ggml.go 26.6 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"
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
	"runtime"
19
20
21
	"slices"
	"strconv"
	"strings"
22
	"sync/atomic"
23
	"unicode"
Michael Yang's avatar
Michael Yang committed
24
25
26
	"unsafe"

	"github.com/ollama/ollama/format"
27
28
	"github.com/ollama/ollama/fs"
	fsggml "github.com/ollama/ollama/fs/ggml"
Michael Yang's avatar
Michael Yang committed
29
	"github.com/ollama/ollama/ml"
30
	ggml "github.com/ollama/ollama/ml/backend/ggml/ggml/src"
Michael Yang's avatar
Michael Yang committed
31
32
33
	"golang.org/x/sync/errgroup"
)

Michael Yang's avatar
Michael Yang committed
34
35
36
37
38
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
39
	}
Michael Yang's avatar
Michael Yang committed
40
41

	return ds
42
}
Michael Yang's avatar
Michael Yang committed
43
44

type Backend struct {
45
	meta    *fsggml.GGML
46
47
	sched   *C.struct_ggml_backend_sched
	tensors map[string]*C.struct_ggml_tensor
Michael Yang's avatar
Michael Yang committed
48
49

	// input is the backend used for inputs
50
	input *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(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend, error) {
62
	meta, n, err := fsggml.Decode(r, -1)
Michael Yang's avatar
Michael Yang committed
63
64
65
66
67
68
69
70
71
72
73
74
75
76
	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
	type tensor struct {
186
		source *fsggml.Tensor
187
188
189
		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)
Michael Yang's avatar
Michael Yang committed
243
244
245
			if _, ok := meta.Tensors().GroupLayers()["output"]; !ok && t.Name == "token_embd.weight" {
				createTensor(tensor{source: t, target: "output.weight"}, output.bts)
			}
246
		case contains(t.Name, "cls", "output", "output_norm"):
247
			createTensor(tensor{source: t}, output.bts)
248
		case strings.HasPrefix(t.Name, "v.") || strings.HasPrefix(t.Name, "mm."):
Michael Yang's avatar
Michael Yang committed
249
			// TODO: assign vision tensors to the gpu if possible
Michael Yang's avatar
Michael Yang committed
250
			createTensor(tensor{source: t}, output.bts)
Michael Yang's avatar
Michael Yang committed
251
252
253
254
255
256
257
258
		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)
			}
259
		default:
Michael Yang's avatar
Michael Yang committed
260
261
262
263
			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
264
				}
Michael Yang's avatar
Michael Yang committed
265
			}
266

Michael Yang's avatar
Michael Yang committed
267
268
			if layerIndex >= 0 {
				createTensor(tensor{source: t}, layers[layerIndex].bts)
269
			} else {
Michael Yang's avatar
Michael Yang committed
270
271
				// load all other tensors on the cpu
				createTensor(tensor{source: t}, input.bts)
272
273
274
			}
		}
	}
Michael Yang's avatar
Michael Yang committed
275

Michael Yang's avatar
Michael Yang committed
276
277
	// allocate buffers for each context
	bbs := make(map[*C.struct_ggml_context]*C.struct_ggml_backend_buffer, len(ctxs))
278
279
280
281
282
283
284
	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
285
		bbs[c] = b
286
287
288
	}

	for bs := range maps.Values(bbs) {
Michael Yang's avatar
Michael Yang committed
289
		slog.Info("model weights", "buffer", C.GoString(C.ggml_backend_buffer_name(bs)), "size", format.HumanBytes2(uint64(C.ggml_backend_buffer_get_size(bs))))
290
291
	}

Michael Yang's avatar
Michael Yang committed
292
	// map tensor names to tensors for easy lookup later
293
294
295
296
297
298
299
	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
		}
	}

300
301
302
303
304
	var doneBytes atomic.Uint64
	totalBytes := uint64(n) - meta.Tensors().Offset

	g, ctx := errgroup.WithContext(ctx)
	g.SetLimit(runtime.GOMAXPROCS(0))
305
	for _, t := range meta.Tensors().Items() {
306
307
308
309
		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]
310
311
312
				if target == "" {
					target = t.Name
				}
313

314
315
316
317
				tt, ok := tensors[target]
				if !ok {
					return fmt.Errorf("unassigned tensor: %s", t.Name)
				}
Michael Yang's avatar
Michael Yang committed
318

319
320
321
				tts[i] = tt
			}

322
323
324
325
326
327
328
329
			// 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 {
				return err
			}
			defer file.Close()
			sr := io.NewSectionReader(file, int64(meta.Tensors().Offset+t.Offset), int64(t.Size()))
330
331
332
333
334
335
336
			bts := make([]byte, 128*format.KibiByte)

			var s uint64
			for s < t.Size() {
				n, err := io.ReadFull(sr, bts[:min(len(bts), int(t.Size()-s))])
				if err != nil {
					return err
337
				}
Michael Yang's avatar
Michael Yang committed
338

339
340
				for _, tt := range tts {
					C.ggml_backend_tensor_set(tt, unsafe.Pointer(&bts[0]), C.size_t(s), C.size_t(n))
341
				}
Michael Yang's avatar
Michael Yang committed
342

343
344
345
346
347
348
349
350
351
352
				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
353
354
	}

355
356
357
358
359
360
361
362
	// start a goroutine to cancel the errgroup if the parent context is done
	go func() {
		<-ctx.Done()
		g.Go(func() error {
			return ctx.Err()
		})
	}()

363
	if err := g.Wait(); err != nil {
Michael Yang's avatar
Michael Yang committed
364
365
366
		return nil, err
	}

367
368
	// 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
369
370
371
372

	// 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
373
374
375
376
	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 {
377
378
			// use the first gpu host buffer type for gpu if possible
			if hbt := C.ggml_backend_dev_host_buffer_type(gpus[0]); hbt != nil {
379
380
381
382
				bt = hbt
			}
		}

383
384
385
		deviceBufferTypes[d] = bt

		schedBackends = append(schedBackends, b)
Michael Yang's avatar
Michael Yang committed
386
		schedBufts = append(schedBufts, bt)
387

388
		slog.Info("compute graph", "backend", C.GoString(C.ggml_backend_name(b)), "buffer_type", C.GoString(C.ggml_backend_buft_name(bt)))
389
390

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

Michael Yang's avatar
Michael Yang committed
396
	maxGraphNodes := max(8192, len(meta.Tensors().Items())*5)
Michael Yang's avatar
Michael Yang committed
397
	return &Backend{
398
		flashAttention: params.FlashAttention,
399
400
		meta:           meta,
		tensors:        tensors,
401
		sched: C.ggml_backend_sched_new(
Michael Yang's avatar
Michael Yang committed
402
403
404
405
			(*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),
406
			C._Bool(len(gpus) > 1 && slices.Contains(gpus, output.d)),
407
		),
408
		input: deviceBufferTypes[input.d],
409
410
		layers: func() map[int]*C.struct_ggml_backend_buffer_type {
			m := make(map[int]*C.struct_ggml_backend_buffer_type)
411
			for i, layer := range layers {
412
				m[i] = deviceBufferTypes[layer.d]
413
414
415
			}
			return m
		}(),
Michael Yang's avatar
Michael Yang committed
416
		maxGraphNodes: maxGraphNodes,
Michael Yang's avatar
Michael Yang committed
417
418
419
420
421
422
423
	}, nil
}

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

424
func (b *Backend) Config() fs.Config {
Michael Yang's avatar
Michael Yang committed
425
426
427
428
	return b.meta.KV()
}

func (b *Backend) Get(name string) ml.Tensor {
429
430
	if t, ok := b.tensors[name]; ok {
		return &Tensor{b: b, t: t}
Michael Yang's avatar
Michael Yang committed
431
432
433
434
435
436
	}

	return nil
}

func (b *Backend) NewContext() ml.Context {
Michael Yang's avatar
Michael Yang committed
437
	return b.NewContextSize(b.maxGraphNodes)
438
439
440
}

func (b *Backend) NewContextSize(n int) ml.Context {
Jesse Gross's avatar
Jesse Gross committed
441
442
443
444
	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
445
	return &Context{
446
447
		b:             b,
		maxGraphNodes: n,
448
		ctx: C.ggml_init(C.struct_ggml_init_params{
449
			mem_size: C.size_t(n)*C.ggml_tensor_overhead() + C.ggml_graph_overhead_custom(C.size_t(n), false),
450
451
			no_alloc: true,
		}),
Michael Yang's avatar
Michael Yang committed
452
453
454
	}
}

455
func (b *Backend) CacheConfig() ml.CacheConfig {
456
457
458
459
460
	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}
	}
461
462
}

Michael Yang's avatar
Michael Yang committed
463
type Context struct {
464
	b *Backend
Michael Yang's avatar
Michael Yang committed
465

466
	ctx   *C.struct_ggml_context
Michael Yang's avatar
Michael Yang committed
467
	graph *C.struct_ggml_cgraph
468

469
470
	// buft is the buffer type used for new tensors
	buft *C.struct_ggml_backend_buffer_type
471

Michael Yang's avatar
Michael Yang committed
472
	// maxGraphNodes is the maximum allowed number of graph nodes in this context
473
	maxGraphNodes int
Michael Yang's avatar
Michael Yang committed
474
475
}

Michael Yang's avatar
Michael Yang committed
476
477
func (c Context) Input() ml.Context {
	if c.b.input != nil {
478
479
480
		return &Context{
			b:             c.b,
			ctx:           c.ctx,
481
			buft:          c.b.input,
482
483
484
485
			maxGraphNodes: c.maxGraphNodes,
		}
	}

Michael Yang's avatar
Michael Yang committed
486
	return &c
487
488
}

Michael Yang's avatar
Michael Yang committed
489
func (c Context) Layer(i int) ml.Context {
490
	if buft, ok := c.b.layers[i]; ok {
491
492
493
		return &Context{
			b:             c.b,
			ctx:           c.ctx,
494
			buft:          buft,
495
496
497
498
			maxGraphNodes: c.maxGraphNodes,
		}
	}

Michael Yang's avatar
Michael Yang committed
499
	return &c
500
501
}

502
func (c *Context) Forward(tensors ...ml.Tensor) ml.Context {
Michael Yang's avatar
Michael Yang committed
503
	if c.graph == nil {
504
		c.graph = C.ggml_new_graph_custom(c.ctx, C.size_t(c.maxGraphNodes), false)
Michael Yang's avatar
Michael Yang committed
505
506
	}

507
508
509
510
511
	for _, tensor := range tensors {
		C.ggml_build_forward_expand(c.graph, tensor.(*Tensor).t)
	}

	return c
Michael Yang's avatar
Michael Yang committed
512
513
}

Michael Yang's avatar
Michael Yang committed
514
func (c Context) Compute(tensors ...ml.Tensor) {
515
	C.ggml_backend_sched_graph_compute_async(c.b.sched, c.graph)
Michael Yang's avatar
Michael Yang committed
516
	C.ggml_backend_sched_reset(c.b.sched)
Michael Yang's avatar
Michael Yang committed
517

518
519
520
	needSync := true
	sync := func() {
		if needSync {
521
			C.ggml_backend_sched_synchronize(c.b.sched)
522
523
524
			needSync = false
		}
	}
Michael Yang's avatar
Michael Yang committed
525

526
527
528
	for _, t := range tensors {
		if C.ggml_nbytes(t.(*Tensor).t) > 0 {
			t.(*Tensor).sync = sync
529
530
		}
	}
Michael Yang's avatar
Michael Yang committed
531
532
}

Michael Yang's avatar
Michael Yang committed
533
func (c Context) MaxGraphNodes() int {
534
	return c.maxGraphNodes
Jesse Gross's avatar
Jesse Gross committed
535
536
}

537
538
539
func shapeToGGML(shape []int) *C.int64_t {
	sh := make([]C.int64_t, len(shape))
	for i, s := range shape {
540
		sh[i] = C.int64_t(s)
541
542
543
544
545
	}

	return &sh[0]
}

546
547
548
549
func pad(length, pad C.size_t) C.size_t {
	return ((length + pad - 1) / pad) * pad
}

550
func (c Context) newTensor(dtype ml.DType, shape []int) ml.Tensor {
551
552
553
554
	if c.buft == nil {
		panic("set Input, Output, or Layer before creating tensors")
	}

Michael Yang's avatar
Michael Yang committed
555
556
557
558
559
560
	var cdtype uint32
	switch dtype {
	case ml.DTypeF32:
		cdtype = C.GGML_TYPE_F32
	case ml.DTypeF16:
		cdtype = C.GGML_TYPE_F16
561
562
563
564
	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
565
566
567
568
569
570
	case ml.DTypeI32:
		cdtype = C.GGML_TYPE_I32
	default:
		panic("unsupported dtype")
	}

Jesse Gross's avatar
Jesse Gross committed
571
	if len(shape) < 1 || shape[0] == 0 {
Michael Yang's avatar
Michael Yang committed
572
573
574
		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
575
576
577
578
579
580
581
582
583
		panic("unsupported number of dimensions")
	}

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

Michael Yang's avatar
Michael Yang committed
584
	t := C.ggml_new_tensor(c.ctx, cdtype, C.int(len(shape)), shapeToGGML(shape))
585
586
	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)
Michael Yang's avatar
Michael Yang committed
587
	C.ggml_backend_tensor_alloc(b, t, C.ggml_backend_buffer_get_base(b))
588
	return &Tensor{b: c.b, t: t}
589
590
591
}

func (c Context) Empty(dtype ml.DType, shape ...int) ml.Tensor {
592
	return c.newTensor(dtype, shape)
593
594
595
}

func (c Context) Zeros(dtype ml.DType, shape ...int) ml.Tensor {
596
	t := c.newTensor(dtype, shape)
597
598
	C.ggml_set_zero(t.(*Tensor).t)
	return t
Michael Yang's avatar
Michael Yang committed
599
600
}

601
func checkShape[S ~[]E, E any](s S, shape ...int) error {
Michael Yang's avatar
Michael Yang committed
602
	n := len(s)
Jesse Gross's avatar
Jesse Gross committed
603
604
605
606
607

	if n == 0 {
		return nil
	}

Michael Yang's avatar
Michael Yang committed
608
609
610
611
612
	for _, v := range shape {
		n /= v
	}

	if n != 1 {
613
		return fmt.Errorf("invalid shape: %v", shape)
Michael Yang's avatar
Michael Yang committed
614
615
	}

616
	return nil
Michael Yang's avatar
Michael Yang committed
617
618
619
}

func (c Context) FromFloatSlice(s []float32, shape ...int) (ml.Tensor, error) {
Jesse Gross's avatar
Jesse Gross committed
620
	if err := checkShape(s, shape...); err != nil {
621
622
623
624
		return nil, err
	}

	t := c.newTensor(ml.DTypeF32, shape)
Jesse Gross's avatar
Jesse Gross committed
625
626
627
628
	if len(s) > 0 {
		C.ggml_backend_tensor_set(t.(*Tensor).t, unsafe.Pointer(&s[0]), 0, C.ggml_nbytes(t.(*Tensor).t))
	}

629
	return t, nil
Michael Yang's avatar
Michael Yang committed
630
631
632
}

func (c Context) FromIntSlice(s []int32, shape ...int) (ml.Tensor, error) {
Jesse Gross's avatar
Jesse Gross committed
633
	if err := checkShape(s, shape...); err != nil {
634
635
636
637
		return nil, err
	}

	t := c.newTensor(ml.DTypeI32, shape)
Jesse Gross's avatar
Jesse Gross committed
638
639
640
641
	if len(s) > 0 {
		C.ggml_backend_tensor_set(t.(*Tensor).t, unsafe.Pointer(&s[0]), 0, C.ggml_nbytes(t.(*Tensor).t))
	}

642
	return t, nil
Michael Yang's avatar
Michael Yang committed
643
644
}

Michael Yang's avatar
Michael Yang committed
645
646
func (c *Context) Close() {
	if c != nil {
647
648
		C.ggml_free(c.ctx)
	}
Michael Yang's avatar
Michael Yang committed
649
650
651
}

type Tensor struct {
652
	b    *Backend
Michael Yang's avatar
Michael Yang committed
653
	t    *C.struct_ggml_tensor
654
	sync func()
Michael Yang's avatar
Michael Yang committed
655
656
657
658
659
660
661
662
663
664
}

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()),
	)
}

665
666
func (t *Tensor) Dim(n int) int {
	return int(t.t.ne[n])
Michael Yang's avatar
Michael Yang committed
667
668
}

669
670
func (t *Tensor) Stride(n int) int {
	return int(t.t.nb[n])
Michael Yang's avatar
Michael Yang committed
671
672
}

673
674
func (t *Tensor) Shape() []int {
	shape := make([]int, C.ggml_n_dims(t.t))
Michael Yang's avatar
Michael Yang committed
675
676
677
678
679
680
681
	for i := range shape {
		shape[i] = t.Dim(i)
	}

	return shape
}

682
683
684
685
686
687
688
689
690
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
691
692
}

693
694
695
696
697
698
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
699
700
701
702
703
704
705
706
707
	}

	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
708
709
	case C.GGML_TYPE_F16:
		return ml.DTypeF16
710
711
712
713
	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
714
715
716
717
718
719
720
	case C.GGML_TYPE_I32:
		return ml.DTypeI32
	default:
		return ml.DTypeOther
	}
}

721
722
723
724
725
726
727
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
728
729
func (t *Tensor) Add(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
	return &Tensor{
730
		b: t.b,
Michael Yang's avatar
Michael Yang committed
731
732
733
734
		t: C.ggml_add(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
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
756
757
758
759
760
761
762
763
764
765
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{
766
		b: t.b,
Michael Yang's avatar
Michael Yang committed
767
768
769
770
771
772
		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{
773
		b: t.b,
Michael Yang's avatar
Michael Yang committed
774
775
776
777
778
779
		t: C.ggml_cont(ctx.(*Context).ctx, t.t),
	}
}

func (t *Tensor) Mul(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
	return &Tensor{
780
		b: t.b,
Michael Yang's avatar
Michael Yang committed
781
782
783
784
785
786
		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{
787
		b: t.b,
Michael Yang's avatar
Michael Yang committed
788
789
790
791
		t: C.ggml_mul_mat(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

792
793
794
795
796
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{
797
		b: t.b,
798
799
800
801
		t: mul,
	}
}

Michael Yang's avatar
Michael Yang committed
802
func (t *Tensor) LayerNorm(ctx ml.Context, w, b ml.Tensor, eps float32) ml.Tensor {
803
	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
804
805
806
807
808
809
810
811
	if b != nil {
		tt = tt.Add(ctx, b)
	}

	return tt
}

func (t *Tensor) RMSNorm(ctx ml.Context, w ml.Tensor, eps float32) ml.Tensor {
812
	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
813
814
}

815
func (t *Tensor) Pad(ctx ml.Context, shape ...int) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
816
817
818
819
820
	if len(shape) != 4 {
		panic("expected 4 dimensions")
	}

	return &Tensor{
821
		b: t.b,
Michael Yang's avatar
Michael Yang committed
822
823
824
825
826
827
828
829
830
831
		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{
832
		b: t.b,
Michael Yang's avatar
Michael Yang committed
833
834
835
836
837
838
		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{
839
		b: t.b,
Michael Yang's avatar
Michael Yang committed
840
841
842
843
844
845
		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{
846
		b: t.b,
Michael Yang's avatar
Michael Yang committed
847
848
849
850
		t: C.ggml_cpy(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

851
func (t *Tensor) Reshape(ctx ml.Context, shape ...int) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
852
853
854
	switch len(shape) {
	case 1:
		return &Tensor{
855
			b: t.b,
Michael Yang's avatar
Michael Yang committed
856
857
858
859
			t: C.ggml_reshape_1d(ctx.(*Context).ctx, t.t, C.int64_t(shape[0])),
		}
	case 2:
		return &Tensor{
860
			b: t.b,
Michael Yang's avatar
Michael Yang committed
861
862
863
864
			t: C.ggml_reshape_2d(ctx.(*Context).ctx, t.t, C.int64_t(shape[0]), C.int64_t(shape[1])),
		}
	case 3:
		return &Tensor{
865
			b: t.b,
Michael Yang's avatar
Michael Yang committed
866
867
868
869
			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{
870
			b: t.b,
Michael Yang's avatar
Michael Yang committed
871
872
873
874
875
876
877
878
879
			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{
880
		b: t.b,
Michael Yang's avatar
Michael Yang committed
881
882
883
884
885
886
		t: C.ggml_scale(ctx.(*Context).ctx, t.t, (C.float)(s)),
	}
}

func (t *Tensor) Softmax(ctx ml.Context) ml.Tensor {
	return &Tensor{
887
		b: t.b,
Michael Yang's avatar
Michael Yang committed
888
889
890
891
		t: C.ggml_soft_max(ctx.(*Context).ctx, t.t),
	}
}

892
893
894
895
896
897
898
899
900
901
902
903
904
905
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
906
907
func (t *Tensor) Tanh(ctx ml.Context) ml.Tensor {
	return &Tensor{
908
		b: t.b,
Michael Yang's avatar
Michael Yang committed
909
910
911
912
		t: C.ggml_tanh_inplace(ctx.(*Context).ctx, t.t),
	}
}

913
func (t *Tensor) Unpad(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
		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{
928
			b: t.b,
Michael Yang's avatar
Michael Yang committed
929
930
931
932
			t: C.ggml_view_1d(ctx.(*Context).ctx, t.t, C.int64_t(shape[0]), C.size_t(offset)),
		}
	case 3:
		return &Tensor{
933
			b: t.b,
Michael Yang's avatar
Michael Yang committed
934
935
936
937
938
939
940
			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{
941
			b: t.b,
Michael Yang's avatar
Michael Yang committed
942
943
944
945
946
947
948
			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{
949
			b: t.b,
Michael Yang's avatar
Michael Yang committed
950
951
952
953
954
955
956
957
958
959
960
			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
961
962
963
964
	ropeTypeNorm   C.int = 0
	ropeTypeNeox   C.int = 2
	ropeTypeMrope  C.int = 8
	ropeTypeVision C.int = 24
Michael Yang's avatar
Michael Yang committed
965
966
)

Patrick Devine's avatar
Patrick Devine committed
967
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
968
	if ropeFactors == nil {
969
		ropeFactors = &Tensor{b: t.b}
Michael Yang's avatar
Michael Yang committed
970
971
	}

Jesse Gross's avatar
Jesse Gross committed
972
973
974
975
976
	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
977
	return &Tensor{
978
		b: t.b,
Michael Yang's avatar
Michael Yang committed
979
		t: C.ggml_rope_ext(
Jesse Gross's avatar
Jesse Gross committed
980
			ctx.(*Context).ctx, dequant, positionIDs.(*Tensor).t, ropeFactors.(*Tensor).t,
Michael Yang's avatar
Michael Yang committed
981
			C.int(ropeDim),
Patrick Devine's avatar
Patrick Devine committed
982
983
			C.int(ropeType),
			131072, // YaRN n_ctx_train
Michael Yang's avatar
Michael Yang committed
984
985
986
987
988
989
990
991
992
993
			C.float(ropeBase),
			C.float(ropeScale),
			0.,  // YaRN ext_factor
			1.,  // YaRN attn_factor
			32., // YaRN beta_fast
			1.,  // YaRN beta_slow
		),
	}
}

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

func (t *Tensor) SILU(ctx ml.Context) ml.Tensor {
	return &Tensor{
1010
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1011
1012
1013
1014
1015
1016
		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{
1017
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1018
1019
1020
		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)),
	}
}
1021

Michael Yang's avatar
Michael Yang committed
1022
func (t *Tensor) AvgPool2D(ctx ml.Context, k, s int, p float32) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
1023
1024
	return &Tensor{
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1025
		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
1026
1027
1028
	}
}

Michael Yang's avatar
Michael Yang committed
1029
1030
1031
1032
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
1033
		tt = C.ggml_set_1d(ctx.(*Context).ctx, t.t, t2.(*Tensor).t, C.size_t(offset))
Michael Yang's avatar
Michael Yang committed
1034
	case 1:
Michael Yang's avatar
Michael Yang committed
1035
		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
1036
1037
1038
1039
1040
1041
1042
	default:
		panic("unsupported number of dimensions")
	}

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

1043
1044
1045
1046
1047
1048
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
	}

1049
1050
1051
	query := t.Permute(ctx, 0, 2, 1, 3)
	key = key.Permute(ctx, 0, 2, 1, 3)

1052
1053
	if t.b.flashAttention {
		value = value.Permute(ctx, 0, 2, 1, 3)
1054

1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
		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)
	}
1068
}
1069
1070
1071
1072
1073
1074
1075

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