backend.go 9.39 KB
Newer Older
Michael Yang's avatar
Michael Yang committed
1
2
3
4
package ml

import (
	"bytes"
5
	"context"
Michael Yang's avatar
Michael Yang committed
6
7
	"encoding/binary"
	"fmt"
8
	"math"
Michael Yang's avatar
Michael Yang committed
9
	"slices"
Michael Yang's avatar
Michael Yang committed
10
11
12
	"strconv"
	"strings"

13
14
	"github.com/ollama/ollama/fs"
)
Michael Yang's avatar
Michael Yang committed
15
16

type Backend interface {
17
	Load(ctx context.Context, progress func(float32)) error
18
	Config() fs.Config
Michael Yang's avatar
Michael Yang committed
19
20
	Get(name string) Tensor
	NewContext() Context
21
	NewContextSize(size int) Context
Michael Yang's avatar
Michael Yang committed
22
23
}

24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
// BackendCacheConfig should be implemented by backends that need special output
// from the cache to meet specific requirements. It is frequently implemented in
// conjunction with ScaledDotProductAttention.
type BackendCacheConfig interface {
	CacheConfig() CacheConfig
}

// CacheConfig controls optimizations (mostly backend-specific) that may transform
// the output the cache to work better with specific kernels.
type CacheConfig struct {
	// CachePadding specifies the multiple for the number of tokens of cache history
	// that will be returned from cache Get for k, v and mask. The capacity of the
	// cache itself will also be increased to a multiple of this size if needed.
	CachePadding int

	// PermutedV performs Permute(ctx, 1, 2, 0, 3) on v tensors stored via Put
	// and return the permuted version via Get. This uses the cache copy operation
	// to avoid a Contiguous call on the permuted tensor.
	PermutedV bool
43
44
45
46
47
48
49
50

	// MaskDType specifies the data type for generating the mask. If unset it will
	// default to DTypeF32.
	MaskDType DType

	// MaskBatchPadding specifies the multiple for the batch size dimension in the mask.
	// Any position that does not correspond to an actual token will be filled with -Inf.
	MaskBatchPadding int
51
52
}

53
54
55
56
// BackendParams controls how the backend loads and executes models
type BackendParams struct {
	// NumThreads sets the number of threads to use if running on the CPU
	NumThreads int
Michael Yang's avatar
Michael Yang committed
57

58
59
60
61
62
63
64
65
	// MainGPU is the index of the primary GPU to use
	MainGPU int

	// NumGPULayers is the number of layers to offload to GPUs
	NumGPULayers int

	// TensorSplit is the fraction of the model to offload to each GPU
	TensorSplit []float32
66
67
68

	// FlashAttention indicates that we should use a fused flash attention kernel
	FlashAttention bool
69
70
}

71
var backends = make(map[string]func(string, BackendParams) (Backend, error))
72

73
func RegisterBackend(name string, f func(string, BackendParams) (Backend, error)) {
Michael Yang's avatar
Michael Yang committed
74
75
76
77
78
79
80
	if _, ok := backends[name]; ok {
		panic("backend: backend already registered")
	}

	backends[name] = f
}

81
func NewBackend(modelPath string, params BackendParams) (Backend, error) {
Michael Yang's avatar
Michael Yang committed
82
	if backend, ok := backends["ggml"]; ok {
83
		return backend(modelPath, params)
Michael Yang's avatar
Michael Yang committed
84
85
86
87
88
89
	}

	return nil, fmt.Errorf("unsupported backend")
}

type Context interface {
90
	Empty(dtype DType, shape ...int) Tensor
Michael Yang's avatar
Michael Yang committed
91
92
93
94
	Zeros(dtype DType, shape ...int) Tensor
	FromFloatSlice(s []float32, shape ...int) (Tensor, error)
	FromIntSlice(s []int32, shape ...int) (Tensor, error)

Michael Yang's avatar
arange  
Michael Yang committed
95
96
97
	// Arange creates a 1D tensor with values within an interval (start, stop] increased by step.
	Arange(start, stop, step float32, dtype DType) Tensor

98
	Forward(...Tensor) Context
99
	Compute(...Tensor)
100
101
102
103
104
105
106

	// Reserve is analogous to Compute but rather than executing a
	// graph, simply preallocates memory. Typically called with a
	// worst case graph to ensure all resources are available for
	// for future inference.
	Reserve() error

107
	MaxGraphNodes() int
108
	Close()
109

110
111
	// Input returns a context appropriate for creating tensors that are
	// inputs to the model (which includes things like output locations)
112
113
114
115
	Input() Context

	// Layer returns a context appropriate for creating intermediate tensors
	Layer(int) Context
Michael Yang's avatar
Michael Yang committed
116
117
}

118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
// RopeOptions contains optional parameters for RoPE function
type RopeOptions struct {
	OriginalContextLen uint32
}

// RopeOption defines a function that modifies RopeOpts
type RopeOption func(*RopeOptions)

// WithContextLen sets a custom context length
func WithContextLen(len uint32) RopeOption {
	return func(opts *RopeOptions) {
		opts.OriginalContextLen = len
	}
}

Michael Yang's avatar
Michael Yang committed
133
type Tensor interface {
134
135
	Dim(n int) int
	Stride(n int) int
Michael Yang's avatar
Michael Yang committed
136

137
	Shape() []int
Michael Yang's avatar
Michael Yang committed
138
139
140
141
142
	DType() DType

	Bytes() []byte
	Floats() []float32

143
	Neg(ctx Context) Tensor
Michael Yang's avatar
Michael Yang committed
144
145
146
	Add(ctx Context, t2 Tensor) Tensor
	Mul(ctx Context, t2 Tensor) Tensor
	Mulmat(ctx Context, t2 Tensor) Tensor
147
	MulmatFullPrec(ctx Context, t2 Tensor) Tensor
Michael Yang's avatar
llama4  
Michael Yang committed
148
	MulmatID(ctx Context, t2, ids Tensor) Tensor
Michael Yang's avatar
Michael Yang committed
149
150
151
152
153
154

	Softmax(ctx Context) Tensor
	LayerNorm(ctx Context, weight, bias Tensor, eps float32) Tensor
	RMSNorm(ctx Context, weight Tensor, eps float32) Tensor
	Scale(ctx Context, s float64) Tensor

Michael Yang's avatar
Michael Yang committed
155
	AvgPool2D(ctx Context, k, s int, p float32) Tensor
Michael Yang's avatar
Michael Yang committed
156
	Conv2D(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
Michael Yang's avatar
Michael Yang committed
157

158
	RoPE(ctx Context, positionIDs, ropeFactors Tensor, dim, ropeType uint32, base, scale float32, options ...RopeOption) Tensor
159
	IM2Col(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
Michael Yang's avatar
Michael Yang committed
160

161
162
	Sin(ctx Context) Tensor
	Cos(ctx Context) Tensor
Michael Yang's avatar
Michael Yang committed
163
164
165
	Tanh(ctx Context) Tensor
	GELU(ctx Context) Tensor
	SILU(ctx Context) Tensor
Michael Yang's avatar
llama4  
Michael Yang committed
166
	Sigmoid(ctx Context) Tensor
Michael Yang's avatar
Michael Yang committed
167

168
	Reshape(ctx Context, shape ...int) Tensor
Michael Yang's avatar
Michael Yang committed
169
170
171
	View(ctx Context, offset int, shape ...int) Tensor
	Permute(ctx Context, shape ...int) Tensor
	Contiguous(ctx Context) Tensor
Michael Yang's avatar
Michael Yang committed
172
	Set(ctx Context, t2 Tensor, offset int, strides ...int) Tensor
Michael Yang's avatar
Michael Yang committed
173

174
	Pad(ctx Context, shape ...int) Tensor
Michael Yang's avatar
Michael Yang committed
175
176

	Stack(ctx Context, dim int, s ...Tensor) Tensor
177
178
179

	// Repeat repeats the tensor n times along dimension dim
	Repeat(ctx Context, dim, n int) Tensor
Michael Yang's avatar
Michael Yang committed
180
181
182
	Concat(ctx Context, t2 Tensor, dim int) Tensor
	Rows(ctx Context, t2 Tensor) Tensor
	Copy(ctx Context, t2 Tensor) Tensor
183
	Duplicate(ctx Context) Tensor
Michael Yang's avatar
llama4  
Michael Yang committed
184
185

	TopK(ctx Context, k int) Tensor
186
	Argsort(ctx Context) Tensor
Michael Yang's avatar
Michael Yang committed
187
188
}

189
190
191
192
// ScaledDotProductAttention implements a fused attention
// operation equivalent to following code on a tensor named
// query:
//
193
194
195
196
// query = query.Permute(ctx, 0, 2, 1, 3)
// key = key.Permute(ctx, 0, 2, 1, 3)
// value = value.Permute(ctx, 1, 2, 0, 3).Contiguous(ctx)
//
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
// kq := key.MulmatFullPrec(ctx, query)
//
// kq = kq.Scale(ctx, scale)
//
//	if mask != nil {
//		kq = kq.Add(ctx, mask)
//	}
//
// kq = kq.Softmax(ctx)
//
// kqv := value.Mulmat(ctx, kq)
// return kqv.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
type ScaledDotProductAttention interface {
	ScaledDotProductAttention(ctx Context, key, value, mask Tensor, scale float64) Tensor
}

Michael Yang's avatar
Michael Yang committed
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
type number interface {
	~int | ~int8 | ~int16 | ~int32 | ~int64 |
		~uint | ~uint8 | ~uint16 | ~uint32 | ~uint64 |
		~float32 | ~float64 |
		~complex64 | ~complex128
}

func mul[T number](s ...T) T {
	p := T(1)
	for _, v := range s {
		p *= v
	}

	return p
}

229
type DumpOptions func(*dumpOptions)
Michael Yang's avatar
Michael Yang committed
230

231
232
233
234
235
// DumpWithPrecision sets the number of decimal places to print. Applies to float32 and float64.
func DumpWithPrecision(n int) DumpOptions {
	return func(opts *dumpOptions) {
		opts.Precision = n
	}
Michael Yang's avatar
Michael Yang committed
236
237
}

238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
// DumpWithThreshold sets the threshold for printing the entire tensor. If the number of elements
// is less than or equal to this value, the entire tensor will be printed. Otherwise, only the
// beginning and end of each dimension will be printed.
func DumpWithThreshold(n int) DumpOptions {
	return func(opts *dumpOptions) {
		opts.Threshold = n
	}
}

// DumpWithEdgeItems sets the number of elements to print at the beginning and end of each dimension.
func DumpWithEdgeItems(n int) DumpOptions {
	return func(opts *dumpOptions) {
		opts.EdgeItems = n
	}
}

type dumpOptions struct {
	Precision, Threshold, EdgeItems int
}

func Dump(ctx Context, t Tensor, optsFuncs ...DumpOptions) string {
	opts := dumpOptions{Precision: 4, Threshold: 1000, EdgeItems: 3}
	for _, optsFunc := range optsFuncs {
		optsFunc(&opts)
	}

	if mul(t.Shape()...) <= opts.Threshold {
		opts.EdgeItems = math.MaxInt
Michael Yang's avatar
Michael Yang committed
266
267
268
269
	}

	switch t.DType() {
	case DTypeF32:
270
271
		return dump[[]float32](ctx, t, opts.EdgeItems, func(f float32) string {
			return strconv.FormatFloat(float64(f), 'f', opts.Precision, 32)
Jesse Gross's avatar
Jesse Gross committed
272
		})
273
	case DTypeF16, DTypeQ80, DTypeQ40:
274
		f32 := ctx.Input().Empty(DTypeF32, t.Shape()...)
Jesse Gross's avatar
Jesse Gross committed
275
		f32 = t.Copy(ctx, f32)
276
277
		return dump[[]float32](ctx, f32, opts.EdgeItems, func(f float32) string {
			return strconv.FormatFloat(float64(f), 'f', opts.Precision, 32)
Michael Yang's avatar
Michael Yang committed
278
279
		})
	case DTypeI32:
280
		return dump[[]int32](ctx, t, opts.EdgeItems, func(i int32) string {
Michael Yang's avatar
Michael Yang committed
281
282
283
284
285
286
287
			return strconv.FormatInt(int64(i), 10)
		})
	default:
		return "<unsupported>"
	}
}

Jesse Gross's avatar
Jesse Gross committed
288
289
func dump[S ~[]E, E number](ctx Context, t Tensor, items int, fn func(E) string) string {
	if t.Bytes() == nil {
290
		ctx.Forward(t).Compute(t)
Michael Yang's avatar
Michael Yang committed
291
292
293
294
295
296
297
298
	}

	s := make(S, mul(t.Shape()...))
	if err := binary.Read(bytes.NewBuffer(t.Bytes()), binary.LittleEndian, &s); err != nil {
		panic(err)
	}

	shape := t.Shape()
Michael Yang's avatar
Michael Yang committed
299
	slices.Reverse(shape)
Michael Yang's avatar
Michael Yang committed
300
301

	var sb strings.Builder
302
303
	var f func([]int, int)
	f = func(dims []int, stride int) {
Michael Yang's avatar
Michael Yang committed
304
		prefix := strings.Repeat(" ", len(shape)-len(dims)+1)
Michael Yang's avatar
Michael Yang committed
305
306
		sb.WriteString("[")
		defer func() { sb.WriteString("]") }()
307
		for i := 0; i < dims[0]; i++ {
Michael Yang's avatar
Michael Yang committed
308
			if i >= items && i < dims[0]-items {
Michael Yang's avatar
Michael Yang committed
309
				sb.WriteString("..., ")
Michael Yang's avatar
Michael Yang committed
310
311
312
313
314
315
316
317
318
319
320
321
322
323
				// skip to next printable element
				skip := dims[0] - 2*items
				if len(dims) > 1 {
					stride += mul(append(dims[1:], skip)...)
					fmt.Fprint(&sb, strings.Repeat("\n", len(dims)-1), prefix)
				}
				i += skip - 1
			} else if len(dims) > 1 {
				f(dims[1:], stride)
				stride += mul(dims[1:]...)
				if i < dims[0]-1 {
					fmt.Fprint(&sb, ",", strings.Repeat("\n", len(dims)-1), prefix)
				}
			} else {
Michael Yang's avatar
Michael Yang committed
324
325
326
327
328
329
				text := fn(s[stride+i])
				if len(text) > 0 && text[0] != '-' {
					sb.WriteString(" ")
				}

				sb.WriteString(text)
Michael Yang's avatar
Michael Yang committed
330
				if i < dims[0]-1 {
Michael Yang's avatar
Michael Yang committed
331
					sb.WriteString(", ")
Michael Yang's avatar
Michael Yang committed
332
333
334
335
336
337
338
339
340
341
342
343
				}
			}
		}
	}
	f(shape, 0)

	return sb.String()
}

type DType int

const (
Jesse Gross's avatar
Jesse Gross committed
344
345
346
	DTypeOther DType = iota
	DTypeF32
	DTypeF16
347
348
	DTypeQ80
	DTypeQ40
Michael Yang's avatar
Michael Yang committed
349
350
	DTypeI32
)