) ml.Tensor {
560
	return c.newTensor(dtype, shape)
561
562
563
}

func (c Context) Zeros(dtype ml.DType, shape ...int) ml.Tensor {
564
	t := c.newTensor(dtype, shape)
565
566
	C.ggml_set_zero(t.(*Tensor).t)
	return t
Michael Yang's avatar
Michael Yang committed
567
568
}

569
func checkShape[S ~[]E, E any](s S, shape ...int) error {
Michael Yang's avatar
Michael Yang committed
570
	n := len(s)
Jesse Gross's avatar
Jesse Gross committed
571
572
573
574
575

	if n == 0 {
		return nil
	}

Michael Yang's avatar
Michael Yang committed
576
577
578
579
580
	for _, v := range shape {
		n /= v
	}

	if n != 1 {
581
		return fmt.Errorf("invalid shape: %v", shape)
Michael Yang's avatar
Michael Yang committed
582
583
	}

584
	return nil
Michael Yang's avatar
Michael Yang committed
585
586
587
}

func (c Context) FromFloatSlice(s []float32, shape ...int) (ml.Tensor, error) {
Jesse Gross's avatar
Jesse Gross committed
588
	if err := checkShape(s, shape...); err != nil {
589
590
591
592
		return nil, err
	}

	t := c.newTensor(ml.DTypeF32, shape)
Jesse Gross's avatar
Jesse Gross committed
593
594
595
596
	if len(s) > 0 {
		C.ggml_backend_tensor_set(t.(*Tensor).t, unsafe.Pointer(&s[0]), 0, C.ggml_nbytes(t.(*Tensor).t))
	}

597
	return t, nil
Michael Yang's avatar
Michael Yang committed
598
599
600
}

func (c Context) FromIntSlice(s []int32, shape ...int) (ml.Tensor, error) {
Jesse Gross's avatar
Jesse Gross committed
601
	if err := checkShape(s, shape...); err != nil {
602
603
604
605
		return nil, err
	}

	t := c.newTensor(ml.DTypeI32, shape)
Jesse Gross's avatar
Jesse Gross committed
606
607
608
609
	if len(s) > 0 {
		C.ggml_backend_tensor_set(t.(*Tensor).t, unsafe.Pointer(&s[0]), 0, C.ggml_nbytes(t.(*Tensor).t))
	}

610
	return t, nil
Michael Yang's avatar
Michael Yang committed
611
612
}

Michael Yang's avatar
Michael Yang committed
613
614
func (c *Context) Close() {
	if c != nil {
615
616
		C.ggml_free(c.ctx)
	}
Michael Yang's avatar
Michael Yang committed
617
618
619
}

type Tensor struct {
620
	b    *Backend
Michael Yang's avatar
Michael Yang committed
621
	t    *C.struct_ggml_tensor
622
	sync func()
Michael Yang's avatar
Michael Yang committed
623
624
625
626
627
628
629
630
631
632
}

func (t *Tensor) LogValue() slog.Value {
	return slog.GroupValue(
		slog.String("name", C.GoString(C.ggml_get_name(t.t))),
		slog.String("type", C.GoString(C.ggml_type_name(t.t._type))),
		slog.Any("shape", t.Shape()),
	)
}

633
634
func (t *Tensor) Dim(n int) int {
	return int(t.t.ne[n])
Michael Yang's avatar
Michael Yang committed
635
636
}

637
638
func (t *Tensor) Stride(n int) int {
	return int(t.t.nb[n])
Michael Yang's avatar
Michael Yang committed
639
640
}

641
642
func (t *Tensor) Shape() []int {
	shape := make([]int, C.ggml_n_dims(t.t))
Michael Yang's avatar
Michael Yang committed
643
644
645
646
647
648
649
	for i := range shape {
		shape[i] = t.Dim(i)
	}

	return shape
}

650
651
652
653
654
655
656
657
658
func (t *Tensor) Bytes() (data []byte) {
	if t.sync != nil {
		data = make([]byte, C.ggml_nbytes(t.t))

		t.sync()
		C.ggml_backend_tensor_get(t.t, unsafe.Pointer(&data[0]), 0, C.ggml_nbytes(t.t))
	}

	return
Michael Yang's avatar
Michael Yang committed
659
660
}

661
662
663
664
665
666
func (t *Tensor) Floats() (data []float32) {
	if t.sync != nil {
		data = make([]float32, C.ggml_nelements(t.t))

		t.sync()
		C.ggml_backend_tensor_get(t.t, unsafe.Pointer(&data[0]), 0, C.ggml_nbytes(t.t))
Michael Yang's avatar
Michael Yang committed
667
668
669
670
671
672
673
674
675
	}

	return
}

func (t *Tensor) DType() ml.DType {
	switch t.t._type {
	case C.GGML_TYPE_F32:
		return ml.DTypeF32
Jesse Gross's avatar
Jesse Gross committed
676
677
	case C.GGML_TYPE_F16:
		return ml.DTypeF16
Michael Yang's avatar
Michael Yang committed
678
679
680
681
682
683
684
685
686
	case C.GGML_TYPE_I32:
		return ml.DTypeI32
	default:
		return ml.DTypeOther
	}
}

func (t *Tensor) Add(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
	return &Tensor{
687
		b: t.b,
Michael Yang's avatar
Michael Yang committed
688
689
690
691
692
693
694
695
696
697
698
699
700
701
		t: C.ggml_add(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

func (t *Tensor) Stack(ctx ml.Context, dim int, s ...ml.Tensor) ml.Tensor {
	if len(s) > 0 {
		return t.Concat(ctx, s[0].Stack(ctx, dim, s[1:]...), dim)
	}

	return t
}

func (t *Tensor) Concat(ctx ml.Context, t2 ml.Tensor, dim int) ml.Tensor {
	return &Tensor{
702
		b: t.b,
Michael Yang's avatar
Michael Yang committed
703
704
705
706
707
708
		t: C.ggml_concat(ctx.(*Context).ctx, t.t, t2.(*Tensor).t, C.int(dim)),
	}
}

func (t *Tensor) Contiguous(ctx ml.Context) ml.Tensor {
	return &Tensor{
709
		b: t.b,
Michael Yang's avatar
Michael Yang committed
710
711
712
713
714
715
		t: C.ggml_cont(ctx.(*Context).ctx, t.t),
	}
}

func (t *Tensor) Mul(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
	return &Tensor{
716
		b: t.b,
Michael Yang's avatar
Michael Yang committed
717
718
719
720
721
722
		t: C.ggml_mul(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

func (t *Tensor) Mulmat(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
	return &Tensor{
723
		b: t.b,
Michael Yang's avatar
Michael Yang committed
724
725
726
727
		t: C.ggml_mul_mat(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

728
729
730
731
732
func (t *Tensor) MulmatFullPrec(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
	mul := C.ggml_mul_mat(ctx.(*Context).ctx, t.t, t2.(*Tensor).t)
	C.ggml_mul_mat_set_prec(mul, C.GGML_PREC_F32)

	return &Tensor{
733
		b: t.b,
734
735
736
737
		t: mul,
	}
}

Michael Yang's avatar
Michael Yang committed
738
func (t *Tensor) LayerNorm(ctx ml.Context, w, b ml.Tensor, eps float32) ml.Tensor {
739
	tt := (&Tensor{b: t.b, t: C.ggml_norm(ctx.(*Context).ctx, t.t, C.float(eps))}).Mul(ctx, w)
Michael Yang's avatar
Michael Yang committed
740
741
742
743
744
745
746
747
	if b != nil {
		tt = tt.Add(ctx, b)
	}

	return tt
}

func (t *Tensor) RMSNorm(ctx ml.Context, w ml.Tensor, eps float32) ml.Tensor {
748
	return (&Tensor{b: t.b, t: C.ggml_rms_norm(ctx.(*Context).ctx, t.t, C.float(eps))}).Mul(ctx, w)
Michael Yang's avatar
Michael Yang committed
749
750
}

751
func (t *Tensor) Pad(ctx ml.Context, shape ...int) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
752
753
754
755
756
	if len(shape) != 4 {
		panic("expected 4 dimensions")
	}

	return &Tensor{
757
		b: t.b,
Michael Yang's avatar
Michael Yang committed
758
759
760
761
762
763
764
765
766
767
		t: C.ggml_pad(ctx.(*Context).ctx, t.t, C.int(shape[0]), C.int(shape[1]), C.int(shape[2]), C.int(shape[3])),
	}
}

func (t *Tensor) Permute(ctx ml.Context, shape ...int) ml.Tensor {
	if len(shape) != 4 {
		panic("expected 4 dimensions")
	}

	return &Tensor{
768
		b: t.b,
Michael Yang's avatar
Michael Yang committed
769
770
771
772
773
774
		t: C.ggml_permute(ctx.(*Context).ctx, t.t, C.int(shape[0]), C.int(shape[1]), C.int(shape[2]), C.int(shape[3])),
	}
}

func (t *Tensor) Rows(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
	return &Tensor{
775
		b: t.b,
Michael Yang's avatar
Michael Yang committed
776
777
778
779
780
781
		t: C.ggml_get_rows(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

func (t *Tensor) Copy(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
	return &Tensor{
782
		b: t.b,
Michael Yang's avatar
Michael Yang committed
783
784
785
786
		t: C.ggml_cpy(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

787
func (t *Tensor) Reshape(ctx ml.Context, shape ...int) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
788
789
790
	switch len(shape) {
	case 1:
		return &Tensor{
791
			b: t.b,
Michael Yang's avatar
Michael Yang committed
792
793
794
795
			t: C.ggml_reshape_1d(ctx.(*Context).ctx, t.t, C.int64_t(shape[0])),
		}
	case 2:
		return &Tensor{
796
			b: t.b,
Michael Yang's avatar
Michael Yang committed
797
798
799
800
			t: C.ggml_reshape_2d(ctx.(*Context).ctx, t.t, C.int64_t(shape[0]), C.int64_t(shape[1])),
		}
	case 3:
		return &Tensor{
801
			b: t.b,
Michael Yang's avatar
Michael Yang committed
802
803
804
805
			t: C.ggml_reshape_3d(ctx.(*Context).ctx, t.t, C.int64_t(shape[0]), C.int64_t(shape[1]), C.int64_t(shape[2])),
		}
	case 4:
		return &Tensor{
806
			b: t.b,
Michael Yang's avatar
Michael Yang committed
807
808
809
810
811
812
813
814
815
			t: C.ggml_reshape_4d(ctx.(*Context).ctx, t.t, C.int64_t(shape[0]), C.int64_t(shape[1]), C.int64_t(shape[2]), C.int64_t(shape[3])),
		}
	default:
		panic("unsupported number of dimensions")
	}
}

func (t *Tensor) Scale(ctx ml.Context, s float64) ml.Tensor {
	return &Tensor{
816
		b: t.b,
Michael Yang's avatar
Michael Yang committed
817
818
819
820
821
822
		t: C.ggml_scale(ctx.(*Context).ctx, t.t, (C.float)(s)),
	}
}

func (t *Tensor) Softmax(ctx ml.Context) ml.Tensor {
	return &Tensor{
823
		b: t.b,
Michael Yang's avatar
Michael Yang committed
824
825
826
827
828
829
		t: C.ggml_soft_max(ctx.(*Context).ctx, t.t),
	}
}

func (t *Tensor) Tanh(ctx ml.Context) ml.Tensor {
	return &Tensor{
830
		b: t.b,
Michael Yang's avatar
Michael Yang committed
831
832
833
834
		t: C.ggml_tanh_inplace(ctx.(*Context).ctx, t.t),
	}
}

835
func (t *Tensor) Unpad(ctx ml.Context, shape ...int) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
836
837
838
839
840
	if len(shape) != 4 {
		panic("expected 4 dimensions")
	}

	return &Tensor{
841
		b: t.b,
Michael Yang's avatar
Michael Yang committed
842
843
844
845
846
847
848
849
		t: C.ggml_unpad(ctx.(*Context).ctx, t.t, C.int(shape[0]), C.int(shape[1]), C.int(shape[2]), C.int(shape[3])),
	}
}

func (t *Tensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
	switch len(shape) {
	case 1:
		return &Tensor{
850
			b: t.b,
Michael Yang's avatar
Michael Yang committed
851
852
853
854
			t: C.ggml_view_1d(ctx.(*Context).ctx, t.t, C.int64_t(shape[0]), C.size_t(offset)),
		}
	case 3:
		return &Tensor{
855
			b: t.b,
Michael Yang's avatar
Michael Yang committed
856
857
858
859
860
861
862
			t: C.ggml_view_2d(ctx.(*Context).ctx, t.t,
				C.int64_t(shape[0]), C.int64_t(shape[2]),
				C.size_t(shape[1]),
				C.size_t(offset)),
		}
	case 5:
		return &Tensor{
863
			b: t.b,
Michael Yang's avatar
Michael Yang committed
864
865
866
867
868
869
870
			t: C.ggml_view_3d(ctx.(*Context).ctx, t.t,
				C.int64_t(shape[0]), C.int64_t(shape[2]), C.int64_t(shape[4]),
				C.size_t(shape[1]), C.size_t(shape[3]),
				C.size_t(offset)),
		}
	case 7:
		return &Tensor{
871
			b: t.b,
Michael Yang's avatar
Michael Yang committed
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
			t: C.ggml_view_4d(ctx.(*Context).ctx, t.t,
				C.int64_t(shape[0]), C.int64_t(shape[2]), C.int64_t(shape[4]), C.int64_t(shape[6]),
				C.size_t(shape[1]), C.size_t(shape[3]), C.size_t(shape[5]),
				C.size_t(offset)),
		}
	default:
		panic("unsupported number of dimensions")
	}
}

const (
	ropeTypeNorm C.int = iota
)

func (t *Tensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, ropeDim uint32, ropeBase, ropeScale float32) ml.Tensor {
	if ropeFactors == nil {
888
		ropeFactors = &Tensor{b: t.b}
Michael Yang's avatar
Michael Yang committed
889
890
	}

Jesse Gross's avatar
Jesse Gross committed
891
892
893
894
895
	dequant := t.t
	if C.ggml_is_quantized(t.t._type) {
		dequant = C.ggml_cast(ctx.(*Context).ctx, t.t, C.GGML_TYPE_F32)
	}

Michael Yang's avatar
Michael Yang committed
896
	return &Tensor{
897
		b: t.b,
Michael Yang's avatar
Michael Yang committed
898
		t: C.ggml_rope_ext(
Jesse Gross's avatar
Jesse Gross committed
899
			ctx.(*Context).ctx, dequant, positionIDs.(*Tensor).t, ropeFactors.(*Tensor).t,
Michael Yang's avatar
Michael Yang committed
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
			C.int(ropeDim),
			131072,       // YaRN n_ctx_train
			ropeTypeNorm, // ROPE_TYPE_NORM
			C.float(ropeBase),
			C.float(ropeScale),
			0.,  // YaRN ext_factor
			1.,  // YaRN attn_factor
			32., // YaRN beta_fast
			1.,  // YaRN beta_slow
		),
	}
}

func (t *Tensor) GELU(ctx ml.Context) ml.Tensor {
	return &Tensor{
915
		b: t.b,
Michael Yang's avatar
Michael Yang committed
916
917
918
919
920
921
		t: C.ggml_gelu_inplace(ctx.(*Context).ctx, t.t),
	}
}

func (t *Tensor) SILU(ctx ml.Context) ml.Tensor {
	return &Tensor{
922
		b: t.b,
Michael Yang's avatar
Michael Yang committed
923
924
925
926
927
928
		t: C.ggml_silu_inplace(ctx.(*Context).ctx, t.t),
	}
}

func (t *Tensor) Conv2D(ctx ml.Context, t2 ml.Tensor, s0, s1, p0, p1, d0, d1 int) ml.Tensor {
	return &Tensor{
929
		b: t.b,
Michael Yang's avatar
Michael Yang committed
930
931
932
		t: C.ggml_conv_2d(ctx.(*Context).ctx, t.t, t2.(*Tensor).t, C.int(s0), C.int(s1), C.int(p0), C.int(p1), C.int(d0), C.int(d1)),
	}
}
933

934
935
936
937
938
939
func (t *Tensor) ScaledDotProductAttention(ctx ml.Context, key, value, mask ml.Tensor, scale float64) ml.Tensor {
	var kqMask *C.struct_ggml_tensor
	if mask != nil {
		kqMask = mask.(*Tensor).t
	}

940
941
942
	query := t.Permute(ctx, 0, 2, 1, 3)
	key = key.Permute(ctx, 0, 2, 1, 3)

943
944
	if t.b.flashAttention {
		value = value.Permute(ctx, 0, 2, 1, 3)
945

946
947
948
949
950
951
952
953
954
955
956
957
958
		kqv := C.ggml_flash_attn_ext(ctx.(*Context).ctx, query.(*Tensor).t, key.(*Tensor).t, value.(*Tensor).t, kqMask, C.float(scale), 0, 0)
		C.ggml_flash_attn_ext_set_prec(kqv, C.GGML_PREC_F32)
		return &Tensor{b: t.b, t: kqv}
	} else {
		kq := key.MulmatFullPrec(ctx, query)
		kq = &Tensor{
			b: t.b,
			t: C.ggml_soft_max_ext(ctx.(*Context).ctx, kq.(*Tensor).t, kqMask, C.float(scale), 0),
		}

		kqv := value.Mulmat(ctx, kq)
		return kqv.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
	}
959
}