ggml.go 29.7 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"
30
	"github.com/ollama/ollama/logutil"
Michael Yang's avatar
Michael Yang committed
31
	"github.com/ollama/ollama/ml"
32
	ggml "github.com/ollama/ollama/ml/backend/ggml/ggml/src"
Michael Yang's avatar
Michael Yang committed
33
34
35
	"golang.org/x/sync/errgroup"
)

Michael Yang's avatar
Michael Yang committed
36
37
38
39
40
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
41
	}
Michael Yang's avatar
Michael Yang committed
42
43

	return ds
44
}
Michael Yang's avatar
Michael Yang committed
45
46

type Backend struct {
47
48
49
50
51
52
	meta *fsggml.GGML

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

53
	tensors map[string]*C.struct_ggml_tensor
Michael Yang's avatar
Michael Yang committed
54
55

	// input is the backend used for inputs
56
	input *C.struct_ggml_backend_buffer_type
Michael Yang's avatar
Michael Yang committed
57
58

	// layers is the backend used for repeating layers
59
	layers map[int]*C.struct_ggml_backend_buffer_type
60

61
	flashAttention bool
Michael Yang's avatar
Michael Yang committed
62
63
64

	// maxGraphNodes is the maximum allowed number of graph nodes in this scheduler
	maxGraphNodes int
Michael Yang's avatar
Michael Yang committed
65
66
}

67
func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend, error) {
68
	meta, n, err := fsggml.Decode(r, -1)
Michael Yang's avatar
Michael Yang committed
69
70
71
72
73
74
75
76
77
78
79
80
81
82
	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()),
	)

83
	type deviceBufferType struct {
84
85
86
87
88
		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
89
	for _, d := range devices() {
90
91
		switch C.ggml_backend_dev_type(d) {
		case C.GGML_BACKEND_DEVICE_TYPE_CPU:
92
93
94
95
			if len(cpus) == 0 {
				// only the first cpu device should be used
				cpus = append(cpus, d)
			}
96
97
		case C.GGML_BACKEND_DEVICE_TYPE_ACCEL:
			accels = append(accels, d)
Michael Yang's avatar
Michael Yang committed
98
		case C.GGML_BACKEND_DEVICE_TYPE_GPU:
99
			gpus = append(gpus, d)
Michael Yang's avatar
Michael Yang committed
100
101
102
		}
	}

Michael Yang's avatar
Michael Yang committed
103
	// create list of buffer types for the cpu
Michael Yang's avatar
Michael Yang committed
104
	cpuDeviceBufferType := deviceBufferType{d: C.ggml_backend_dev_by_type(C.GGML_BACKEND_DEVICE_TYPE_CPU)}
105
106
107
108
	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
109
			cpuDeviceBufferType.bts = append(cpuDeviceBufferType.bts, C.ggml_backend_dev_buffer_type(d))
Michael Yang's avatar
Michael Yang committed
110
		}
111
112
	}

Michael Yang's avatar
Michael Yang committed
113
	// create list of buffer types for each gpu
114
	var gpuDeviceBufferTypes []deviceBufferType
115
116
	for _, d := range gpus {
		bt := C.ggml_backend_dev_buffer_type(d)
117
		gpuDeviceBufferTypes = append(gpuDeviceBufferTypes, deviceBufferType{
118
			d:   d,
Michael Yang's avatar
Michael Yang committed
119
			bts: append([]*C.struct_ggml_backend_buffer_type{bt}, cpuDeviceBufferType.bts...),
120
		})
Michael Yang's avatar
Michael Yang committed
121
122
	}

Michael Yang's avatar
Michael Yang committed
123
124
125
126
127
	useDefaultSplit := true
	for _, s := range params.TensorSplit {
		if s != 0 {
			useDefaultSplit = false
			break
128
		}
Michael Yang's avatar
Michael Yang committed
129
	}
130

Michael Yang's avatar
Michael Yang committed
131
132
133
134
	// calculate splits
	splits := make([]float32, len(gpus))
	if useDefaultSplit {
		// default: split on free memory
135
136
137
138
139
		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
140
141
	} else {
		splits = params.TensorSplit
142
143
144
	}

	var sum float32
Michael Yang's avatar
Michael Yang committed
145
	// cumulative sum of all splits
146
147
148
149
150
	for i := range splits {
		sum += splits[i]
		splits[i] = sum
	}

Michael Yang's avatar
Michael Yang committed
151
	// normalize splits
152
	for i := range splits {
153
		splits[i] /= sum
154
155
	}

Michael Yang's avatar
Michael Yang committed
156
	// inputs always use cpu
Michael Yang's avatar
Michael Yang committed
157
	input := cpuDeviceBufferType
158

159
	blocks := int(meta.KV().BlockCount())
Michael Yang's avatar
Michael Yang committed
160
161
162
163

	// 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
164
	assignLayer := func(i int) deviceBufferType {
Michael Yang's avatar
Michael Yang committed
165
		if i < gpuRangeStart || i >= gpuRangeStop {
Michael Yang's avatar
Michael Yang committed
166
			return cpuDeviceBufferType
167
		}
168

Michael Yang's avatar
Michael Yang committed
169
		index := slices.IndexFunc(splits, func(f float32) bool { return float32(i-gpuRangeStart)/float32(gpuRangeStop-gpuRangeStart) < f })
170
		if index < 0 || index >= len(gpuDeviceBufferTypes) {
Michael Yang's avatar
Michael Yang committed
171
			return cpuDeviceBufferType
172
173
174
		}

		return gpuDeviceBufferTypes[index]
175
176
	}

Michael Yang's avatar
Michael Yang committed
177
	// repeating layers are assigned based on their index in reverse order, e.g. i / (block_count + 1)
178
	layers := make([]deviceBufferType, blocks)
179
	for i := range layers {
180
		layers[i] = assignLayer(i)
181
182
	}

Michael Yang's avatar
Michael Yang committed
183
	// outputs are assigned iff allowed by splits and configured number of gpu layers
184
	output := assignLayer(blocks)
185
186
187

	maxTensors := len(meta.Tensors().Items())
	maxTensors += 1
Michael Yang's avatar
Michael Yang committed
188
	// each layer has at most 2 extra tensors for rope operations
189
190
	maxTensors += blocks * 2

191
	type tensor struct {
192
		source *fsggml.Tensor
193
194
195
		target string
	}

Michael Yang's avatar
Michael Yang committed
196
	// some tensors are mapped to different names so keep a list
197
198
	targets := make(map[string][]string)

Michael Yang's avatar
Michael Yang committed
199
	// contexts are shared by tensors of the same buffer type
200
	ctxs := make(map[*C.struct_ggml_backend_buffer_type]*C.struct_ggml_context)
201
	createTensor := func(t tensor, bts []*C.struct_ggml_backend_buffer_type) *C.struct_ggml_tensor {
202
203
204
205
206
207
208
		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
209

210
211
212
213
214
215
216
217
			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
218
			defer C.free(unsafe.Pointer(cname))
219
220
221
222
			if tt := C.ggml_get_tensor(ctxs[bt], cname); tt != nil {
				return tt
			}

223
			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
224
225
			C.ggml_set_name(tt, cname)

226
			slog.Log(context.TODO(), logutil.LevelTrace, "created tensor", "name", name, "shape", t.source.Shape, "dtype", t.source.Kind, "buffer_type", C.GoString(C.ggml_backend_buft_name(bt)))
227
228
229
230
231
			//nolint:staticcheck // TODO: check if buffer type supports this tensor
			return tt
		}

		return nil
Michael Yang's avatar
Michael Yang committed
232
233
	}

234
	contains := func(s string, parts ...string) bool {
235
236
237
238
239
240
241
242
		split := strings.Split(s, ".")
		for _, part := range parts {
			if slices.Contains(split, part) {
				return true
			}
		}

		return false
Michael Yang's avatar
Michael Yang committed
243
244
	}

245
246
	for _, t := range meta.Tensors().Items() {
		switch {
247
		case contains(t.Name, "position_embd", "token_embd", "token_norm_embd", "token_types"):
248
			createTensor(tensor{source: t}, input.bts)
Michael Yang's avatar
Michael Yang committed
249
250
251
			if _, ok := meta.Tensors().GroupLayers()["output"]; !ok && t.Name == "token_embd.weight" {
				createTensor(tensor{source: t, target: "output.weight"}, output.bts)
			}
252
		case contains(t.Name, "cls", "output", "output_norm"):
253
			createTensor(tensor{source: t}, output.bts)
254
		case strings.HasPrefix(t.Name, "v.") || strings.HasPrefix(t.Name, "mm."):
Michael Yang's avatar
Michael Yang committed
255
			// TODO: assign vision tensors to the gpu if possible
Michael Yang's avatar
Michael Yang committed
256
			createTensor(tensor{source: t}, output.bts)
Michael Yang's avatar
Michael Yang committed
257
258
259
260
261
262
263
264
		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)
			}
265
		default:
Michael Yang's avatar
Michael Yang committed
266
267
268
269
			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
270
				}
Michael Yang's avatar
Michael Yang committed
271
			}
272

Michael Yang's avatar
Michael Yang committed
273
274
			if layerIndex >= 0 {
				createTensor(tensor{source: t}, layers[layerIndex].bts)
275
			} else {
Michael Yang's avatar
Michael Yang committed
276
277
				// load all other tensors on the cpu
				createTensor(tensor{source: t}, input.bts)
278
279
280
			}
		}
	}
Michael Yang's avatar
Michael Yang committed
281

Michael Yang's avatar
Michael Yang committed
282
283
	// allocate buffers for each context
	bbs := make(map[*C.struct_ggml_context]*C.struct_ggml_backend_buffer, len(ctxs))
284
285
286
287
288
289
	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)
290
291
292
293
		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)))
		}

294
		C.ggml_backend_buffer_set_usage(b, C.GGML_BACKEND_BUFFER_USAGE_WEIGHTS)
Michael Yang's avatar
Michael Yang committed
295
		bbs[c] = b
296
297
298
	}

	for bs := range maps.Values(bbs) {
Michael Yang's avatar
Michael Yang committed
299
		slog.Info("model weights", "buffer", C.GoString(C.ggml_backend_buffer_name(bs)), "size", format.HumanBytes2(uint64(C.ggml_backend_buffer_get_size(bs))))
300
301
	}

Michael Yang's avatar
Michael Yang committed
302
	// map tensor names to tensors for easy lookup later
303
304
305
306
307
308
309
	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
		}
	}

310
311
312
313
314
	var doneBytes atomic.Uint64
	totalBytes := uint64(n) - meta.Tensors().Offset

	g, ctx := errgroup.WithContext(ctx)
	g.SetLimit(runtime.GOMAXPROCS(0))
315
	for _, t := range meta.Tensors().Items() {
316
		t := t
317
318
319
320
		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]
321
322
323
				if target == "" {
					target = t.Name
				}
324

325
326
327
328
				tt, ok := tensors[target]
				if !ok {
					return fmt.Errorf("unassigned tensor: %s", t.Name)
				}
Michael Yang's avatar
Michael Yang committed
329

330
331
332
				tts[i] = tt
			}

333
334
335
336
			// 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
337
				slog.Warn("file open error", "file", r.Name(), "error", err)
338
339
340
341
				return err
			}
			defer file.Close()
			sr := io.NewSectionReader(file, int64(meta.Tensors().Offset+t.Offset), int64(t.Size()))
342
343
344
345
			bts := make([]byte, 128*format.KibiByte)

			var s uint64
			for s < t.Size() {
346
347
348
349
350
				// 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
				}

351
352
				n, err := io.ReadFull(sr, bts[:min(len(bts), int(t.Size()-s))])
				if err != nil {
Jesse Gross's avatar
Jesse Gross committed
353
					slog.Warn("file read error", "file", r.Name(), "error", err)
354
					return err
355
				}
Michael Yang's avatar
Michael Yang committed
356

357
358
				for _, tt := range tts {
					C.ggml_backend_tensor_set(tt, unsafe.Pointer(&bts[0]), C.size_t(s), C.size_t(n))
359
				}
Michael Yang's avatar
Michael Yang committed
360

361
362
363
364
365
366
367
368
369
370
				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
371
372
	}

373
	if err := g.Wait(); err != nil {
Michael Yang's avatar
Michael Yang committed
374
375
376
		return nil, err
	}

377
378
	// 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
379
380
381
382

	// 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
383
384
385
386
	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)

387
388
389
		deviceBufferTypes[d] = bt

		schedBackends = append(schedBackends, b)
Michael Yang's avatar
Michael Yang committed
390
		schedBufts = append(schedBufts, bt)
391

392
		if C.ggml_backend_is_cpu(b) {
Michael Yang's avatar
Michael Yang committed
393
			// set number of threads for cpu backend
Michael Yang's avatar
Michael Yang committed
394
			C.ggml_backend_cpu_set_n_threads(b, C.int(Threads(params.NumThreads)))
395
		}
396
397
	}

Michael Yang's avatar
Michael Yang committed
398
	maxGraphNodes := max(8192, len(meta.Tensors().Items())*5)
Michael Yang's avatar
Michael Yang committed
399
	return &Backend{
400
		flashAttention: params.FlashAttention,
401
402
		meta:           meta,
		tensors:        tensors,
403
		sched: C.ggml_backend_sched_new(
Michael Yang's avatar
Michael Yang committed
404
405
406
407
			(*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),
408
			C._Bool(len(gpus) > 1 && slices.Contains(gpus, output.d)),
409
			C._Bool(false),
410
		),
411
412
413
		schedBackends: schedBackends,
		schedBufts:    schedBufts,
		input:         deviceBufferTypes[input.d],
414
415
		layers: func() map[int]*C.struct_ggml_backend_buffer_type {
			m := make(map[int]*C.struct_ggml_backend_buffer_type)
416
			for i, layer := range layers {
417
				m[i] = deviceBufferTypes[layer.d]
418
419
420
			}
			return m
		}(),
Michael Yang's avatar
Michael Yang committed
421
		maxGraphNodes: maxGraphNodes,
Michael Yang's avatar
Michael Yang committed
422
423
424
425
426
427
428
	}, nil
}

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

429
func (b *Backend) Config() fs.Config {
Michael Yang's avatar
Michael Yang committed
430
431
432
433
	return b.meta.KV()
}

func (b *Backend) Get(name string) ml.Tensor {
434
435
	if t, ok := b.tensors[name]; ok {
		return &Tensor{b: b, t: t}
Michael Yang's avatar
Michael Yang committed
436
437
438
439
440
441
	}

	return nil
}

func (b *Backend) NewContext() ml.Context {
Michael Yang's avatar
Michael Yang committed
442
	return b.NewContextSize(b.maxGraphNodes)
443
444
445
}

func (b *Backend) NewContextSize(n int) ml.Context {
Jesse Gross's avatar
Jesse Gross committed
446
447
448
449
	if n > b.maxGraphNodes {
		panic(fmt.Errorf("requested number of graph nodes (%v) for new context exceeds maximum (%v)", n, b.maxGraphNodes))
	}

450
451
	var allocatedBuffers []*C.struct_ggml_backend_buffer

Michael Yang's avatar
Michael Yang committed
452
	return &Context{
453
454
		b:             b,
		maxGraphNodes: n,
455
		ctx: C.ggml_init(C.struct_ggml_init_params{
456
			mem_size: C.size_t(n)*C.ggml_tensor_overhead() + C.ggml_graph_overhead_custom(C.size_t(n), false),
457
458
			no_alloc: true,
		}),
459
		allocatedBuffers: &allocatedBuffers,
Michael Yang's avatar
Michael Yang committed
460
461
462
	}
}

463
func (b *Backend) CacheConfig() ml.CacheConfig {
464
465
466
467
468
	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}
	}
469
470
}

Michael Yang's avatar
Michael Yang committed
471
type Context struct {
472
	b *Backend
Michael Yang's avatar
Michael Yang committed
473

474
	ctx   *C.struct_ggml_context
Michael Yang's avatar
Michael Yang committed
475
	graph *C.struct_ggml_cgraph
476

477
478
	// buft is the buffer type used for new tensors
	buft *C.struct_ggml_backend_buffer_type
479

480
481
482
483
	// 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
484
	// maxGraphNodes is the maximum allowed number of graph nodes in this context
485
	maxGraphNodes int
Michael Yang's avatar
Michael Yang committed
486
487
}

488
func (c *Context) Input() ml.Context {
Michael Yang's avatar
Michael Yang committed
489
	if c.b.input != nil {
490
		return &Context{
491
492
493
494
495
			b:                c.b,
			ctx:              c.ctx,
			buft:             c.b.input,
			allocatedBuffers: c.allocatedBuffers,
			maxGraphNodes:    c.maxGraphNodes,
496
497
498
		}
	}

499
	return c
500
501
}

502
func (c *Context) Layer(i int) ml.Context {
503
	if buft, ok := c.b.layers[i]; ok {
504
		return &Context{
505
506
507
508
509
			b:                c.b,
			ctx:              c.ctx,
			buft:             buft,
			allocatedBuffers: c.allocatedBuffers,
			maxGraphNodes:    c.maxGraphNodes,
510
511
512
		}
	}

513
	return c
514
515
}

516
func (c *Context) Forward(tensors ...ml.Tensor) ml.Context {
Michael Yang's avatar
Michael Yang committed
517
	if c.graph == nil {
518
		c.graph = C.ggml_new_graph_custom(c.ctx, C.size_t(c.maxGraphNodes), false)
Michael Yang's avatar
Michael Yang committed
519
520
	}

521
522
523
524
525
	for _, tensor := range tensors {
		C.ggml_build_forward_expand(c.graph, tensor.(*Tensor).t)
	}

	return c
Michael Yang's avatar
Michael Yang committed
526
527
}

528
func (c *Context) Compute(tensors ...ml.Tensor) {
529
	C.ggml_backend_sched_graph_compute_async(c.b.sched, c.graph)
Michael Yang's avatar
Michael Yang committed
530
	C.ggml_backend_sched_reset(c.b.sched)
Michael Yang's avatar
Michael Yang committed
531

532
533
534
	needSync := true
	sync := func() {
		if needSync {
535
			C.ggml_backend_sched_synchronize(c.b.sched)
536
537
538
			needSync = false
		}
	}
Michael Yang's avatar
Michael Yang committed
539

540
541
542
	for _, t := range tensors {
		if C.ggml_nbytes(t.(*Tensor).t) > 0 {
			t.(*Tensor).sync = sync
543
544
		}
	}
Michael Yang's avatar
Michael Yang committed
545
546
}

547
func (c *Context) Reserve() error {
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
	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
}

565
func (c *Context) MaxGraphNodes() int {
566
	return c.maxGraphNodes
Jesse Gross's avatar
Jesse Gross committed
567
568
}

569
570
571
func shapeToGGML(shape []int) *C.int64_t {
	sh := make([]C.int64_t, len(shape))
	for i, s := range shape {
572
		sh[i] = C.int64_t(s)
573
574
575
576
577
	}

	return &sh[0]
}

578
579
580
581
func pad(length, pad C.size_t) C.size_t {
	return ((length + pad - 1) / pad) * pad
}

582
func (c *Context) newTensor(dtype ml.DType, shape []int) (ml.Tensor, error) {
583
	if c.buft == nil {
584
		panic("set Input or Layer before creating tensors")
585
586
	}

Michael Yang's avatar
Michael Yang committed
587
588
589
590
591
592
	var cdtype uint32
	switch dtype {
	case ml.DTypeF32:
		cdtype = C.GGML_TYPE_F32
	case ml.DTypeF16:
		cdtype = C.GGML_TYPE_F16
593
594
595
596
	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
597
598
599
600
601
602
	case ml.DTypeI32:
		cdtype = C.GGML_TYPE_I32
	default:
		panic("unsupported dtype")
	}

Jesse Gross's avatar
Jesse Gross committed
603
	if len(shape) < 1 || shape[0] == 0 {
Michael Yang's avatar
Michael Yang committed
604
		var shape C.int64_t = 0
605
		return &Tensor{b: c.b, t: C.ggml_new_tensor(c.ctx, cdtype, 1, &shape)}, nil
Michael Yang's avatar
Michael Yang committed
606
	} else if len(shape) > 4 {
Michael Yang's avatar
Michael Yang committed
607
608
609
610
611
612
613
614
615
		panic("unsupported number of dimensions")
	}

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

Michael Yang's avatar
Michael Yang committed
616
	t := C.ggml_new_tensor(c.ctx, cdtype, C.int(len(shape)), shapeToGGML(shape))
617
618
	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)
619
620
621
	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)))
	}
622
	*c.allocatedBuffers = append(*c.allocatedBuffers, b)
623

Michael Yang's avatar
Michael Yang committed
624
	C.ggml_backend_tensor_alloc(b, t, C.ggml_backend_buffer_get_base(b))
625
	return &Tensor{b: c.b, t: t}, nil
626
627
}

628
func (c *Context) Empty(dtype ml.DType, shape ...int) ml.Tensor {
629
630
631
632
633
634
	t, err := c.newTensor(dtype, shape)
	if err != nil {
		panic(err)
	}

	return t
635
636
}

637
func (c *Context) Zeros(dtype ml.DType, shape ...int) ml.Tensor {
638
639
640
641
642
	t, err := c.newTensor(dtype, shape)
	if err != nil {
		panic(err)
	}

643
644
	C.ggml_set_zero(t.(*Tensor).t)
	return t
Michael Yang's avatar
Michael Yang committed
645
646
}

647
func checkShape[S ~[]E, E any](s S, shape ...int) error {
Michael Yang's avatar
Michael Yang committed
648
	n := len(s)
Jesse Gross's avatar
Jesse Gross committed
649
650
651
652
653

	if n == 0 {
		return nil
	}

Michael Yang's avatar
Michael Yang committed
654
655
656
657
658
	for _, v := range shape {
		n /= v
	}

	if n != 1 {
659
		return fmt.Errorf("invalid shape: %v", shape)
Michael Yang's avatar
Michael Yang committed
660
661
	}

662
	return nil
Michael Yang's avatar
Michael Yang committed
663
664
}

665
func (c *Context) FromFloatSlice(s []float32, shape ...int) (ml.Tensor, error) {
Jesse Gross's avatar
Jesse Gross committed
666
	if err := checkShape(s, shape...); err != nil {
667
668
669
		return nil, err
	}

670
671
672
673
674
	t, err := c.newTensor(ml.DTypeF32, shape)
	if err != nil {
		return nil, err
	}

Jesse Gross's avatar
Jesse Gross committed
675
676
677
678
	if len(s) > 0 {
		C.ggml_backend_tensor_set(t.(*Tensor).t, unsafe.Pointer(&s[0]), 0, C.ggml_nbytes(t.(*Tensor).t))
	}

679
	return t, nil
Michael Yang's avatar
Michael Yang committed
680
681
}

682
func (c *Context) FromIntSlice(s []int32, shape ...int) (ml.Tensor, error) {
Jesse Gross's avatar
Jesse Gross committed
683
	if err := checkShape(s, shape...); err != nil {
684
685
686
		return nil, err
	}

687
688
689
690
691
	t, err := c.newTensor(ml.DTypeI32, shape)
	if err != nil {
		return nil, err
	}

Jesse Gross's avatar
Jesse Gross committed
692
693
694
695
	if len(s) > 0 {
		C.ggml_backend_tensor_set(t.(*Tensor).t, unsafe.Pointer(&s[0]), 0, C.ggml_nbytes(t.(*Tensor).t))
	}

696
	return t, nil
Michael Yang's avatar
Michael Yang committed
697
698
}

Michael Yang's avatar
arange  
Michael Yang committed
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
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
725
726
func (c *Context) Close() {
	if c != nil {
727
728
729
730
731
		for _, b := range *c.allocatedBuffers {
			C.ggml_backend_buffer_free(b)
		}
		*c.allocatedBuffers = nil

732
733
		C.ggml_free(c.ctx)
	}
Michael Yang's avatar
Michael Yang committed
734
735
736
}

type Tensor struct {
737
	b    *Backend
Michael Yang's avatar
Michael Yang committed
738
	t    *C.struct_ggml_tensor
739
	sync func()
Michael Yang's avatar
Michael Yang committed
740
741
742
743
744
745
746
747
748
749
}

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()),
	)
}

750
751
func (t *Tensor) Dim(n int) int {
	return int(t.t.ne[n])
Michael Yang's avatar
Michael Yang committed
752
753
}

754
755
func (t *Tensor) Stride(n int) int {
	return int(t.t.nb[n])
Michael Yang's avatar
Michael Yang committed
756
757
}

758
759
func (t *Tensor) Shape() []int {
	shape := make([]int, C.ggml_n_dims(t.t))
Michael Yang's avatar
Michael Yang committed
760
761
762
763
764
765
766
	for i := range shape {
		shape[i] = t.Dim(i)
	}

	return shape
}

767
768
769
770
771
772
773
774
775
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
776
777
}

778
779
780
781
782
783
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
784
785
786
787
788
789
790
791
792
	}

	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
793
794
	case C.GGML_TYPE_F16:
		return ml.DTypeF16
795
796
797
798
	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
799
800
801
802
803
804
805
	case C.GGML_TYPE_I32:
		return ml.DTypeI32
	default:
		return ml.DTypeOther
	}
}

806
807
808
809
810
811
812
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
813
814
func (t *Tensor) Add(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
	return &Tensor{
815
		b: t.b,
Michael Yang's avatar
Michael Yang committed
816
817
818
819
		t: C.ggml_add(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
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
841
842
843
844
845
846
847
848
849
850
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{
851
		b: t.b,
Michael Yang's avatar
Michael Yang committed
852
853
854
855
856
857
		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{
858
		b: t.b,
Michael Yang's avatar
Michael Yang committed
859
860
861
862
863
864
		t: C.ggml_cont(ctx.(*Context).ctx, t.t),
	}
}

func (t *Tensor) Mul(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
	return &Tensor{
865
		b: t.b,
Michael Yang's avatar
Michael Yang committed
866
867
868
869
870
871
		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{
872
		b: t.b,
Michael Yang's avatar
Michael Yang committed
873
874
875
876
		t: C.ggml_mul_mat(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

877
878
879
880
881
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{
882
		b: t.b,
883
884
885
886
		t: mul,
	}
}

Michael Yang's avatar
llama4  
Michael Yang committed
887
888
889
890
891
892
893
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
894
func (t *Tensor) LayerNorm(ctx ml.Context, w, b ml.Tensor, eps float32) ml.Tensor {
Michael Yang's avatar
llama4  
Michael Yang committed
895
896
897
898
899
900
	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
901
902
	}

Michael Yang's avatar
llama4  
Michael Yang committed
903
	return &Tensor{b: t.b, t: tt}
Michael Yang's avatar
Michael Yang committed
904
905
906
}

func (t *Tensor) RMSNorm(ctx ml.Context, w ml.Tensor, eps float32) ml.Tensor {
Michael Yang's avatar
llama4  
Michael Yang committed
907
908
909
910
911
912
	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
913
914
}

915
func (t *Tensor) Pad(ctx ml.Context, shape ...int) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
916
917
	if len(shape) != 4 {
		panic("expected 4 dimensions")
918
919
	} else if shape[3] != 0 {
		panic("cuda does not support 4d tensors")
Michael Yang's avatar
Michael Yang committed
920
921
922
	}

	return &Tensor{
923
		b: t.b,
Michael Yang's avatar
Michael Yang committed
924
925
926
927
928
929
930
931
932
933
		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{
934
		b: t.b,
Michael Yang's avatar
Michael Yang committed
935
936
937
938
939
940
		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{
941
		b: t.b,
Michael Yang's avatar
Michael Yang committed
942
943
944
945
946
947
		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{
948
		b: t.b,
Michael Yang's avatar
Michael Yang committed
949
950
951
952
		t: C.ggml_cpy(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

953
func (t *Tensor) Reshape(ctx ml.Context, shape ...int) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
954
955
956
	switch len(shape) {
	case 1:
		return &Tensor{
957
			b: t.b,
Michael Yang's avatar
Michael Yang committed
958
959
960
961
			t: C.ggml_reshape_1d(ctx.(*Context).ctx, t.t, C.int64_t(shape[0])),
		}
	case 2:
		return &Tensor{
962
			b: t.b,
Michael Yang's avatar
Michael Yang committed
963
964
965
966
			t: C.ggml_reshape_2d(ctx.(*Context).ctx, t.t, C.int64_t(shape[0]), C.int64_t(shape[1])),
		}
	case 3:
		return &Tensor{
967
			b: t.b,
Michael Yang's avatar
Michael Yang committed
968
969
970
971
			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{
972
			b: t.b,
Michael Yang's avatar
Michael Yang committed
973
974
975
976
977
978
979
980
981
			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{
982
		b: t.b,
Michael Yang's avatar
Michael Yang committed
983
984
985
986
987
988
		t: C.ggml_scale(ctx.(*Context).ctx, t.t, (C.float)(s)),
	}
}

func (t *Tensor) Softmax(ctx ml.Context) ml.Tensor {
	return &Tensor{
989
		b: t.b,
Michael Yang's avatar
Michael Yang committed
990
991
992
993
		t: C.ggml_soft_max(ctx.(*Context).ctx, t.t),
	}
}

994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
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
1008
1009
func (t *Tensor) Tanh(ctx ml.Context) ml.Tensor {
	return &Tensor{
1010
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1011
1012
1013
1014
		t: C.ggml_tanh_inplace(ctx.(*Context).ctx, t.t),
	}
}

Michael Yang's avatar
llama4  
Michael Yang committed
1015
1016
1017
1018
1019
1020
1021
func (t *Tensor) Sigmoid(ctx ml.Context) ml.Tensor {
	return &Tensor{
		b: t.b,
		t: C.ggml_sigmoid_inplace(ctx.(*Context).ctx, t.t),
	}
}

Michael Yang's avatar
Michael Yang committed
1022
1023
1024
1025
func (t *Tensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
	switch len(shape) {
	case 1:
		return &Tensor{
1026
			b: t.b,
Michael Yang's avatar
Michael Yang committed
1027
1028
1029
1030
			t: C.ggml_view_1d(ctx.(*Context).ctx, t.t, C.int64_t(shape[0]), C.size_t(offset)),
		}
	case 3:
		return &Tensor{
1031
			b: t.b,
Michael Yang's avatar
Michael Yang committed
1032
1033
1034
1035
1036
1037
1038
			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{
1039
			b: t.b,
Michael Yang's avatar
Michael Yang committed
1040
1041
1042
1043
1044
1045
1046
			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{
1047
			b: t.b,
Michael Yang's avatar
Michael Yang committed
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
			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
1059
1060
1061
1062
	ropeTypeNorm   C.int = 0
	ropeTypeNeox   C.int = 2
	ropeTypeMrope  C.int = 8
	ropeTypeVision C.int = 24
Michael Yang's avatar
Michael Yang committed
1063
1064
)

1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
func (t *Tensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, ropeDim, ropeType uint32, ropeBase, ropeScale float32, options ...ml.RopeOption) ml.Tensor {
	// Default options
	opts := &ml.RopeOptions{
		OriginalContextLen: 131072,
	}

	// Apply any provided options
	for _, option := range options {
		option(opts)
	}

Michael Yang's avatar
Michael Yang committed
1076
	if ropeFactors == nil {
1077
		ropeFactors = &Tensor{b: t.b}
Michael Yang's avatar
Michael Yang committed
1078
1079
	}

Jesse Gross's avatar
Jesse Gross committed
1080
1081
1082
1083
1084
	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
1085
	return &Tensor{
1086
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1087
		t: C.ggml_rope_ext(
1088
1089
1090
1091
			ctx.(*Context).ctx,
			dequant,
			positionIDs.(*Tensor).t,
			ropeFactors.(*Tensor).t,
Michael Yang's avatar
Michael Yang committed
1092
			C.int(ropeDim),
Patrick Devine's avatar
Patrick Devine committed
1093
			C.int(ropeType),
1094
			C.int(opts.OriginalContextLen),
Michael Yang's avatar
Michael Yang committed
1095
1096
			C.float(ropeBase),
			C.float(ropeScale),
1097
1098
1099
1100
			C.float(0.0),
			C.float(1.0),
			C.float(32.0),
			C.float(1.0),
Michael Yang's avatar
Michael Yang committed
1101
1102
1103
1104
		),
	}
}

1105
1106
1107
1108
1109
1110
1111
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
1112
1113
func (t *Tensor) GELU(ctx ml.Context) ml.Tensor {
	return &Tensor{
1114
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1115
1116
1117
1118
1119
1120
		t: C.ggml_gelu_inplace(ctx.(*Context).ctx, t.t),
	}
}

func (t *Tensor) SILU(ctx ml.Context) ml.Tensor {
	return &Tensor{
1121
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1122
1123
1124
1125
1126
1127
		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{
1128
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1129
1130
1131
		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)),
	}
}
1132

Michael Yang's avatar
Michael Yang committed
1133
func (t *Tensor) AvgPool2D(ctx ml.Context, k, s int, p float32) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
1134
1135
	return &Tensor{
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1136
		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
1137
1138
1139
	}
}

Michael Yang's avatar
Michael Yang committed
1140
1141
1142
1143
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
1144
		tt = C.ggml_set_1d(ctx.(*Context).ctx, t.t, t2.(*Tensor).t, C.size_t(offset))
Michael Yang's avatar
Michael Yang committed
1145
	case 1:
Michael Yang's avatar
Michael Yang committed
1146
		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
1147
1148
1149
1150
1151
1152
1153
	default:
		panic("unsupported number of dimensions")
	}

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

1154
1155
1156
1157
1158
1159
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
	}

1160
1161
1162
	query := t.Permute(ctx, 0, 2, 1, 3)
	key = key.Permute(ctx, 0, 2, 1, 3)

1163
1164
	if t.b.flashAttention {
		value = value.Permute(ctx, 0, 2, 1, 3)
1165

1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
		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)
	}
1179
}
1180
1181
1182
1183
1184
1185
1186

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
1187
1188
1189
1190
1191
1192
1193

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)),
	}
}
1194
1195
1196
1197
1198
1199
1200

func (t *Tensor) Argsort(ctx ml.Context) ml.Tensor {
	return &Tensor{
		b: t.b,
		t: C.ggml_argsort(ctx.(*Context).ctx, t.t, C.GGML_SORT_ORDER_ASC),
	}
}