backend.go 8.3 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
8
	"encoding/binary"
	"fmt"
	"os"
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
	Config() fs.Config
Michael Yang's avatar
Michael Yang committed
18
19
	Get(name string) Tensor
	NewContext() Context
20
	NewContextSize(size int) Context
Michael Yang's avatar
Michael Yang committed
21
22
}

23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
// 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
42
43
44
45
46
47
48
49

	// 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
50
51
}

52
53
// BackendParams controls how the backend loads and executes models
type BackendParams struct {
54
55
56
57
	// Progress is a callback function that allows reporting percentage completion
	// of model loading
	Progress func(float32)

58
59
	// NumThreads sets the number of threads to use if running on the CPU
	NumThreads int
Michael Yang's avatar
Michael Yang committed
60

61
62
63
64
65
66
67
68
	// 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
69
70
71

	// FlashAttention indicates that we should use a fused flash attention kernel
	FlashAttention bool
72
73
}

74
var backends = make(map[string]func(context.Context, *os.File, BackendParams) (Backend, error))
75

76
func RegisterBackend(name string, f func(context.Context, *os.File, BackendParams) (Backend, error)) {
Michael Yang's avatar
Michael Yang committed
77
78
79
80
81
82
83
	if _, ok := backends[name]; ok {
		panic("backend: backend already registered")
	}

	backends[name] = f
}

84
func NewBackend(ctx context.Context, f *os.File, params BackendParams) (Backend, error) {
Michael Yang's avatar
Michael Yang committed
85
	if backend, ok := backends["ggml"]; ok {
86
		return backend(ctx, f, params)
Michael Yang's avatar
Michael Yang committed
87
88
89
90
91
92
	}

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

type Context interface {
93
	Empty(dtype DType, shape ...int) Tensor
Michael Yang's avatar
Michael Yang committed
94
95
96
97
	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
98
99
100
	// Arange creates a 1D tensor with values within an interval (start, stop] increased by step.
	Arange(start, stop, step float32, dtype DType) Tensor

101
	Forward(...Tensor) Context
102
	Compute(...Tensor)
103
104
105
106
107
108
109

	// 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

110
	MaxGraphNodes() int
111
	Close()
112

113
114
	// Input returns a context appropriate for creating tensors that are
	// inputs to the model (which includes things like output locations)
115
116
117
118
	Input() Context

	// Layer returns a context appropriate for creating intermediate tensors
	Layer(int) Context
Michael Yang's avatar
Michael Yang committed
119
120
121
}

type Tensor interface {
122
123
	Dim(n int) int
	Stride(n int) int
Michael Yang's avatar
Michael Yang committed
124

125
	Shape() []int
Michael Yang's avatar
Michael Yang committed
126
127
128
129
130
	DType() DType

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

131
	Neg(ctx Context) Tensor
Michael Yang's avatar
Michael Yang committed
132
133
134
	Add(ctx Context, t2 Tensor) Tensor
	Mul(ctx Context, t2 Tensor) Tensor
	Mulmat(ctx Context, t2 Tensor) Tensor
135
	MulmatFullPrec(ctx Context, t2 Tensor) Tensor
Michael Yang's avatar
Michael Yang committed
136
137
138
139
140
141

	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
142
	AvgPool2D(ctx Context, k, s int, p float32) Tensor
Michael Yang's avatar
Michael Yang committed
143
	Conv2D(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
Michael Yang's avatar
Michael Yang committed
144

Patrick Devine's avatar
Patrick Devine committed
145
	RoPE(ctx Context, positionIDs, ropeFactors Tensor, dim, ropeType uint32, base, scale float32) Tensor
146
	IM2Col(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
Michael Yang's avatar
Michael Yang committed
147

148
149
	Sin(ctx Context) Tensor
	Cos(ctx Context) Tensor
Michael Yang's avatar
Michael Yang committed
150
151
152
153
	Tanh(ctx Context) Tensor
	GELU(ctx Context) Tensor
	SILU(ctx Context) Tensor

154
	Reshape(ctx Context, shape ...int) Tensor
Michael Yang's avatar
Michael Yang committed
155
156
157
	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
158
	Set(ctx Context, t2 Tensor, offset int, strides ...int) Tensor
Michael Yang's avatar
Michael Yang committed
159

160
161
	Pad(ctx Context, shape ...int) Tensor
	Unpad(ctx Context, shape ...int) Tensor
Michael Yang's avatar
Michael Yang committed
162
163

	Stack(ctx Context, dim int, s ...Tensor) Tensor
164
165
166

	// Repeat repeats the tensor n times along dimension dim
	Repeat(ctx Context, dim, n int) Tensor
Michael Yang's avatar
Michael Yang committed
167
168
169
	Concat(ctx Context, t2 Tensor, dim int) Tensor
	Rows(ctx Context, t2 Tensor) Tensor
	Copy(ctx Context, t2 Tensor) Tensor
170
	Duplicate(ctx Context) Tensor
Michael Yang's avatar
Michael Yang committed
171
172
}

173
174
175
176
// ScaledDotProductAttention implements a fused attention
// operation equivalent to following code on a tensor named
// query:
//
177
178
179
180
// 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)
//
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
// 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
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
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
}

type DumpOptions struct {
	// Items is the number of elements to print at the beginning and end of each dimension.
215
	Items int
Michael Yang's avatar
Michael Yang committed
216
217
218
219
220

	// Precision is the number of decimal places to print. Applies to float32 and float64.
	Precision int
}

Jesse Gross's avatar
Jesse Gross committed
221
func Dump(ctx Context, t Tensor, opts ...DumpOptions) string {
Michael Yang's avatar
Michael Yang committed
222
223
224
225
226
227
228
229
230
	if len(opts) < 1 {
		opts = append(opts, DumpOptions{
			Items:     3,
			Precision: 4,
		})
	}

	switch t.DType() {
	case DTypeF32:
Jesse Gross's avatar
Jesse Gross committed
231
232
233
		return dump[[]float32](ctx, t, opts[0].Items, func(f float32) string {
			return strconv.FormatFloat(float64(f), 'f', opts[0].Precision, 32)
		})
234
	case DTypeF16, DTypeQ80, DTypeQ40:
235
		f32 := ctx.Input().Empty(DTypeF32, t.Shape()...)
Jesse Gross's avatar
Jesse Gross committed
236
237
		f32 = t.Copy(ctx, f32)
		return dump[[]float32](ctx, f32, opts[0].Items, func(f float32) string {
Michael Yang's avatar
Michael Yang committed
238
239
240
			return strconv.FormatFloat(float64(f), 'f', opts[0].Precision, 32)
		})
	case DTypeI32:
Jesse Gross's avatar
Jesse Gross committed
241
		return dump[[]int32](ctx, t, opts[0].Items, func(i int32) string {
Michael Yang's avatar
Michael Yang committed
242
243
244
245
246
247
248
			return strconv.FormatInt(int64(i), 10)
		})
	default:
		return "<unsupported>"
	}
}

Jesse Gross's avatar
Jesse Gross committed
249
250
func dump[S ~[]E, E number](ctx Context, t Tensor, items int, fn func(E) string) string {
	if t.Bytes() == nil {
251
		ctx.Forward(t).Compute(t)
Michael Yang's avatar
Michael Yang committed
252
253
254
255
256
257
258
259
	}

	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
260
	slices.Reverse(shape)
Michael Yang's avatar
Michael Yang committed
261
262

	var sb strings.Builder
263
264
	var f func([]int, int)
	f = func(dims []int, stride int) {
Michael Yang's avatar
Michael Yang committed
265
		prefix := strings.Repeat(" ", len(shape)-len(dims)+1)
Michael Yang's avatar
Michael Yang committed
266
267
		sb.WriteString("[")
		defer func() { sb.WriteString("]") }()
268
		for i := 0; i < dims[0]; i++ {
Michael Yang's avatar
Michael Yang committed
269
			if i >= items && i < dims[0]-items {
Michael Yang's avatar
Michael Yang committed
270
				sb.WriteString("..., ")
Michael Yang's avatar
Michael Yang committed
271
272
273
274
275
276
277
278
279
280
281
282
283
284
				// 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
285
286
287
288
289
290
				text := fn(s[stride+i])
				if len(text) > 0 && text[0] != '-' {
					sb.WriteString(" ")
				}

				sb.WriteString(text)
Michael Yang's avatar
Michael Yang committed
291
				if i < dims[0]-1 {
Michael Yang's avatar
Michael Yang committed
292
					sb.WriteString(", ")
Michael Yang's avatar
Michael Yang committed
293
294
295
296
297
298
299
300
301
302
303
304
				}
			}
		}
	}
	f(shape, 0)

	return sb.String()
}

type DType int

const (
Jesse Gross's avatar
Jesse Gross committed
305
306
307
	DTypeOther DType = iota
	DTypeF32
	DTypeF16
308
309
	DTypeQ80
	DTypeQ40
Michael Yang's avatar
Michael Yang committed
310
311
	DTypeI32
)