ggml.go 40.2 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"
Jesse Gross's avatar
Jesse Gross committed
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"
Jesse Gross's avatar
Jesse Gross committed
23
	"sync"
24
	"sync/atomic"
25
	"unicode"
Michael Yang's avatar
Michael Yang committed
26
27
28
	"unsafe"

	"github.com/ollama/ollama/format"
29
30
	"github.com/ollama/ollama/fs"
	fsggml "github.com/ollama/ollama/fs/ggml"
31
	"github.com/ollama/ollama/logutil"
Michael Yang's avatar
Michael Yang committed
32
	"github.com/ollama/ollama/ml"
33
	ggml "github.com/ollama/ollama/ml/backend/ggml/ggml/src"
34
	"github.com/ollama/ollama/ml/nn/rope"
Michael Yang's avatar
Michael Yang committed
35
36
37
	"golang.org/x/sync/errgroup"
)

Jesse Gross's avatar
Jesse Gross committed
38
39
40
41
42
43
var (
	cpus, accels, gpus []C.ggml_backend_dev_t
	backends           map[C.ggml_backend_dev_t]C.ggml_backend_t
)

var initDevices = sync.OnceFunc(func() {
Michael Yang's avatar
Michael Yang committed
44
45
	ggml.OnceLoad()

Jesse Gross's avatar
Jesse Gross committed
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
	backends = make(map[C.ggml_backend_dev_t]C.ggml_backend_t)
	for i := range C.ggml_backend_dev_count() {
		d := C.ggml_backend_dev_get(i)

		switch C.ggml_backend_dev_type(d) {
		case C.GGML_BACKEND_DEVICE_TYPE_CPU:
			if len(cpus) == 0 {
				// only the first cpu device should be used
				cpus = append(cpus, d)
			}
		case C.GGML_BACKEND_DEVICE_TYPE_ACCEL:
			accels = append(accels, d)
		case C.GGML_BACKEND_DEVICE_TYPE_GPU:
			gpus = append(gpus, d)
		}

		backends[d] = C.ggml_backend_dev_init(d, nil)
	}
})
Michael Yang's avatar
Michael Yang committed
65

Jesse Gross's avatar
Jesse Gross committed
66
67
68
69
70
type layerDevice struct {
	d  C.ggml_backend_dev_t
	bt C.ggml_backend_buffer_type_t
}

Michael Yang's avatar
Michael Yang committed
71
type Backend struct {
72
73
74
	// modelPath is the location of the model data
	modelPath string

75
76
	meta *fsggml.GGML

Jesse Gross's avatar
Jesse Gross committed
77
78
79
80
	// allocMemory means that memory should be allocated for tensors and not
	// just a dry run
	allocMemory bool

81
82
83
84
	// tensorLoadTargets maps from the name of the tensor in the file
	// to the name that is used by the model definition
	tensorLoadTargets map[string][]string

85
86
87
	sched         C.ggml_backend_sched_t
	schedBackends []C.ggml_backend_t
	schedBufts    []C.ggml_backend_buffer_type_t
88

89
	tensors map[string]*C.struct_ggml_tensor
Michael Yang's avatar
Michael Yang committed
90

Jesse Gross's avatar
Jesse Gross committed
91
	// input is the backend buffer type used for inputs
92
	input C.ggml_backend_buffer_type_t
Michael Yang's avatar
Michael Yang committed
93

Jesse Gross's avatar
Jesse Gross committed
94
95
96
	// output is the backend device used for outputs
	output C.ggml_backend_dev_t

Michael Yang's avatar
Michael Yang committed
97
	// layers is the backend used for repeating layers
Jesse Gross's avatar
Jesse Gross committed
98
	layers map[int]layerDevice
99

100
101
102
103
	// requiredMemory is the cumulative memory allocations needed by the backend
	requiredMemory *ml.BackendMemory

	// btDeviceMemory maps from a buffer type to the memory allocations associated with that device
104
	btDeviceMemory map[C.ggml_backend_buffer_type_t]*ml.DeviceMemory
105

106
	flashAttention bool
Michael Yang's avatar
Michael Yang committed
107
108
109

	// maxGraphNodes is the maximum allowed number of graph nodes in this scheduler
	maxGraphNodes int
Jesse Gross's avatar
Jesse Gross committed
110
111
112

	// weightBuffers are the GGML contexts and buffers for allocating weights
	weightBuffers map[*C.struct_ggml_context]C.ggml_backend_buffer_t
Michael Yang's avatar
Michael Yang committed
113
114
}

Jesse Gross's avatar
Jesse Gross committed
115
116
var once sync.Once

117
118
119
120
121
122
123
124
func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
	r, err := os.Open(modelPath)
	if err != nil {
		return nil, err
	}
	defer r.Close()

	meta, err := fsggml.Decode(r, -1)
Michael Yang's avatar
Michael Yang committed
125
126
127
128
	if err != nil {
		return nil, err
	}

Jesse Gross's avatar
Jesse Gross committed
129
130
131
132
133
134
135
136
137
138
139
	once.Do(func() {
		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()),
		)
	})
Michael Yang's avatar
Michael Yang committed
140

Jesse Gross's avatar
Jesse Gross committed
141
142
	initDevices()

143
	var requiredMemory ml.BackendMemory
144
	btDeviceMemory := make(map[C.ggml_backend_buffer_type_t]*ml.DeviceMemory)
145

146
	type deviceBufferType struct {
147
148
		d   C.ggml_backend_dev_t
		bts []C.ggml_backend_buffer_type_t
149
150
	}

151
152
	blocks := int(meta.KV().BlockCount())

Michael Yang's avatar
Michael Yang committed
153
	// create list of buffer types for the cpu
Michael Yang's avatar
Michael Yang committed
154
	cpuDeviceBufferType := deviceBufferType{d: C.ggml_backend_dev_by_type(C.GGML_BACKEND_DEVICE_TYPE_CPU)}
155
156
157
158
	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:
Jesse Gross's avatar
Jesse Gross committed
159
160
161
162
			bt := C.ggml_backend_dev_buffer_type(d)
			cpuDeviceBufferType.bts = append(cpuDeviceBufferType.bts, bt)
			C.ggml_backend_buft_set_alloc(bt, C.bool(params.AllocMemory))

163
			btDeviceMemory[C.ggml_backend_dev_buffer_type(d)] = &requiredMemory.CPU
Michael Yang's avatar
Michael Yang committed
164
		}
165
166
	}

167
	requiredMemory.CPU.Name = C.GoString(C.ggml_backend_dev_name(cpuDeviceBufferType.d))
168
169
	var props C.struct_ggml_backend_dev_props
	C.ggml_backend_dev_get_props(cpuDeviceBufferType.d, &props)
170
	requiredMemory.CPU.ID = C.GoString(props.id)
171
172
173
	requiredMemory.CPU.Weights = make([]ml.Memory, blocks+1)
	requiredMemory.CPU.Cache = make([]ml.Memory, blocks+1)

Michael Yang's avatar
Michael Yang committed
174
	// create list of buffer types for each gpu
175
	var gpuDeviceBufferTypes []deviceBufferType
176
177
	requiredMemory.GPUs = make([]ml.DeviceMemory, len(gpus))
	for i, d := range gpus {
178
		bt := C.ggml_backend_dev_buffer_type(d)
179
		gpuDeviceBufferTypes = append(gpuDeviceBufferTypes, deviceBufferType{
180
			d:   d,
181
			bts: append([]C.ggml_backend_buffer_type_t{bt}, cpuDeviceBufferType.bts...),
182
		})
Jesse Gross's avatar
Jesse Gross committed
183
184
		C.ggml_backend_buft_set_alloc(bt, C.bool(params.AllocMemory))

185
186
		btDeviceMemory[bt] = &requiredMemory.GPUs[i]
		requiredMemory.GPUs[i].Name = C.GoString(C.ggml_backend_dev_name(d))
187
188
		var props C.struct_ggml_backend_dev_props
		C.ggml_backend_dev_get_props(d, &props)
189
		requiredMemory.GPUs[i].ID = C.GoString(props.id)
190
191
		requiredMemory.GPUs[i].Weights = make([]ml.Memory, blocks+1)
		requiredMemory.GPUs[i].Cache = make([]ml.Memory, blocks+1)
Michael Yang's avatar
Michael Yang committed
192
193
	}

Michael Yang's avatar
Michael Yang committed
194
	// inputs always use cpu
Michael Yang's avatar
Michael Yang committed
195
	input := cpuDeviceBufferType
196

Jesse Gross's avatar
Jesse Gross committed
197
198
199
200
201
202
203
204
205
	assignLayer := func(layer int) deviceBufferType {
		for _, p := range params.GPULayers {
			for _, l := range p.Layers {
				if l == layer {
					for i := range requiredMemory.GPUs {
						if requiredMemory.GPUs[i].ID == p.ID {
							return gpuDeviceBufferTypes[i]
						}
					}
206

Jesse Gross's avatar
Jesse Gross committed
207
208
209
					return cpuDeviceBufferType
				}
			}
210
211
		}

Jesse Gross's avatar
Jesse Gross committed
212
		return cpuDeviceBufferType
213
214
	}

Michael Yang's avatar
Michael Yang committed
215
	// repeating layers are assigned based on their index in reverse order, e.g. i / (block_count + 1)
216
	layers := make([]deviceBufferType, blocks)
217
	for i := range layers {
218
		layers[i] = assignLayer(i)
219
220
	}

Michael Yang's avatar
Michael Yang committed
221
	// outputs are assigned iff allowed by splits and configured number of gpu layers
222
	output := assignLayer(blocks)
223
224
225

	maxTensors := len(meta.Tensors().Items())
	maxTensors += 1
Michael Yang's avatar
Michael Yang committed
226
	// each layer has at most 2 extra tensors for rope operations
227
228
	maxTensors += blocks * 2

229
	type tensor struct {
230
		source *fsggml.Tensor
231
232
233
		target string
	}

Michael Yang's avatar
Michael Yang committed
234
	// some tensors are mapped to different names so keep a list
235
236
	targets := make(map[string][]string)

Michael Yang's avatar
Michael Yang committed
237
	// contexts are shared by tensors of the same buffer type
238
239
	ctxs := make(map[C.ggml_backend_buffer_type_t]*C.struct_ggml_context)
	createTensor := func(t tensor, bts []C.ggml_backend_buffer_type_t, layer int) *C.struct_ggml_tensor {
240
241
242
243
244
245
246
		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
247

248
249
250
251
252
253
254
255
			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
256
			defer C.free(unsafe.Pointer(cname))
257
258
259
260
			if tt := C.ggml_get_tensor(ctxs[bt], cname); tt != nil {
				return tt
			}

261
262
263
264
265
266
267
268
269
270
			kind := t.source.Kind
			if t.source.Kind == 4 {
				// transform raw mxfp4 stream to ggml mxfp4 format
				kind = 39
			} else if t.source.Kind == uint32(fsggml.TensorTypeBF16) && strings.HasSuffix(t.source.Name, "_exps.bias") {
				// transform "_exps.bias" from bf16 to fp32; add_ids only supports fp32 tensors
				kind = uint32(fsggml.TensorTypeF32)
			}

			tt := C.ggml_new_tensor(ctxs[bt], kind, C.int(len(t.source.Shape)), (*C.int64_t)(unsafe.Pointer(&t.source.Shape[0])))
Michael Yang's avatar
Michael Yang committed
271
272
			C.ggml_set_name(tt, cname)

273
			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)))
274
275
276
277

			size := pad(C.ggml_backend_buft_get_alloc_size(bt, tt), C.ggml_backend_buft_get_alignment(bt))
			if layer == -1 {
				// Assume that InputWeights can be allocated - they're always in system memory and can't be moved in any case
Jesse Gross's avatar
Jesse Gross committed
278
279
280
				if params.AllocMemory {
					requiredMemory.InputWeights.Status = ml.Allocated
				}
281
282
283
284
285
				requiredMemory.InputWeights.Size += uint64(size)
			} else {
				btDeviceMemory[bt].Weights[layer].Size += uint64(size)
			}

286
287
288
289
290
			//nolint:staticcheck // TODO: check if buffer type supports this tensor
			return tt
		}

		return nil
Michael Yang's avatar
Michael Yang committed
291
292
	}

293
	contains := func(s string, parts ...string) bool {
294
295
296
297
298
299
300
301
		split := strings.Split(s, ".")
		for _, part := range parts {
			if slices.Contains(split, part) {
				return true
			}
		}

		return false
Michael Yang's avatar
Michael Yang committed
302
303
	}

304
305
	for _, t := range meta.Tensors().Items() {
		switch {
306
		case contains(t.Name, "position_embd", "token_embd", "token_norm_embd", "token_types"):
307
			createTensor(tensor{source: t}, input.bts, -1)
Michael Yang's avatar
Michael Yang committed
308
			if _, ok := meta.Tensors().GroupLayers()["output"]; !ok && t.Name == "token_embd.weight" {
309
				createTensor(tensor{source: t, target: "output.weight"}, output.bts, blocks)
Michael Yang's avatar
Michael Yang committed
310
			}
Michael Yang's avatar
Michael Yang committed
311
312
313
		case contains(t.Name, "cls", "output", "output_norm",
			"altup_proj", "altup_unembd_proj",
			"per_layer_token_embd", "per_layer_model_proj", "per_layer_proj_norm"):
314
			createTensor(tensor{source: t}, output.bts, blocks)
315
		case strings.HasPrefix(t.Name, "v.") || strings.HasPrefix(t.Name, "mm."):
Michael Yang's avatar
Michael Yang committed
316
			// TODO: assign vision tensors to the gpu if possible
317
			createTensor(tensor{source: t}, output.bts, blocks)
Michael Yang's avatar
Michael Yang committed
318
319
320
321
322
323
		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,
324
				}, layer.bts, i)
Michael Yang's avatar
Michael Yang committed
325
			}
326
		default:
Michael Yang's avatar
Michael Yang committed
327
328
329
330
			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
331
				}
Michael Yang's avatar
Michael Yang committed
332
			}
333

Michael Yang's avatar
Michael Yang committed
334
			if layerIndex >= 0 {
335
				createTensor(tensor{source: t}, layers[layerIndex].bts, layerIndex)
336
			} else {
Michael Yang's avatar
Michael Yang committed
337
				// load all other tensors on the cpu
338
				createTensor(tensor{source: t}, input.bts, -1)
339
340
341
			}
		}
	}
Michael Yang's avatar
Michael Yang committed
342

Michael Yang's avatar
Michael Yang committed
343
	// allocate buffers for each context
344
	bbs := make(map[*C.struct_ggml_context]C.ggml_backend_buffer_t, len(ctxs))
345
346
347
348
349
350
	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)
Jesse Gross's avatar
Jesse Gross committed
351
352
353
354
355
356
357
358
		if params.AllocMemory {
			for i := range btDeviceMemory[bt].Weights {
				if btDeviceMemory[bt].Weights[i].Size != 0 {
					if b != nil {
						btDeviceMemory[bt].Weights[i].Status = ml.Allocated
					} else {
						btDeviceMemory[bt].Weights[i].Status = ml.Failed
					}
359
360
361
362
				}
			}
		}

363
		if b == nil {
Jesse Gross's avatar
Jesse Gross committed
364
365
366
367
368
369
370
371
			for _, b := range bbs {
				C.ggml_backend_buffer_free(b)
			}

			for _, ctx := range ctxs {
				C.ggml_free(ctx)
			}

372
			panic(ml.ErrNoMem{BackendMemory: requiredMemory})
373
374
		}

375
		C.ggml_backend_buffer_set_usage(b, C.GGML_BACKEND_BUFFER_USAGE_WEIGHTS)
Michael Yang's avatar
Michael Yang committed
376
		bbs[c] = b
377
378
379
	}

	for bs := range maps.Values(bbs) {
Jesse Gross's avatar
Jesse Gross committed
380
381
		slog.Log(context.TODO(), logutil.LevelTrace, "model weights", "buffer", C.GoString(C.ggml_backend_buffer_name(bs)),
			"size", format.HumanBytes2(uint64(C.ggml_backend_buffer_get_size(bs))))
382
383
	}

Michael Yang's avatar
Michael Yang committed
384
	// map tensor names to tensors for easy lookup later
385
386
387
388
389
390
391
	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
		}
	}

392
	// map devices to backend buffer types so new tensors can be assigned to the correct device
393
	deviceBufferTypes := make(map[C.ggml_backend_dev_t]C.ggml_backend_buffer_type_t)
394
395

	// create backends and buffer types used for the compute graph scheduler
396
397
	var schedBackends []C.ggml_backend_t
	var schedBufts []C.ggml_backend_buffer_type_t
398
	for _, d := range append(gpus, append(accels, cpus...)...) {
Jesse Gross's avatar
Jesse Gross committed
399
		b := backends[d]
400
401
		bt := C.ggml_backend_get_default_buffer_type(b)

Jesse Gross's avatar
Jesse Gross committed
402
403
404
405
406
407
408
		// Always include CPU as a fallback but otherwise, just use the devices where we assigned layers
		if !slices.Contains(cpuDeviceBufferType.bts, bt) {
			if c, ok := ctxs[bt]; !ok || C.ggml_get_first_tensor(c) == nil {
				continue
			}
		}

409
410
411
412
413
414
415
416
417
418
419
420
421
422
		deviceBufferTypes[d] = bt

		schedBackends = append(schedBackends, b)
		schedBufts = append(schedBufts, bt)

		if C.ggml_backend_is_cpu(b) {
			// set number of threads for cpu backend
			C.ggml_backend_cpu_set_n_threads(b, C.int(Threads(params.NumThreads)))
		}
	}

	maxGraphNodes := max(8192, len(meta.Tensors().Items())*5)
	return &Backend{
		modelPath:         modelPath,
Jesse Gross's avatar
Jesse Gross committed
423
		allocMemory:       params.AllocMemory,
424
425
426
427
428
429
430
431
432
		flashAttention:    params.FlashAttention,
		meta:              meta,
		tensorLoadTargets: targets,
		tensors:           tensors,
		sched: C.ggml_backend_sched_new(
			(*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),
433
			C._Bool(false),
434
435
436
437
438
			C._Bool(false),
		),
		schedBackends: schedBackends,
		schedBufts:    schedBufts,
		input:         deviceBufferTypes[input.d],
Jesse Gross's avatar
Jesse Gross committed
439
440
441
		output:        output.d,
		layers: func() map[int]layerDevice {
			m := make(map[int]layerDevice)
442
			for i, layer := range layers {
Jesse Gross's avatar
Jesse Gross committed
443
444
445
446
				m[i] = layerDevice{
					d:  layer.d,
					bt: deviceBufferTypes[layer.d],
				}
447
448
449
			}
			return m
		}(),
450
451
452
		requiredMemory: &requiredMemory,
		btDeviceMemory: btDeviceMemory,
		maxGraphNodes:  maxGraphNodes,
Jesse Gross's avatar
Jesse Gross committed
453
		weightBuffers:  bbs,
454
455
456
457
458
459
460
	}, nil
}

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

Jesse Gross's avatar
Jesse Gross committed
461
462
463
464
465
466
467
468
469
470
471
472
473
func (b *Backend) Close() {
	if b == nil {
		return
	}

	for ctx, b := range b.weightBuffers {
		C.ggml_backend_buffer_free(b)
		C.ggml_free(ctx)
	}

	C.ggml_backend_sched_free(b.sched)
}

474
func (b *Backend) Load(ctx context.Context, progress func(float32)) error {
Jesse Gross's avatar
Jesse Gross committed
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
	if !b.allocMemory {
		return errors.New("cannot load model without memory allocation")
	}

	// Mimic llama runner logs summarizing layers and memory
	gpuLayers := 0
	for layer := range maps.Values(b.layers) {
		if C.ggml_backend_dev_type(layer.d) == C.GGML_BACKEND_DEVICE_TYPE_GPU {
			gpuLayers++
		}
	}
	slog.Info(fmt.Sprintf("offloading %d repeating layers to GPU", gpuLayers))

	switch C.ggml_backend_dev_type(b.output) {
	case C.GGML_BACKEND_DEVICE_TYPE_CPU:
		slog.Info("offloading output layer to CPU")
	case C.GGML_BACKEND_DEVICE_TYPE_GPU:
		slog.Info("offloading output layer to GPU")
		gpuLayers++
	case C.GGML_BACKEND_DEVICE_TYPE_ACCEL:
		slog.Info("offloading output layer to ACCEL")
	}
	slog.Info(fmt.Sprintf("offloaded %d/%d layers to GPU", gpuLayers, len(b.layers)+1))

499
	var doneBytes atomic.Uint64
500
	totalBytes := uint64(b.meta.Length) - b.meta.Tensors().Offset
501
502
503

	g, ctx := errgroup.WithContext(ctx)
	g.SetLimit(runtime.GOMAXPROCS(0))
504
	for _, t := range b.meta.Tensors().Items() {
505
		t := t
506
		g.Go(func() error {
507
			tts := make([]*C.struct_ggml_tensor, max(1, len(b.tensorLoadTargets[t.Name])))
508
			for i := range tts {
509
				target := b.tensorLoadTargets[t.Name][i]
510
511
512
				if target == "" {
					target = t.Name
				}
513

514
				tt, ok := b.tensors[target]
515
516
517
				if !ok {
					return fmt.Errorf("unassigned tensor: %s", t.Name)
				}
Michael Yang's avatar
Michael Yang committed
518

519
520
521
				tts[i] = tt
			}

522
523
			// 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.
524
			file, err := os.Open(b.modelPath)
525
			if err != nil {
526
				slog.Warn("file open error", "file", b.modelPath, "error", err)
527
528
529
				return err
			}
			defer file.Close()
530
			sr := io.NewSectionReader(file, int64(b.meta.Tensors().Offset+t.Offset), int64(t.Size()))
531
532
533
534
535
536
537

			if t.Kind == 4 && tts[0]._type == 39 {
				// source is mxfp4, target is ggml mxfp4

				const BS = 17                             // MXFP4 block size
				bts := make([]byte, 8*BS*format.KibiByte) // ~128k block aligned
				var s uint64
538
				var tmp [16]byte
539
540
541
542
543
544
545
546
547
548
549
550
				for s < t.Size() {
					// 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
					}
					n, err := io.ReadFull(sr, bts[:min(len(bts), int(t.Size()-s))])
					if err != nil {
						slog.Warn("file read error", "file", b.modelPath, "error", err)
						return err
					}
					for j := range n / BS {
						for i := 1; i < 9; i++ {
551
552
553
554
							// transform a1b2c3 ... x7y8z9 -> 71xa82yb93zc
							a, b := bts[j*BS+i], bts[j*BS+i+8]
							tmp[2*(i-1)] = (a & 0x0F) | (b << 4)
							tmp[2*(i-1)+1] = (a >> 4) | (b & 0xF0)
555
						}
556
						copy(bts[j*BS+1:j*BS+17], tmp[:])
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
					}

					for _, tt := range tts {
						C.ggml_backend_tensor_set(tt, unsafe.Pointer(&bts[0]), C.size_t(s), C.size_t(n))
					}

					s += uint64(n)

					if progress != nil {
						done := doneBytes.Add(uint64(n))
						progress(float32(done) / float32(totalBytes))
					}
				}
				return nil
			} else if strings.HasSuffix(t.Name, "_exps.bias") && t.Kind == 30 && tts[0]._type == 0 {
				// source is bf16, target is ggml fp32

				// data is bf16 but we need to convert to fp32
				bts := make([]byte, 128*format.KibiByte)
				var e uint64
				for e < t.Elements() {
					// 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
					}
					n, err := io.ReadFull(sr, bts[:min(len(bts), int(t.Elements()-e)*2)])
					if err != nil {
						slog.Warn("file read error", "file", b.modelPath, "error", err)
						return err
					}
					fp32 := ConvertToF32(bts, uint32(fsggml.TensorTypeBF16), uint64(n/2))

					for _, tt := range tts {
						C.ggml_backend_tensor_set(tt, unsafe.Pointer(&fp32[0]), C.size_t(e*4), C.size_t(n*2))
					}
					e += uint64(n / 2)
					if progress != nil {
						done := doneBytes.Add(uint64(n))
						progress(float32(done) / float32(totalBytes))
					}
				}
				return nil
			}

601
602
603
604
			bts := make([]byte, 128*format.KibiByte)

			var s uint64
			for s < t.Size() {
605
606
607
608
609
				// 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
				}

610
611
				n, err := io.ReadFull(sr, bts[:min(len(bts), int(t.Size()-s))])
				if err != nil {
612
					slog.Warn("file read error", "file", b.modelPath, "error", err)
613
					return err
614
				}
Michael Yang's avatar
Michael Yang committed
615

616
617
				for _, tt := range tts {
					C.ggml_backend_tensor_set(tt, unsafe.Pointer(&bts[0]), C.size_t(s), C.size_t(n))
618
				}
Michael Yang's avatar
Michael Yang committed
619

620
621
				s += uint64(n)

622
				if progress != nil {
623
					done := doneBytes.Add(uint64(n))
624
					progress(float32(done) / float32(totalBytes))
625
626
627
628
629
				}
			}

			return nil
		})
Michael Yang's avatar
Michael Yang committed
630
631
	}

632
633
634
635
636
637
638
639
640
641
642
643
	// Cleanup any backend state from devices that we didn't end up using
nextDevice:
	for _, d := range append(gpus, append(accels, cpus...)...) {
		for _, backend := range b.schedBackends {
			if d == C.ggml_backend_get_device(backend) {
				continue nextDevice
			}
		}

		C.ggml_backend_dev_reset(d)
	}

644
	if err := g.Wait(); err != nil {
645
		return err
646
647
	}

648
	return nil
Michael Yang's avatar
Michael Yang committed
649
650
}

651
652
653
654
func (b *Backend) BackendMemory() ml.BackendMemory {
	return *b.requiredMemory
}

655
func (b *Backend) Config() fs.Config {
Michael Yang's avatar
Michael Yang committed
656
657
658
659
	return b.meta.KV()
}

func (b *Backend) Get(name string) ml.Tensor {
660
661
	if t, ok := b.tensors[name]; ok {
		return &Tensor{b: b, t: t}
Michael Yang's avatar
Michael Yang committed
662
663
664
665
666
667
	}

	return nil
}

func (b *Backend) NewContext() ml.Context {
Michael Yang's avatar
Michael Yang committed
668
	return b.NewContextSize(b.maxGraphNodes)
669
670
671
}

func (b *Backend) NewContextSize(n int) ml.Context {
Jesse Gross's avatar
Jesse Gross committed
672
673
674
675
	if n > b.maxGraphNodes {
		panic(fmt.Errorf("requested number of graph nodes (%v) for new context exceeds maximum (%v)", n, b.maxGraphNodes))
	}

676
	var allocatedBuffers []C.ggml_backend_buffer_t
677

Michael Yang's avatar
Michael Yang committed
678
	return &Context{
679
680
		b:             b,
		maxGraphNodes: n,
681
		ctx: C.ggml_init(C.struct_ggml_init_params{
682
			mem_size: C.size_t(n)*C.ggml_tensor_overhead() + C.ggml_graph_overhead_custom(C.size_t(n), false),
683
684
			no_alloc: true,
		}),
685
		allocatedBuffers: &allocatedBuffers,
686
		layer:            -1,
Michael Yang's avatar
Michael Yang committed
687
688
689
	}
}

690
func (b *Backend) CacheConfig() ml.CacheConfig {
691
692
693
694
695
	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}
	}
696
697
}

Michael Yang's avatar
Michael Yang committed
698
type Context struct {
699
	b *Backend
Michael Yang's avatar
Michael Yang committed
700

701
	ctx   *C.struct_ggml_context
Michael Yang's avatar
Michael Yang committed
702
	graph *C.struct_ggml_cgraph
703

704
	// buft is the buffer type used for new tensors
705
	buft C.ggml_backend_buffer_type_t
706

707
708
	// allocatedBuffers are buffers for tensors that we have allocated in this context
	// so that we can free them when we close the context
709
	allocatedBuffers *[]C.ggml_backend_buffer_t
710

Michael Yang's avatar
Michael Yang committed
711
	// maxGraphNodes is the maximum allowed number of graph nodes in this context
712
	maxGraphNodes int
713
714
715

	// layer is the graph layer that this context is allocating for - assumed to be cache
	layer int
Michael Yang's avatar
Michael Yang committed
716
717
}

718
func (c *Context) Input() ml.Context {
Michael Yang's avatar
Michael Yang committed
719
	if c.b.input != nil {
720
		return &Context{
721
722
723
724
725
			b:                c.b,
			ctx:              c.ctx,
			buft:             c.b.input,
			allocatedBuffers: c.allocatedBuffers,
			maxGraphNodes:    c.maxGraphNodes,
726
			layer:            -1,
727
728
729
		}
	}

730
	return c
731
732
}

733
func (c *Context) Layer(i int) ml.Context {
Jesse Gross's avatar
Jesse Gross committed
734
	if layer, ok := c.b.layers[i]; ok {
735
		return &Context{
736
737
			b:                c.b,
			ctx:              c.ctx,
Jesse Gross's avatar
Jesse Gross committed
738
			buft:             layer.bt,
739
740
			allocatedBuffers: c.allocatedBuffers,
			maxGraphNodes:    c.maxGraphNodes,
741
			layer:            i,
742
743
744
		}
	}

745
	return c
746
747
}

748
func (c *Context) Forward(tensors ...ml.Tensor) ml.Context {
Michael Yang's avatar
Michael Yang committed
749
	if c.graph == nil {
750
		c.graph = C.ggml_new_graph_custom(c.ctx, C.size_t(c.maxGraphNodes), false)
Michael Yang's avatar
Michael Yang committed
751
752
	}

753
754
755
756
757
	for _, tensor := range tensors {
		C.ggml_build_forward_expand(c.graph, tensor.(*Tensor).t)
	}

	return c
Michael Yang's avatar
Michael Yang committed
758
759
}

760
func (c *Context) Compute(tensors ...ml.Tensor) {
761
762
763
	if status := C.ggml_backend_sched_graph_compute_async(c.b.sched, c.graph); status != C.GGML_STATUS_SUCCESS {
		panic(fmt.Errorf("error computing ggml graph: %v", status))
	}
Michael Yang's avatar
Michael Yang committed
764
	C.ggml_backend_sched_reset(c.b.sched)
Michael Yang's avatar
Michael Yang committed
765

766
767
768
	needSync := true
	sync := func() {
		if needSync {
769
			C.ggml_backend_sched_synchronize(c.b.sched)
770
771
772
			needSync = false
		}
	}
Michael Yang's avatar
Michael Yang committed
773

774
775
776
	for _, t := range tensors {
		if C.ggml_nbytes(t.(*Tensor).t) > 0 {
			t.(*Tensor).sync = sync
777
778
		}
	}
Michael Yang's avatar
Michael Yang committed
779
780
}

781
782
func (c *Context) Reserve() {
	reserved := C.ggml_backend_sched_reserve(c.b.sched, c.graph)
783
784

	slog.Debug("compute graph", "nodes", C.ggml_graph_n_nodes(c.graph), "splits", C.ggml_backend_sched_get_n_splits(c.b.sched))
785
786
787
788
789
790

	// Reserve may get called multiple times for different graphs - we just want the last run, which will contain the max allocations
	for _, bt := range c.b.schedBufts {
		c.b.btDeviceMemory[bt].Graph = ml.Memory{}
	}

791
	for i := range c.b.schedBackends {
792
793
794
795
		bufferStatus := C.ggml_backend_sched_get_attempted_buffer_size(c.b.sched, c.b.schedBackends[i])

		graph := &c.b.btDeviceMemory[c.b.schedBufts[i]].Graph
		graph.Size += uint64(bufferStatus.size)
Jesse Gross's avatar
Jesse Gross committed
796
797
798
799
800
801
		if c.b.allocMemory {
			if bufferStatus.allocated && graph.Status != ml.Failed {
				graph.Status = ml.Allocated
			} else {
				graph.Status = ml.Failed
			}
802
803
		}

Jesse Gross's avatar
Jesse Gross committed
804
805
		slog.Log(context.TODO(), logutil.LevelTrace, "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(bufferStatus.size)))
806
807
	}

808
809
810
	if !reserved {
		panic(ml.ErrNoMem{BackendMemory: *c.b.requiredMemory})
	}
811
812
}

813
func (c *Context) MaxGraphNodes() int {
814
	return c.maxGraphNodes
Jesse Gross's avatar
Jesse Gross committed
815
816
}

817
818
819
func shapeToGGML(shape []int) *C.int64_t {
	sh := make([]C.int64_t, len(shape))
	for i, s := range shape {
820
		sh[i] = C.int64_t(s)
821
822
823
824
825
	}

	return &sh[0]
}

826
827
828
829
func pad(length, pad C.size_t) C.size_t {
	return ((length + pad - 1) / pad) * pad
}

830
func (c *Context) newTensor(dtype ml.DType, shape []int) ml.Tensor {
831
	if c.buft == nil {
832
		panic("set Input or Layer before creating tensors")
833
834
	}

835
	cdtype := ggmlDType(dtype)
Michael Yang's avatar
Michael Yang committed
836

Jesse Gross's avatar
Jesse Gross committed
837
	if len(shape) < 1 || shape[0] == 0 {
Michael Yang's avatar
Michael Yang committed
838
		var shape C.int64_t = 0
839
		return &Tensor{b: c.b, t: C.ggml_new_tensor(c.ctx, cdtype, 1, &shape)}
Michael Yang's avatar
Michael Yang committed
840
	} else if len(shape) > 4 {
Michael Yang's avatar
Michael Yang committed
841
842
843
844
845
846
847
848
849
		panic("unsupported number of dimensions")
	}

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

Michael Yang's avatar
Michael Yang committed
850
	t := C.ggml_new_tensor(c.ctx, cdtype, C.int(len(shape)), shapeToGGML(shape))
851
	size := pad(C.ggml_backend_buft_get_alloc_size(c.buft, t), C.ggml_backend_buft_get_alignment(c.buft))
852

853
	b := C.ggml_backend_buft_alloc_buffer(c.buft, size)
854
855
856
857
	if c.layer >= 0 {
		cache := &c.b.btDeviceMemory[c.buft].Cache[c.layer]

		cache.Size += uint64(size)
Jesse Gross's avatar
Jesse Gross committed
858
859
860
861
862
863
		if c.b.allocMemory {
			if b != nil {
				cache.Status = ml.Allocated
			} else {
				cache.Status = ml.Failed
			}
864
865
866
		}
	}

867
	if b == nil {
868
		panic(ml.ErrNoMem{BackendMemory: *c.b.requiredMemory})
869
870
	}

871
	*c.allocatedBuffers = append(*c.allocatedBuffers, b)
Michael Yang's avatar
Michael Yang committed
872
	C.ggml_backend_tensor_alloc(b, t, C.ggml_backend_buffer_get_base(b))
873
	return &Tensor{b: c.b, t: t}
874
875
}

876
func (c *Context) Empty(dtype ml.DType, shape ...int) ml.Tensor {
877
	return c.newTensor(dtype, shape)
878
879
}

880
func (c *Context) Zeros(dtype ml.DType, shape ...int) ml.Tensor {
881
	t := c.newTensor(dtype, shape)
Jesse Gross's avatar
Jesse Gross committed
882
883
884
	if c.b.allocMemory {
		C.ggml_set_zero(t.(*Tensor).t)
	}
885
	return t
Michael Yang's avatar
Michael Yang committed
886
887
}

888
func checkShape[S ~[]E, E any](s S, shape ...int) {
Michael Yang's avatar
Michael Yang committed
889
	n := len(s)
Jesse Gross's avatar
Jesse Gross committed
890
891

	if n == 0 {
892
		return
Jesse Gross's avatar
Jesse Gross committed
893
894
	}

Michael Yang's avatar
Michael Yang committed
895
896
897
898
899
	for _, v := range shape {
		n /= v
	}

	if n != 1 {
900
		panic(fmt.Errorf("invalid shape: %v", shape))
Michael Yang's avatar
Michael Yang committed
901
902
903
	}
}

904
905
func (c *Context) FromFloatSlice(s []float32, shape ...int) ml.Tensor {
	checkShape(s, shape...)
906

907
	t := c.newTensor(ml.DTypeF32, shape)
908

Jesse Gross's avatar
Jesse Gross committed
909
	if c.b.allocMemory && len(s) > 0 {
Jesse Gross's avatar
Jesse Gross committed
910
911
912
		C.ggml_backend_tensor_set(t.(*Tensor).t, unsafe.Pointer(&s[0]), 0, C.ggml_nbytes(t.(*Tensor).t))
	}

913
	return t
Michael Yang's avatar
Michael Yang committed
914
915
}

916
917
func (c *Context) FromIntSlice(s []int32, shape ...int) ml.Tensor {
	checkShape(s, shape...)
918

919
	t := c.newTensor(ml.DTypeI32, shape)
920

Jesse Gross's avatar
Jesse Gross committed
921
	if c.b.allocMemory && len(s) > 0 {
Jesse Gross's avatar
Jesse Gross committed
922
923
924
		C.ggml_backend_tensor_set(t.(*Tensor).t, unsafe.Pointer(&s[0]), 0, C.ggml_nbytes(t.(*Tensor).t))
	}

925
	return t
Michael Yang's avatar
Michael Yang committed
926
927
}

Michael Yang's avatar
arange  
Michael Yang committed
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
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))
		}

943
		return c.Input().FromIntSlice(arange, len(arange))
Michael Yang's avatar
arange  
Michael Yang committed
944
945
946
947
948
	default:
		panic("unsupported dtype for arange")
	}
}

Michael Yang's avatar
Michael Yang committed
949
950
func (c *Context) Close() {
	if c != nil {
951
952
953
954
955
		for _, b := range *c.allocatedBuffers {
			C.ggml_backend_buffer_free(b)
		}
		*c.allocatedBuffers = nil

956
957
		C.ggml_free(c.ctx)
	}
Michael Yang's avatar
Michael Yang committed
958
959
960
}

type Tensor struct {
961
	b    *Backend
Michael Yang's avatar
Michael Yang committed
962
	t    *C.struct_ggml_tensor
963
	sync func()
Michael Yang's avatar
Michael Yang committed
964
965
966
967
968
969
970
971
972
973
}

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()),
	)
}

974
975
func (t *Tensor) Dim(n int) int {
	return int(t.t.ne[n])
Michael Yang's avatar
Michael Yang committed
976
977
}

978
979
func (t *Tensor) Stride(n int) int {
	return int(t.t.nb[n])
Michael Yang's avatar
Michael Yang committed
980
981
}

982
983
func (t *Tensor) Shape() []int {
	shape := make([]int, C.ggml_n_dims(t.t))
Michael Yang's avatar
Michael Yang committed
984
985
986
987
988
989
990
	for i := range shape {
		shape[i] = t.Dim(i)
	}

	return shape
}

991
992
993
994
995
996
997
998
999
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
1000
1001
}

1002
1003
1004
1005
1006
1007
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
1008
1009
1010
1011
1012
1013
1014
1015
1016
	}

	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
1017
1018
	case C.GGML_TYPE_F16:
		return ml.DTypeF16
1019
1020
1021
1022
	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
1023
1024
	case C.GGML_TYPE_I32:
		return ml.DTypeI32
Michael Yang's avatar
Michael Yang committed
1025
1026
	case C.GGML_TYPE_MXFP4:
		return ml.DTypeMXFP4
Michael Yang's avatar
Michael Yang committed
1027
1028
1029
1030
1031
	default:
		return ml.DTypeOther
	}
}

1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
func ggmlDType(dtype ml.DType) uint32 {
	switch dtype {
	case ml.DTypeF32:
		return C.GGML_TYPE_F32
	case ml.DTypeF16:
		return C.GGML_TYPE_F16
	case ml.DTypeQ80:
		return C.GGML_TYPE_Q8_0
	case ml.DTypeQ40:
		return C.GGML_TYPE_Q4_0
	case ml.DTypeI32:
		return C.GGML_TYPE_I32
	case ml.DTypeMXFP4:
		return C.GGML_TYPE_MXFP4
	default:
		panic("unsupported dtype")
	}
}

func (t *Tensor) Cast(ctx ml.Context, dtype ml.DType) ml.Tensor {
	return &Tensor{
		b: t.b,
		t: C.ggml_cast(ctx.(*Context).ctx, t.t, ggmlDType(dtype)),
	}
}

1058
1059
1060
1061
1062
1063
1064
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
1065
1066
func (t *Tensor) Add(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
	return &Tensor{
1067
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1068
1069
1070
1071
		t: C.ggml_add(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

Michael Yang's avatar
Michael Yang committed
1072
1073
1074
1075
1076
1077
1078
func (t *Tensor) Sub(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
	return &Tensor{
		b: t.b,
		t: C.ggml_sub(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
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
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
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{
1110
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1111
1112
1113
1114
		t: C.ggml_concat(ctx.(*Context).ctx, t.t, t2.(*Tensor).t, C.int(dim)),
	}
}

Michael Yang's avatar
Michael Yang committed
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
func (t *Tensor) Contiguous(ctx ml.Context, shape ...int) ml.Tensor {
	switch len(shape) {
	case 0:
		return &Tensor{
			b: t.b,
			t: C.ggml_cont(ctx.(*Context).ctx, t.t),
		}
	case 1:
		return &Tensor{
			b: t.b,
			t: C.ggml_cont_1d(ctx.(*Context).ctx, t.t, C.int64_t(shape[0])),
		}
	case 2:
		return &Tensor{
			b: t.b,
			t: C.ggml_cont_2d(ctx.(*Context).ctx, t.t, C.int64_t(shape[0]), C.int64_t(shape[1])),
		}
	case 3:
		return &Tensor{
			b: t.b,
			t: C.ggml_cont_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{
			b: t.b,
			t: C.ggml_cont_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")
Michael Yang's avatar
Michael Yang committed
1144
1145
1146
1147
1148
	}
}

func (t *Tensor) Mul(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
	return &Tensor{
1149
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1150
1151
1152
1153
		t: C.ggml_mul(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

1154
1155
1156
1157
1158
1159
1160
func (t *Tensor) Div(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
	return &Tensor{
		b: t.b,
		t: C.ggml_div(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

Michael Yang's avatar
Michael Yang committed
1161
1162
func (t *Tensor) Mulmat(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
	return &Tensor{
1163
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1164
1165
1166
1167
		t: C.ggml_mul_mat(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

1168
1169
1170
1171
1172
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{
1173
		b: t.b,
1174
1175
1176
1177
		t: mul,
	}
}

Michael Yang's avatar
llama4  
Michael Yang committed
1178
1179
1180
1181
1182
1183
1184
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),
	}
}

1185
1186
1187
1188
1189
1190
1191
func (t *Tensor) AddID(ctx ml.Context, t2, ids ml.Tensor) ml.Tensor {
	return &Tensor{
		b: t.b,
		t: C.ggml_add_id(ctx.(*Context).ctx, t.t, t2.(*Tensor).t, ids.(*Tensor).t),
	}
}

Michael Yang's avatar
Michael Yang committed
1192
func (t *Tensor) LayerNorm(ctx ml.Context, w, b ml.Tensor, eps float32) ml.Tensor {
Michael Yang's avatar
llama4  
Michael Yang committed
1193
1194
1195
1196
1197
1198
	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
1199
1200
	}

Michael Yang's avatar
llama4  
Michael Yang committed
1201
	return &Tensor{b: t.b, t: tt}
Michael Yang's avatar
Michael Yang committed
1202
1203
1204
}

func (t *Tensor) RMSNorm(ctx ml.Context, w ml.Tensor, eps float32) ml.Tensor {
Michael Yang's avatar
llama4  
Michael Yang committed
1205
1206
1207
1208
1209
1210
	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
1211
1212
}

1213
func (t *Tensor) Pad(ctx ml.Context, shape ...int) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
1214
1215
	if len(shape) != 4 {
		panic("expected 4 dimensions")
1216
1217
	} else if shape[3] != 0 {
		panic("cuda does not support 4d tensors")
Michael Yang's avatar
Michael Yang committed
1218
1219
1220
	}

	return &Tensor{
1221
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
		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{
1232
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1233
1234
1235
1236
1237
1238
		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{
1239
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1240
1241
1242
1243
1244
1245
		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{
1246
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1247
1248
1249
1250
		t: C.ggml_cpy(ctx.(*Context).ctx, t.t, t2.(*Tensor).t),
	}
}

1251
func (t *Tensor) Reshape(ctx ml.Context, shape ...int) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
1252
1253
1254
	switch len(shape) {
	case 1:
		return &Tensor{
1255
			b: t.b,
Michael Yang's avatar
Michael Yang committed
1256
1257
1258
1259
			t: C.ggml_reshape_1d(ctx.(*Context).ctx, t.t, C.int64_t(shape[0])),
		}
	case 2:
		return &Tensor{
1260
			b: t.b,
Michael Yang's avatar
Michael Yang committed
1261
1262
1263
1264
			t: C.ggml_reshape_2d(ctx.(*Context).ctx, t.t, C.int64_t(shape[0]), C.int64_t(shape[1])),
		}
	case 3:
		return &Tensor{
1265
			b: t.b,
Michael Yang's avatar
Michael Yang committed
1266
1267
1268
1269
			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{
1270
			b: t.b,
Michael Yang's avatar
Michael Yang committed
1271
1272
1273
1274
1275
1276
1277
1278
1279
			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{
1280
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1281
1282
1283
1284
		t: C.ggml_scale(ctx.(*Context).ctx, t.t, (C.float)(s)),
	}
}

1285
1286
1287
1288
1289
1290
1291
func (t *Tensor) SumRows(ctx ml.Context) ml.Tensor {
	return &Tensor{
		b: t.b,
		t: C.ggml_sum_rows(ctx.(*Context).ctx, t.t),
	}
}

Michael Yang's avatar
Michael Yang committed
1292
1293
func (t *Tensor) Softmax(ctx ml.Context) ml.Tensor {
	return &Tensor{
1294
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1295
1296
1297
1298
		t: C.ggml_soft_max(ctx.(*Context).ctx, t.t),
	}
}

1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
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
1313
1314
func (t *Tensor) Tanh(ctx ml.Context) ml.Tensor {
	return &Tensor{
1315
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1316
1317
1318
1319
		t: C.ggml_tanh_inplace(ctx.(*Context).ctx, t.t),
	}
}

Michael Yang's avatar
llama4  
Michael Yang committed
1320
1321
1322
1323
1324
1325
1326
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
1327
1328
1329
1330
func (t *Tensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
	switch len(shape) {
	case 1:
		return &Tensor{
1331
			b: t.b,
Michael Yang's avatar
Michael Yang committed
1332
1333
1334
1335
			t: C.ggml_view_1d(ctx.(*Context).ctx, t.t, C.int64_t(shape[0]), C.size_t(offset)),
		}
	case 3:
		return &Tensor{
1336
			b: t.b,
Michael Yang's avatar
Michael Yang committed
1337
1338
1339
1340
1341
1342
1343
			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{
1344
			b: t.b,
Michael Yang's avatar
Michael Yang committed
1345
1346
1347
1348
1349
1350
1351
			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{
1352
			b: t.b,
Michael Yang's avatar
Michael Yang committed
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
			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")
	}
}

1363
func (t *Tensor) RoPE(ctx ml.Context, positions ml.Tensor, ropeDim int, ropeBase, ropeScale float32, options ...func(*rope.Options)) ml.Tensor {
1364
	// Default options
Michael Yang's avatar
Michael Yang committed
1365
1366
1367
1368
1369
1370
1371
1372
	opts := rope.Options{
		Factors:               &Tensor{},
		OriginalContextLength: 131072,
		ExtrapolationFactor:   0.,
		AttentionFactor:       1.,
		BetaFast:              32.,
		BetaSlow:              1.,
	}
1373
1374
1375

	// Apply any provided options
	for _, option := range options {
Michael Yang's avatar
Michael Yang committed
1376
		option(&opts)
1377
1378
	}

Jesse Gross's avatar
Jesse Gross committed
1379
1380
1381
1382
1383
	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
1384
	return &Tensor{
1385
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1386
		t: C.ggml_rope_ext(
1387
1388
			ctx.(*Context).ctx,
			dequant,
1389
1390
			positions.(*Tensor).t,
			opts.Factors.(*Tensor).t,
Michael Yang's avatar
Michael Yang committed
1391
			C.int(ropeDim),
1392
1393
			C.int(opts.Type),
			C.int(opts.OriginalContextLength),
Michael Yang's avatar
Michael Yang committed
1394
1395
			C.float(ropeBase),
			C.float(ropeScale),
Michael Yang's avatar
Michael Yang committed
1396
1397
1398
1399
			C.float(opts.ExtrapolationFactor),
			C.float(opts.AttentionFactor),
			C.float(opts.BetaFast),
			C.float(opts.BetaSlow),
Michael Yang's avatar
Michael Yang committed
1400
1401
1402
1403
		),
	}
}

1404
1405
1406
1407
1408
1409
1410
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
1411
1412
func (t *Tensor) GELU(ctx ml.Context) ml.Tensor {
	return &Tensor{
1413
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1414
1415
1416
1417
		t: C.ggml_gelu_inplace(ctx.(*Context).ctx, t.t),
	}
}

Michael Yang's avatar
Michael Yang committed
1418
1419
1420
1421
1422
1423
1424
func (t *Tensor) QuickGELU(ctx ml.Context) ml.Tensor {
	return &Tensor{
		b: t.b,
		t: C.ggml_gelu_quick_inplace(ctx.(*Context).ctx, t.t),
	}
}

Michael Yang's avatar
Michael Yang committed
1425
1426
func (t *Tensor) SILU(ctx ml.Context) ml.Tensor {
	return &Tensor{
1427
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1428
1429
1430
1431
		t: C.ggml_silu_inplace(ctx.(*Context).ctx, t.t),
	}
}

Michael Yang's avatar
Michael Yang committed
1432
1433
1434
1435
1436
1437
1438
func (t *Tensor) RELU(ctx ml.Context) ml.Tensor {
	return &Tensor{
		b: t.b,
		t: C.ggml_relu_inplace(ctx.(*Context).ctx, t.t),
	}
}

1439
1440
1441
1442
1443
1444
1445
func (t *Tensor) SwiGLU(ctx ml.Context, up ml.Tensor, alpha, limit float32) ml.Tensor {
	return &Tensor{
		b: t.b,
		t: C.ggml_swiglu_oai(ctx.(*Context).ctx, t.t, up.(*Tensor).t, C.float(alpha), C.float(limit)),
	}
}

Michael Yang's avatar
Michael Yang committed
1446
1447
func (t *Tensor) Conv2D(ctx ml.Context, t2 ml.Tensor, s0, s1, p0, p1, d0, d1 int) ml.Tensor {
	return &Tensor{
1448
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1449
1450
1451
		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)),
	}
}
1452

Michael Yang's avatar
Michael Yang committed
1453
func (t *Tensor) AvgPool2D(ctx ml.Context, k, s int, p float32) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
1454
1455
	return &Tensor{
		b: t.b,
Michael Yang's avatar
Michael Yang committed
1456
		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
1457
1458
1459
	}
}

Michael Yang's avatar
Michael Yang committed
1460
1461
1462
1463
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
1464
		tt = C.ggml_set_1d(ctx.(*Context).ctx, t.t, t2.(*Tensor).t, C.size_t(offset))
Michael Yang's avatar
Michael Yang committed
1465
	case 1:
Michael Yang's avatar
Michael Yang committed
1466
		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
1467
1468
1469
1470
1471
1472
1473
	default:
		panic("unsupported number of dimensions")
	}

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

1474
func (t *Tensor) ScaledDotProductAttention(ctx ml.Context, key, value, mask, sinks ml.Tensor, scale float64) ml.Tensor {
1475
1476
1477
1478
1479
	var kqMask *C.struct_ggml_tensor
	if mask != nil {
		kqMask = mask.(*Tensor).t
	}

1480
1481
1482
	query := t.Permute(ctx, 0, 2, 1, 3)
	key = key.Permute(ctx, 0, 2, 1, 3)

1483
1484
	if t.b.flashAttention {
		value = value.Permute(ctx, 0, 2, 1, 3)
1485

1486
		kqv := C.ggml_flash_attn_ext(ctx.(*Context).ctx, query.(*Tensor).t, key.(*Tensor).t, value.(*Tensor).t, kqMask, C.float(scale), 0, 0)
1487
1488
1489
		if sinks != nil {
			C.ggml_flash_attn_ext_add_sinks(kqv, sinks.(*Tensor).t)
		}
1490
1491
1492
1493
1494
1495
1496
1497
		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),
		}
1498
1499
1500
		if sinks != nil {
			C.ggml_soft_max_add_sinks(kq.(*Tensor).t, sinks.(*Tensor).t)
		}
1501
1502
1503
1504

		kqv := value.Mulmat(ctx, kq)
		return kqv.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
	}
1505
}
1506
1507
1508
1509
1510
1511
1512

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
1513
1514
1515
1516
1517
1518
1519

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)),
	}
}
1520
1521
1522
1523
1524
1525
1526

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),
	}
}
Michael Yang's avatar
Michael Yang committed
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565

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

func (t *Tensor) Variance(ctx ml.Context) ml.Tensor {
	return t.Add(ctx, t.Mean(ctx).Scale(ctx, -1)).
		Sqr(ctx).
		SumRows(ctx).
		Scale(ctx, 1/float64(t.Dim(0)))
}

func (t *Tensor) Stddev(ctx ml.Context) ml.Tensor {
	return t.Variance(ctx).Sqrt(ctx)
}

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

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

func (t *Tensor) Clamp(ctx ml.Context, min, max float32) ml.Tensor {
	return &Tensor{
		b: t.b,
		t: C.ggml_clamp(ctx.(*Context).ctx, t.t, C.float(min), C.float(max)),
	}
}
Michael Yang's avatar
Michael Yang committed
1566
1567
1568
1569

func (c Context) FromBytes(dtype ml.DType, s []uint8, shape ...int) ml.Tensor {
	// Unchecked to handle quantized types
	t := c.newTensor(dtype, shape)
Jesse Gross's avatar
Jesse Gross committed
1570
	if c.b.allocMemory && len(s) > 0 {
Michael Yang's avatar
Michael Yang committed
1571
1572
1573
1574
1575
		C.ggml_backend_tensor_set(t.(*Tensor).t, unsafe.Pointer(&s[0]), 0, C.ggml_nbytes(t.(*Tensor).t))
	}

	return t
}