ggml.go 29.5 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
	"context"
13
	"errors"
Michael Yang's avatar
Michael Yang committed
14
15
16
	"fmt"
	"io"
	"log/slog"
17
	"maps"
Michael Yang's avatar
Michael Yang committed
18
	"os"
19
	"runtime"
20
21
22
	"slices"
	"strconv"
	"strings"
23
	"sync/atomic"
24
	"unicode"
Michael Yang's avatar
Michael Yang committed
25
26
27
	"unsafe"

	"github.com/ollama/ollama/format"
28
29
	"github.com/ollama/ollama/fs"
	fsggml "github.com/ollama/ollama/fs/ggml"
Michael Yang's avatar
Michael Yang committed
30
	"github.com/ollama/ollama/ml"
31
	ggml "github.com/ollama/ollama/ml/backend/ggml/ggml/src"
Michael Yang's avatar
Michael Yang committed
32
33
34
	"golang.org/x/sync/errgroup"
)

Michael Yang's avatar
Michael Yang committed
35
36
37
38
39
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
40
	}
Michael Yang's avatar
Michael Yang committed
41
42

	return ds
43
}
Michael Yang's avatar
Michael Yang committed
44
45

type Backend struct {
46
47
48
49
50
51
	meta *fsggml.GGML

	sched         *C.struct_ggml_backend_sched
	schedBackends []*C.struct_ggml_backend
	schedBufts    []*C.struct_ggml_backend_buffer_type

52
	tensors map[string]*C.struct_ggml_tensor
Michael Yang's avatar
Michael Yang committed
53
54

	// input is the backend used for inputs
55
	input *C.struct_ggml_backend_buffer_type
Michael Yang's avatar
Michael Yang committed
56
57

	// layers is the backend used for repeating layers
58
	layers map[int]*C.struct_ggml_backend_buffer_type
59

60
	flashAttention bool
Michael Yang's avatar
Michael Yang committed
61
62
63

	// maxGraphNodes is the maximum allowed number of graph nodes in this scheduler
	maxGraphNodes int
Michael Yang's avatar
Michael Yang committed
64
65
}

66
func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend, error) {
67
	meta, n, err := fsggml.Decode(r, -1)
Michael Yang's avatar
Michael Yang committed
68
69
70
71
72
73
74
75
76
77
78
79
80
81
	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()),
	)

82
	type deviceBufferType struct {
83
84
85
86
87
		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
88
	for _, d := range devices() {
89
90
		switch C.ggml_backend_dev_type(d) {
		case C.GGML_BACKEND_DEVICE_TYPE_CPU:
91
92
93
94
			if len(cpus) == 0 {
				// only the first cpu device should be used
				cpus = append(cpus, d)
			}
95
96
		case C.GGML_BACKEND_DEVICE_TYPE_ACCEL:
			accels = append(accels, d)
Michael Yang's avatar
Michael Yang committed
97
		case C.GGML_BACKEND_DEVICE_TYPE_GPU:
98
			gpus = append(gpus, d)
Michael Yang's avatar
Michael Yang committed
99
100
101
		}
	}

Michael Yang's avatar
Michael Yang committed
102
	// create list of buffer types for the cpu
Michael Yang's avatar
Michael Yang committed
103
	cpuDeviceBufferType := deviceBufferType{d: C.ggml_backend_dev_by_type(C.GGML_BACKEND_DEVICE_TYPE_CPU)}
104
105
106
107
	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
108
			cpuDeviceBufferType.bts = append(cpuDeviceBufferType.bts, C.ggml_backend_dev_buffer_type(d))
Michael Yang's avatar
Michael Yang committed
109
		}
110
111
	}

Michael Yang's avatar
Michael Yang committed
112
	// create list of buffer types for each gpu
113
	var gpuDeviceBufferTypes []deviceBufferType
114
115
	for _, d := range gpus {
		bt := C.ggml_backend_dev_buffer_type(d)
116
		gpuDeviceBufferTypes = append(gpuDeviceBufferTypes, deviceBufferType{
117
			d:   d,
Michael Yang's avatar
Michael Yang committed
118
			bts: append([]*C.struct_ggml_backend_buffer_type{bt}, cpuDeviceBufferType.bts...),
119
		})
Michael Yang's avatar
Michael Yang committed
120
121
	}

Michael Yang's avatar
Michael Yang committed
122
123
124
125
126
	useDefaultSplit := true
	for _, s := range params.TensorSplit {
		if s != 0 {
			useDefaultSplit = false
			break
127
		}
Michael Yang's avatar
Michael Yang committed
128
	}
129

Michael Yang's avatar
Michael Yang committed
130
131
132
133
	// calculate splits
	splits := make([]float32, len(gpus))
	if useDefaultSplit {
		// default: split on free memory
134
135
136
137
138
		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
139
140
	} else {
		splits = params.TensorSplit
141
142
143
	}

	var sum float32
Michael Yang's avatar
Michael Yang committed
144
	// cumulative sum of all splits
145
146
147
148
149
	for i := range splits {
		sum += splits[i]
		splits[i] = sum
	}

Michael Yang's avatar
Michael Yang committed
150
	// normalize splits
151
	for i := range splits {
152
		splits[i] /= sum
153
154
	}

Michael Yang's avatar
Michael Yang committed
155
	// inputs always use cpu
Michael Yang's avatar
Michael Yang committed
156
	input := cpuDeviceBufferType
157

158
	blocks := int(meta.KV().BlockCount())
Michael Yang's avatar
Michael Yang committed
159
160
161
162

	// 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
163
	assignLayer := func(i int) deviceBufferType {
Michael Yang's avatar
Michael Yang committed
164
		if i < gpuRangeStart || i >= gpuRangeStop {
Michael Yang's avatar
Michael Yang committed
165
			return cpuDeviceBufferType
166
		}
167

Michael Yang's avatar
Michael Yang committed
168
		index := slices.IndexFunc(splits, func(f float32) bool { return float32(i-gpuRangeStart)/float32(gpuRangeStop-gpuRangeStart) < f })
169
		if index < 0 || index >= len(gpuDeviceBufferTypes) {
Michael Yang's avatar
Michael Yang committed
170
			return cpuDeviceBufferType
171
172
173
		}

		return gpuDeviceBufferTypes[index]
174
175
	}

Michael Yang's avatar
Michael Yang committed
176
	// repeating layers are assigned based on their index in reverse order, e.g. i / (block_count + 1)
177
	layers := make([]deviceBufferType, blocks)
178
	for i := range layers {
179
		layers[i] = assignLayer(i)
180
181
	}

Michael Yang's avatar
Michael Yang committed
182
	// outputs are assigned iff allowed by splits and configured number of gpu layers
183
	output := assignLayer(blocks)
184
185
186

	maxTensors := len(meta.Tensors().Items())
	maxTensors += 1
Michael Yang's avatar
Michael Yang committed
187
	// each layer has at most 2 extra tensors for rope operations
188
189
	maxTensors += blocks * 2

190
	type tensor struct {
191
		source *fsggml.Tensor
192
193
194
		target string
	}

Michael Yang's avatar
Michael Yang committed
195
	// some tensors are mapped to different names so keep a list
196
197
	targets := make(map[string][]string)

Michael Yang's avatar
Michael Yang committed
198
	// contexts are shared by tensors of the same buffer type
199
	ctxs := make(map[*C.struct_ggml_backend_buffer_type]*C.struct_ggml_context)
200
	createTensor := func(t tensor, bts []*C.struct_ggml_backend_buffer_type) *C.struct_ggml_tensor {
201
202
203
204
205
206
207
		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
208

209
210
211
212
213
214
215
216
			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
217
			defer C.free(unsafe.Pointer(cname))
218
219
220
221
			if tt := C.ggml_get_tensor(ctxs[bt], cname); tt != nil {
				return tt
			}

222
			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
223
224
			C.ggml_set_name(tt, cname)

225
			slog.Debug("created tensor", "name", name, "shape", t.source.Shape, "dtype", t.source.Kind, "buffer_type", C.GoString(C.ggml_backend_buft_name(bt)))
226
227
228
229
230
			//nolint:staticcheck // TODO: check if buffer type supports this tensor
			return tt
		}

		return nil
Michael Yang's avatar
Michael Yang committed
231
232
	}

233
	contains := func(s string, parts ...string) bool {
234
235
236
237
238
239
240
241
		split := strings.Split(s, ".")
		for _, part := range parts {
			if slices.Contains(split, part) {
				return true
			}
		}

		return false
Michael Yang's avatar
Michael Yang committed
242
243
	}

244
245
	for _, t := range meta.Tensors().Items() {
		switch {
246
		case contains(t.Name, "position_embd", "token_embd", "token_norm_embd", "token_types"):
247
			createTensor(tensor{source: t}, input.bts)
Michael Yang's avatar
Michael Yang committed
248
249
250
			if _, ok := meta.Tensors().GroupLayers()["output"]; !ok && t.Name == "token_embd.weight" {
				createTensor(tensor{source: t, target: "output.weight"}, output.bts)
			}
251
		case contains(t.Name, "cls", "output", "output_norm"):
252
			createTensor(tensor{source: t}, output.bts)
253
		case strings.HasPrefix(t.Name, "v.") || strings.HasPrefix(t.Name, "mm."):
Michael Yang's avatar
Michael Yang committed
254
			// TODO: assign vision tensors to the gpu if possible
Michael Yang's avatar
Michael Yang committed
255
			createTensor(tensor{source: t}, output.bts)
Michael Yang's avatar
Michael Yang committed
256
257
258
259
260
261
262
263
		case contains(t.Name, "rope_freqs", "rope_factors_long", "rope_factors_short"):
			// these tensors should be repeated per layer
			for i, layer := range layers {
				createTensor(tensor{
					source: t,
					target: "blk." + strconv.Itoa(i) + "." + t.Name,
				}, layer.bts)
			}
264
		default:
Michael Yang's avatar
Michael Yang committed
265
266
267
268
			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
269
				}
Michael Yang's avatar
Michael Yang committed
270
			}
271

Michael Yang's avatar
Michael Yang committed
272
273
			if layerIndex >= 0 {
				createTensor(tensor{source: t}, layers[layerIndex].bts)
274
			} else {
Michael Yang's avatar
Michael Yang committed
275
276
				// load all other tensors on the cpu
				createTensor(tensor{source: t}, input.bts)
277
278
279
			}
		}
	}
Michael Yang's avatar
Michael Yang committed
280

Michael Yang's avatar
Michael Yang committed
281
282
	// allocate buffers for each context
	bbs := make(map[*C.struct_ggml_context]*C.struct_ggml_backend_buffer, len(ctxs))
283
284
285
286
287
288
	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)
289
290
291
292
		if b == nil {
			return nil, fmt.Errorf("unable to allocate memory from device %v for model weights", C.GoString(C.ggml_backend_buft_name(bt)))
		}

293
		C.ggml_backend_buffer_set_usage(b, C.GGML_BACKEND_BUFFER_USAGE_WEIGHTS)
Michael Yang's avatar
Michael Yang committed
294
		bbs[c] = b
295
296
297
	}

	for bs := range maps.Values(bbs) {
Michael Yang's avatar
Michael Yang committed
298
		slog.Info("model weights", "buffer", C.GoString(C.ggml_backend_buffer_name(bs)), "size", format.HumanBytes2(uint64(C.ggml_backend_buffer_get_size(bs))))
299
300
	}

Michael Yang's avatar
Michael Yang committed
301
	// map tensor names to tensors for easy lookup later
302
303
304
305
306
307
308
	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
		}
	}

309
310
311
312
313
	var doneBytes atomic.Uint64
	totalBytes := uint64(n) - meta.Tensors().Offset

	g, ctx := errgroup.WithContext(ctx)
	g.SetLimit(runtime.GOMAXPROCS(0))
314
	for _, t := range meta.Tensors().Items() {
315
		t := t
316
317
318
319
		g.Go(func() error {
			tts := make([]*C.struct_ggml_tensor, max(1, len(targets[t.Name])))
			for i := range tts {
				target := targets[t.Name][i]
320
321
322
				if target == "" {
					target = t.Name
				}
323

324
325
326
327
				tt, ok := tensors[target]
				if !ok {
					return fmt.Errorf("unassigned tensor: %s", t.Name)
				}
Michael Yang's avatar
Michael Yang committed
328

329
330
331
				tts[i] = tt
			}

332
333
334
335
			// Create a new FD for each goroutine so that each FD is read sequentially, rather than
			// seeking around within an FD shared between all goroutines.
			file, err := os.Open(r.Name())
			if err != nil {
Jesse Gross's avatar
Jesse Gross committed
336
				slog.Warn("file open error", "file", r.Name(), "error", err)
337
338
339
340
				return err
			}
			defer file.Close()
			sr := io.NewSectionReader(file, int64(meta.Tensors().Offset+t.Offset), int64(t.Size()))
341
342
343
344
			bts := make([]byte, 128*format.KibiByte)

			var s uint64
			for s < t.Size() {
345
346
347
348
349
				// Stop if either the parent context has been canceled or if any of the other tensors returned an error
				if err := ctx.Err(); err != nil {
					return err
				}

350
351
				n, err := io.ReadFull(sr, bts[:min(len(bts), int(t.Size()-s))])
				if err != nil {
Jesse Gross's avatar
Jesse Gross committed
352
					slog.Warn("file read error", "file", r.Name(), "error", err)
353
					return err
354
				}
Michael Yang's avatar
Michael Yang committed
355

356
357
				for _, tt := range tts {
					C.ggml_backend_tensor_set(tt, unsafe.Pointer(&bts[0]), C.size_t(s), C.size_t(n))
358
				}
Michael Yang's avatar
Michael Yang committed
359

360
361
362
363
364
365
366
367
368
369
				s += uint64(n)

				if params.Progress != nil {
					done := doneBytes.Add(uint64(n))
					params.Progress(float32(done) / float32(totalBytes))
				}
			}

			return nil
		})
Michael Yang's avatar
Michael Yang committed
370
371
	}

372
	if err := g.Wait(); err != nil {
Michael Yang's avatar
Michael Yang committed
373
374
375
		return nil, err
	}

376
377
	// 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
378
379
380
381

	// 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
382
383
384
385
	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)

386
387
388
		deviceBufferTypes[d] = bt

		schedBackends = append(schedBackends, b)
Michael Yang's avatar
Michael Yang committed
389
		schedBufts = append(schedBufts, bt)
390

391
		if C.ggml_backend_is_cpu(b) {
Michael Yang's avatar
Michael Yang committed
392
			// set number of threads for cpu backend
Michael Yang's avatar
Michael Yang committed
393
			C.ggml_backend_cpu_set_n_threads(b, C.int(Threads(params.NumThreads)))
394
		}
395
396
	}

Michael Yang's avatar
Michael Yang committed
397
	maxGraphNodes := max(8192, len(meta.Tensors().Items())*5)
Michael Yang's avatar
Michael Yang committed
398
	return &Backend{
399
		flashAttention: params.FlashAttention,
400
401
		meta:           meta,
		tensors:        tensors,
402
		sched: C.ggml_backend_sched_new(
Michael Yang's avatar
Michael Yang committed
403
404
405
406
			(*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),
407
			C._Bool(len(gpus) > 1 && slices.Contains(gpus, output.d)),
408
		),
409
410
411
		schedBackends: schedBackends,
		schedBufts:    schedBufts,
		input:         deviceBufferTypes[input.d],
412
413
		layers: func() map[int]*C.struct_ggml_backend_buffer_type {
			m := make(map[int]*C.struct_ggml_backend_buffer_type)
414
			for i, layer := range layers {
415
				m[i] = deviceBufferTypes[layer.d]
416
417
418
			}
			return m
		}(),
Michael Yang's avatar
Michael Yang committed
419
		maxGraphNodes: maxGraphNodes,
Michael Yang's avatar
Michael Yang committed
420
421
422
423
424
425
426
	}, nil
}

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

427
func (b *Backend) Config() fs.Config {
Michael Yang's avatar
Michael Yang committed
428
429
430
431
	return b.meta.KV()
}

func (b *Backend) Get(name string) ml.Tensor {
432
433
	if t, ok := b.tensors[name]; ok {
		return &Tensor{b: b, t: t}
Michael Yang's avatar
Michael Yang committed
434
435
436
437
438
439
	}

	return nil
}

func (b *Backend) NewContext() ml.Context {
Michael Yang's avatar
Michael Yang committed
440
	return b.NewContextSize(b.maxGraphNodes)
441
442
443
}

func (b *Backend) NewContextSize(n int) ml.Context {
Jesse Gross's avatar
Jesse Gross committed
444
445
446
447
	if n > b.maxGraphNodes {
		panic(fmt.Errorf("requested number of graph nodes (%v) for new context exceeds maximum (%v)", n, b.maxGraphNodes))
	}

448
449
	var allocatedBuffers []*C.struct_ggml_backend_buffer

Michael Yang's avatar
Michael Yang committed
450
	return &Context{
451
452
		b:             b,
		maxGraphNodes: n,
453
		ctx: C.ggml_init(C.struct_ggml_init_params{
454
			mem_size: C.size_t(n)*C.ggml_tensor_overhead() + C.ggml_graph_overhead_custom(C.size_t(n), false),
455
456
			no_alloc: true,
		}),
457
		allocatedBuffers: &allocatedBuffers,
Michael Yang's avatar
Michael Yang committed
458
459
460
	}
}

461
func (b *Backend) CacheConfig() ml.CacheConfig {
462
463
464
465
466
	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}
	}
467
468
}

Michael Yang's avatar
Michael Yang committed
469
type Context struct {
470
	b *Backend
Michael Yang's avatar
Michael Yang committed
471

472
	ctx   *C.struct_ggml_context
Michael Yang's avatar
Michael Yang committed
473
	graph *C.struct_ggml_cgraph
474

475
476
	// buft is the buffer type used for new tensors
	buft *C.struct_ggml_backend_buffer_type
477

478
479
480
481
	// allocatedBuffers are buffers for tensors that we have allocated in this context
	// so that we can free them when we close the context
	allocatedBuffers *[]*C.struct_ggml_backend_buffer

Michael Yang's avatar
Michael Yang committed
482
	// maxGraphNodes is the maximum allowed number of graph nodes in this context
483
	maxGraphNodes int
Michael Yang's avatar
Michael Yang committed
484
485
}

486
func (c *Context) Input() ml.Context {
Michael Yang's avatar
Michael Yang committed
487
	if c.b.input != nil {
488
		return &Context{
489
490
491
492
493
			b:                c.b,
			ctx:              c.ctx,
			buft:             c.b.input,
			allocatedBuffers: c.allocatedBuffers,
			maxGraphNodes:    c.maxGraphNodes,
494
495
496
		}
	}

497
	return c
498
499
}

500
func (c *Context) Layer(i int) ml.Context {
501
	if buft, ok := c.b.layers[i]; ok {
502
		return &Context{
503
504
505
506
507
			b:                c.b,
			ctx:              c.ctx,
			buft:             buft,
			allocatedBuffers: c.allocatedBuffers,
			maxGraphNodes:    c.maxGraphNodes,
508
509
510
		}
	}

511
	return c
512
513
}

514
func (c *Context) Forward(tensors ...ml.Tensor) ml.Context {
Michael Yang's avatar
Michael Yang committed
515
	if c.graph == nil {
516
		c.graph = C.ggml_new_graph_custom(c.ctx, C.size_t(c.maxGraphNodes), false)
Michael Yang's avatar
Michael Yang committed
517
518
	}

519
520
521
522
523
	for _, tensor := range tensors {
		C.ggml_build_forward_expand(c.graph, tensor.(*Tensor).t)
	}

	return c
Michael Yang's avatar
Michael Yang committed
524
525
}

526
func (c *Context) Compute(tensors ...ml.Tensor) {
527
	C.ggml_backend_sched_graph_compute_async(c.b.sched, c.graph)
Michael Yang's avatar
Michael Yang committed
528
	C.ggml_backend_sched_reset(c.b.sched)
Michael Yang's avatar
Michael Yang committed
529

530
531
532
	needSync := true
	sync := func() {
		if needSync {
533
			C.ggml_backend_sched_synchronize(c.b.sched)
534
535
536
			needSync = false
		}
	}
Michael Yang's avatar
Michael Yang committed
537

538
539
540
	for _, t := range tensors {
		if C.ggml_nbytes(t.(*Tensor).t) > 0 {
			t.(*Tensor).sync = sync
541
542
		}
	}
Michael Yang's avatar
Michael Yang committed
543
544
}

545
func (c *Context) Reserve() error {
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
	if !C.ggml_backend_sched_reserve(c.b.sched, c.graph) {
		C.ggml_backend_sched_reset(c.b.sched)
		return errors.New("failed to reserve graph")
	}

	slog.Debug("compute graph", "nodes", C.ggml_graph_n_nodes(c.graph), "splits", C.ggml_backend_sched_get_n_splits(c.b.sched))
	for i := range c.b.schedBackends {
		size := C.ggml_backend_sched_get_buffer_size(c.b.sched, c.b.schedBackends[i])
		slog.Info("compute graph", "backend", C.GoString(C.ggml_backend_name(c.b.schedBackends[i])), "buffer_type", C.GoString(C.ggml_backend_buft_name(c.b.schedBufts[i])),
			"size", format.HumanBytes2(uint64(size)))
	}

	C.ggml_backend_sched_reset(c.b.sched)

	return nil
}

563
func (c *Context) MaxGraphNodes() int {
564
	return c.maxGraphNodes
Jesse Gross's avatar
Jesse Gross committed
565
566
}

567
568
569
func shapeToGGML(shape []int) *C.int64_t {
	sh := make([]C.int64_t, len(shape))
	for i, s := range shape {
570
		sh[i] = C.int64_t(s)
571
572
573
574
575
	}

	return &sh[0]
}

576
577
578
579
func pad(length, pad C.size_t) C.size_t {
	return ((length + pad - 1) / pad) * pad
}

580
func (c *Context) newTensor(dtype ml.DType, shape []int) (ml.Tensor, error) {
581
	if c.buft == nil {
582
		panic("set Input or Layer before creating tensors")
583
584
	}

Michael Yang's avatar
Michael Yang committed
585
586
587
588
589
590
	var cdtype uint32
	switch dtype {
	case ml.DTypeF32:
		cdtype = C.GGML_TYPE_F32
	case ml.DTypeF16:
		cdtype = C.GGML_TYPE_F16
591
592
593
594
	case ml.DTypeQ80:
		cdtype = C.GGML_TYPE_Q8_0
	case ml.DTypeQ40:
		cdtype = C.GGML_TYPE_Q4_0
Michael Yang's avatar
Michael Yang committed
595
596
597
598
599
600
	case ml.DTypeI32:
		cdtype = C.GGML_TYPE_I32
	default:
		panic("unsupported dtype")
	}

Jesse Gross's avatar
Jesse Gross committed
601
	if len(shape) < 1 || shape[0] == 0 {
Michael Yang's avatar
Michael Yang committed
602
		var shape C.int64_t = 0
603
		return &Tensor{b: c.b, t: C.ggml_new_tensor(c.ctx, cdtype, 1, &shape)}, nil
Michael Yang's avatar
Michael Yang committed
604
	} else if len(shape) > 4 {
Michael Yang's avatar
Michael Yang committed
605
606
607
608
609
610
611
612
613
		panic("unsupported number of dimensions")
	}

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

Michael Yang's avatar
Michael Yang committed
614
	t := C.ggml_new_tensor(c.ctx, cdtype, C.int(len(shape)), shapeToGGML(shape))
615
616
	size := pad(C.ggml_backend_buft_get_alloc_size(c.buft, t), C.ggml_backend_buft_get_alignment(c.buft))
	b := C.ggml_backend_buft_alloc_buffer(c.buft, size)
617
618
619
	if b == nil {
		return nil, fmt.Errorf("unable to allocate %v from device %v for new tensor", format.HumanBytes2(uint64(size)), C.GoString(C.ggml_backend_buft_name(c.buft)))
	}
620
	*c.allocatedBuffers = append(*c.allocatedBuffers, b)
621

Michael Yang's avatar
Michael Yang committed
622
	C.ggml_backend_tensor_alloc(b, t, C.ggml_backend_buffer_get_base(b))
623
	return &Tensor{b: c.b, t: t}, nil
624
625
}

626
func (c *Context) Empty(dtype ml.DType, shape ...int) ml.Tensor {
627
628
629
630
631
632
	t, err := c.newTensor(dtype, shape)
	if err != nil {
		panic(err)
	}

	return t
633
634
}

635
func (c *Context) Zeros(dtype ml.DType, shape ...int) ml.Tensor {
636
637
638
639
640
	t, err := c.newTensor(dtype, shape)
	if err != nil {
		panic(err)
	}

641
642
	C.ggml_set_zero(t.(*Tensor).t)
	return t
Michael Yang's avatar
Michael Yang committed
643
644
}

645
func checkShape[S ~[]E, E any](s S, shape ...int) error {
Michael Yang's avatar
Michael Yang committed
646
	n := len(s)
Jesse Gross's avatar
Jesse Gross committed
647
648
649
650
651

	if n == 0 {
		return nil
	}

Michael Yang's avatar
Michael Yang committed
652
653
654
655
656
	for _, v := range shape {
		n /= v
	}

	if n != 1 {
657
		return fmt.Errorf("invalid shape: %v", shape)
Michael Yang's avatar
Michael Yang committed
658
659
	}

660
	return nil
Michael Yang's avatar
Michael Yang committed
661
662
}

663
func (c *Context) FromFloatSlice(s []float32, shape ...int) (ml.Tensor, error) {
Jesse Gross's avatar
Jesse Gross committed
664
	if err := checkShape(s, shape...); err != nil {
665
666
667
		return nil, err
	}

668
669
670
671
672
	t, err := c.newTensor(ml.DTypeF32, shape)
	if err != nil {
		return nil, err
	}

Jesse Gross's avatar
Jesse Gross committed
673
674
675
676
	if len(s) > 0 {
		C.ggml_backend_tensor_set(t.(*Tensor).t, unsafe.Pointer(&s[0]), 0, C.ggml_nbytes(t.(*Tensor).t))
	}

677
	return t, nil
Michael Yang's avatar
Michael Yang committed
678
679
}

680
func (c *Context) FromIntSlice(s []int32, shape ...int) (ml.Tensor, error) {
Jesse Gross's avatar
Jesse Gross committed
681
	if err := checkShape(s, shape...); err != nil {
682
683
684
		return nil, err
	}

685
686
687
688
689
	t, err := c.newTensor(ml.DTypeI32, shape)
	if err != nil {
		return nil, err
	}

Jesse Gross's avatar
Jesse Gross committed
690
691
692
693
	if len(s) > 0 {
		C.ggml_backend_tensor_set(t.(*Tensor).t, unsafe.Pointer(&s[0]), 0, C.ggml_nbytes(t.(*Tensor).t))
	}

694
	return t, nil
Michael Yang's avatar
Michael Yang committed
695
696
}

Michael Yang's avatar
arange  
Michael Yang committed
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
func (c Context) Arange(start, stop, step float32, dtype ml.DType) ml.Tensor {
	switch dtype {
	case ml.DTypeF32:
		// ggml_arange creates a float32 tensor
		return &Tensor{
			b: c.b,
			t: C.ggml_arange(c.ctx, C.float(start), C.float(stop), C.float(step)),
		}
	case ml.DTypeI32:
		// ggml_cast does not support float32 to int32 conversion
		arange := make([]int32, 0, int((stop-start)/step))
		for i := start; i < stop; i += step {
			arange = append(arange, int32(i))
		}

		t, err := c.Input().FromIntSlice(arange, len(arange))
		if err != nil {
			panic(err)
		}

		return t
	default:
		panic("unsupported dtype for arange")
	}
}

Michael Yang's avatar
Michael Yang committed
723
724
func (c *Context) Close() {
	if c != nil {
725
726
727
728
729
		for _, b := range *c.allocatedBuffers {
			C.ggml_backend_buffer_free(b)
		}
		*c.allocatedBuffers = nil

730
731
		C.ggml_free(c.ctx)
	}
Michael Yang's avatar
Michael Yang committed
732
733
734
}

type Tensor struct {
735
	b    *Backend
Michael Yang's avatar
Michael Yang committed
736
	t    *C.struct_ggml_tensor
737
	sync func()
Michael Yang's avatar
Michael Yang committed
738
739
740
741
742
743
744
745
746
747
}

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()),
	)
}

748
749
func (t *Tensor) Dim(n int) int {
	return int(t.t.ne[n])
Michael Yang's avatar
Michael Yang committed
750
751
}

752
753
func (t *Tensor) Stride(n int) int {
	return int(t.t.nb[n])
Michael Yang's avatar
Michael Yang committed
754
755
}

756
757
func (t *Tensor) Shape() []int {
	shape := make([]int, C.ggml_n_dims(t.t))
Michael Yang's avatar
Michael Yang committed
758
759
760
761
762
763
764
	for i := range shape {
		shape[i] = t.Dim(i)
	}

	return shape
}

765
766
767
768
769
770
771
772
773
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
774
775
}

776
777
778
779
780
781
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
782
783
784
785
786
787
788
789
790
	}

	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
791
792
	case C.GGML_TYPE_F16:
		return ml.DTypeF16
793
794
795
796
	case C.GGML_TYPE_Q8_0:
		return ml.DTypeQ80
	case C.GGML_TYPE_Q4_0:
		return ml.DTypeQ40
Michael Yang's avatar
Michael Yang committed
797
798
799
800
801
802
803
	case C.GGML_TYPE_I32:
		return ml.DTypeI32
	default:
		return ml.DTypeOther
	}
}

804
805
806
807
808
809
810
func (t *Tensor) Neg(ctx ml.Context) ml.Tensor {
	return &Tensor{
		b: t.b,
		t: C.ggml_neg(ctx.(*Context).ctx, t.t),
	}
}

Michael Yang's avatar
Michael Yang committed
811
812
func (t *Tensor) Add(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
	return &Tensor{
813
		b: t.b,
Michael Yang's avatar
Michael Yang committed
814
815
816
817
		t: C.ggml_add(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
func (t *Tensor) Repeat(ctx ml.Context, dim, n int) ml.Tensor {
	if dim < 0 || dim >= C.GGML_MAX_DIMS {
		panic("invalid dimension")
	}

	shape := make([]C.int64_t, C.GGML_MAX_DIMS)
	for i := range C.GGML_MAX_DIMS {
		if i == dim {
			shape[i] = C.int64_t(t.Dim(i) * n)
		} else {
			shape[i] = C.int64_t(t.Dim(i))
		}
	}

	tmpl := C.ggml_new_tensor(ctx.(*Context).ctx, t.t._type, C.int(len(shape)), unsafe.SliceData(shape))
	return &Tensor{
		b: t.b,
		t: C.ggml_repeat(ctx.(*Context).ctx, t.t, tmpl),
	}
}

Michael Yang's avatar
Michael Yang committed
839
840
841
842
843
844
845
846
847
848
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{
849
		b: t.b,
Michael Yang's avatar
Michael Yang committed
850
851
852
853
854
855
		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{
856
		b: t.b,
Michael Yang's avatar
Michael Yang committed
857
858
859
860
861
862
		t: C.ggml_cont(ctx.(*Context).ctx, t.t),
	}
}

func (t *Tensor) Mul(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
	return &Tensor{
863
		b: t.b,
Michael Yang's avatar
Michael Yang committed
864
865
866
867
868
869
		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{
870
		b: t.b,
Michael Yang's avatar
Michael Yang committed
871
872
873
874
		t: C.ggml_mul_mat(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

875
876
877
878
879
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{
880
		b: t.b,
881
882
883
884
		t: mul,
	}
}

Michael Yang's avatar
llama4  
Michael Yang committed
885
886
887
888
889
890
891
func (t *Tensor) MulmatID(ctx ml.Context, t2, ids ml.Tensor) ml.Tensor {
	return &Tensor{
		b: t.b,
		t: C.ggml_mul_mat_id(ctx.(*Context).ctx, t.t, t2.(*Tensor).t, ids.(*Tensor).t),
	}
}

Michael Yang's avatar
Michael Yang committed
892
func (t *Tensor) LayerNorm(ctx ml.Context, w, b ml.Tensor, eps float32) ml.Tensor {
Michael Yang's avatar
llama4  
Michael Yang committed
893
894
895
896
897
898
	tt := C.ggml_norm(ctx.(*Context).ctx, t.t, C.float(eps))
	if w != nil {
		tt = C.ggml_mul(ctx.(*Context).ctx, tt, w.(*Tensor).t)
		if b != nil {
			tt = C.ggml_add(ctx.(*Context).ctx, tt, b.(*Tensor).t)
		}
Michael Yang's avatar
Michael Yang committed
899
900
	}

Michael Yang's avatar
llama4  
Michael Yang committed
901
	return &Tensor{b: t.b, t: tt}
Michael Yang's avatar
Michael Yang committed
902
903
904
}

func (t *Tensor) RMSNorm(ctx ml.Context, w ml.Tensor, eps float32) ml.Tensor {
Michael Yang's avatar
llama4  
Michael Yang committed
905
906
907
908
909
910
	tt := C.ggml_rms_norm(ctx.(*Context).ctx, t.t, C.float(eps))
	if w != nil {
		tt = C.ggml_mul(ctx.(*Context).ctx, tt, w.(*Tensor).t)
	}

	return &Tensor{b: t.b, t: tt}
Michael Yang's avatar
Michael Yang committed
911
912
}

913
func (t *Tensor) Pad(ctx ml.Context, shape ...int) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
914
915
916
917
918
	if len(shape) != 4 {
		panic("expected 4 dimensions")
	}

	return &Tensor{
919
		b: t.b,
Michael Yang's avatar
Michael Yang committed
920
921
922
923
924
925
926
927
928
929
		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{
930
		b: t.b,
Michael Yang's avatar
Michael Yang committed
931
932
933
934
935
936
		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{
937
		b: t.b,
Michael Yang's avatar
Michael Yang committed
938
939
940
941
942
943
		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{
944
		b: t.b,
Michael Yang's avatar
Michael Yang committed
945
946
947
948
		t: C.ggml_cpy(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

949
func (t *Tensor) Reshape(ctx ml.Context, shape ...int) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
950
951
952
	switch len(shape) {
	case 1:
		return &Tensor{
953
			b: t.b,
Michael Yang's avatar
Michael Yang committed
954
955
956
957
			t: C.ggml_reshape_1d(ctx.(*Context).ctx, t.t, C.int64_t(shape[0])),
		}
	case 2:
		return &Tensor{
958
			b: t.b,
Michael Yang's avatar
Michael Yang committed
959
960
961
962
			t: C.ggml_reshape_2d(ctx.(*Context).ctx, t.t, C.int64_t(shape[0]), C.int64_t(shape[1])),
		}
	case 3:
		return &Tensor{
963
			b: t.b,
Michael Yang's avatar
Michael Yang committed
964
965
966
967
			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{
968
			b: t.b,
Michael Yang's avatar
Michael Yang committed
969
970
971
972
973
974
975
976
977
			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{
978
		b: t.b,
Michael Yang's avatar
Michael Yang committed
979
980
981
982
983
984
		t: C.ggml_scale(ctx.(*Context).ctx, t.t, (C.float)(s)),
	}
}

func (t *Tensor) Softmax(ctx ml.Context) ml.Tensor {
	return &Tensor{
985
		b: t.b,
Michael Yang's avatar
Michael Yang committed
986
987
988
989
		t: C.ggml_soft_max(ctx.(*Context).ctx, t.t),
	}
}

990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
func (t *Tensor) Sin(ctx ml.Context) ml.Tensor {
	return &Tensor{
		b: t.b,
		t: C.ggml_sin(ctx.(*Context).ctx, t.t),
	}
}

func (t *Tensor) Cos(ctx ml.Context) ml.Tensor {
	return &Tensor{
		b: t.b,
		t: C.ggml_cos(ctx.(*Context).ctx, t.t),
	}
}

Michael Yang's avatar
Michael Yang committed
1004
1005
func (t *Tensor) Tanh(ctx ml.Context) ml.Tensor {
	return &Tensor{
1006
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1007
1008
1009
1010
		t: C.ggml_tanh_inplace(ctx.(*Context).ctx, t.t),
	}
}

Michael Yang's avatar
llama4  
Michael Yang committed
1011
1012
1013
1014
1015
1016
1017
func (t *Tensor) Sigmoid(ctx ml.Context) ml.Tensor {
	return &Tensor{
		b: t.b,
		t: C.ggml_sigmoid_inplace(ctx.(*Context).ctx, t.t),
	}
}

1018
func (t *Tensor) Unpad(ctx ml.Context, shape ...int) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
1019
1020
1021
1022
1023
	if len(shape) != 4 {
		panic("expected 4 dimensions")
	}

	return &Tensor{
1024
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1025
1026
1027
1028
1029
1030
1031
1032
		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{
1033
			b: t.b,
Michael Yang's avatar
Michael Yang committed
1034
1035
1036
1037
			t: C.ggml_view_1d(ctx.(*Context).ctx, t.t, C.int64_t(shape[0]), C.size_t(offset)),
		}
	case 3:
		return &Tensor{
1038
			b: t.b,
Michael Yang's avatar
Michael Yang committed
1039
1040
1041
1042
1043
1044
1045
			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{
1046
			b: t.b,
Michael Yang's avatar
Michael Yang committed
1047
1048
1049
1050
1051
1052
1053
			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{
1054
			b: t.b,
Michael Yang's avatar
Michael Yang committed
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
			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 (
Patrick Devine's avatar
Patrick Devine committed
1066
1067
1068
1069
	ropeTypeNorm   C.int = 0
	ropeTypeNeox   C.int = 2
	ropeTypeMrope  C.int = 8
	ropeTypeVision C.int = 24
Michael Yang's avatar
Michael Yang committed
1070
1071
)

Patrick Devine's avatar
Patrick Devine committed
1072
func (t *Tensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, ropeDim, ropeType uint32, ropeBase, ropeScale float32) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
1073
	if ropeFactors == nil {
1074
		ropeFactors = &Tensor{b: t.b}
Michael Yang's avatar
Michael Yang committed
1075
1076
	}

Jesse Gross's avatar
Jesse Gross committed
1077
1078
1079
1080
1081
	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
1082
	return &Tensor{
1083
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1084
		t: C.ggml_rope_ext(
Jesse Gross's avatar
Jesse Gross committed
1085
			ctx.(*Context).ctx, dequant, positionIDs.(*Tensor).t, ropeFactors.(*Tensor).t,
Michael Yang's avatar
Michael Yang committed
1086
			C.int(ropeDim),
Patrick Devine's avatar
Patrick Devine committed
1087
1088
			C.int(ropeType),
			131072, // YaRN n_ctx_train
Michael Yang's avatar
Michael Yang committed
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
			C.float(ropeBase),
			C.float(ropeScale),
			0.,  // YaRN ext_factor
			1.,  // YaRN attn_factor
			32., // YaRN beta_fast
			1.,  // YaRN beta_slow
		),
	}
}

1099
1100
1101
1102
1103
1104
1105
func (t *Tensor) IM2Col(ctx ml.Context, t2 ml.Tensor, s0, s1, p0, p1, d0, d1 int) ml.Tensor {
	return &Tensor{
		b: t.b,
		t: C.ggml_im2col(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), true, C.GGML_TYPE_F32),
	}
}

Michael Yang's avatar
Michael Yang committed
1106
1107
func (t *Tensor) GELU(ctx ml.Context) ml.Tensor {
	return &Tensor{
1108
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1109
1110
1111
1112
1113
1114
		t: C.ggml_gelu_inplace(ctx.(*Context).ctx, t.t),
	}
}

func (t *Tensor) SILU(ctx ml.Context) ml.Tensor {
	return &Tensor{
1115
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1116
1117
1118
1119
1120
1121
		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{
1122
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1123
1124
1125
		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)),
	}
}
1126

Michael Yang's avatar
Michael Yang committed
1127
func (t *Tensor) AvgPool2D(ctx ml.Context, k, s int, p float32) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
1128
1129
	return &Tensor{
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1130
		t: C.ggml_pool_2d(ctx.(*Context).ctx, t.t, C.GGML_OP_POOL_AVG, C.int(k), C.int(k), C.int(s), C.int(s), C.float(p), C.float(p)),
Michael Yang's avatar
Michael Yang committed
1131
1132
1133
	}
}

Michael Yang's avatar
Michael Yang committed
1134
1135
1136
1137
func (t *Tensor) Set(ctx ml.Context, t2 ml.Tensor, offset int, strides ...int) ml.Tensor {
	var tt *C.struct_ggml_tensor
	switch len(strides) {
	case 0:
Michael Yang's avatar
Michael Yang committed
1138
		tt = C.ggml_set_1d(ctx.(*Context).ctx, t.t, t2.(*Tensor).t, C.size_t(offset))
Michael Yang's avatar
Michael Yang committed
1139
	case 1:
Michael Yang's avatar
Michael Yang committed
1140
		tt = C.ggml_set_2d(ctx.(*Context).ctx, t.t, t2.(*Tensor).t, C.size_t(offset), C.size_t(strides[0]))
Michael Yang's avatar
Michael Yang committed
1141
1142
1143
1144
1145
1146
1147
	default:
		panic("unsupported number of dimensions")
	}

	return &Tensor{b: t.b, t: tt}
}

1148
1149
1150
1151
1152
1153
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
	}

1154
1155
1156
	query := t.Permute(ctx, 0, 2, 1, 3)
	key = key.Permute(ctx, 0, 2, 1, 3)

1157
1158
	if t.b.flashAttention {
		value = value.Permute(ctx, 0, 2, 1, 3)
1159

1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
		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)
	}
1173
}
1174
1175
1176
1177
1178
1179
1180

func (t *Tensor) Duplicate(ctx ml.Context) ml.Tensor {
	return &Tensor{
		b: t.b,
		t: C.ggml_dup(ctx.(*Context).ctx, t.t),
	}
}
Michael Yang's avatar
llama4  
Michael Yang committed
1181
1182
1183
1184
1185
1186
1187

func (t *Tensor) TopK(ctx ml.Context, k int) ml.Tensor {
	return &Tensor{
		b: t.b,
		t: C.ggml_top_k(ctx.(*Context).ctx, t.t, C.int(k)),
	}
}