backend.go 7.4 KB
Newer Older
Michael Yang's avatar
Michael Yang committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
package ml

import (
	"bytes"
	"encoding/binary"
	"fmt"
	"os"
	"strconv"
	"strings"
)

type Config interface {
	Architecture() string
	String(string, ...string) string
	Uint(string, ...uint32) uint32
	Float(string, ...float32) float32
17
	Bool(string, ...bool) bool
Michael Yang's avatar
Michael Yang committed
18
19
20
21
22
23
24
25
26

	Strings(string, ...[]string) []string
	Uints(string, ...[]uint32) []uint32
}

type Backend interface {
	Config() Config
	Get(name string) Tensor
	NewContext() Context
27
	NewContextSize(size int) Context
Michael Yang's avatar
Michael Yang committed
28
29
}

30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
// 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
49
50
51
52
53
54
55
56

	// 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
57
58
}

59
60
61
62
// 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
63

64
65
66
67
68
69
70
71
	// 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
72
73
74

	// FlashAttention indicates that we should use a fused flash attention kernel
	FlashAttention bool
75
76
77
78
79
}

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

func RegisterBackend(name string, f func(*os.File, BackendParams) (Backend, error)) {
Michael Yang's avatar
Michael Yang committed
80
81
82
83
84
85
86
	if _, ok := backends[name]; ok {
		panic("backend: backend already registered")
	}

	backends[name] = f
}

87
func NewBackend(f *os.File, params BackendParams) (Backend, error) {
Michael Yang's avatar
Michael Yang committed
88
	if backend, ok := backends["ggml"]; ok {
89
		return backend(f, params)
Michael Yang's avatar
Michael Yang committed
90
91
92
93
94
95
	}

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

type Context interface {
96
	Empty(dtype DType, shape ...int) Tensor
Michael Yang's avatar
Michael Yang committed
97
98
99
100
	Zeros(dtype DType, shape ...int) Tensor
	FromFloatSlice(s []float32, shape ...int) (Tensor, error)
	FromIntSlice(s []int32, shape ...int) (Tensor, error)

101
	Forward(...Tensor) Context
102
	Compute(...Tensor)
103
	MaxGraphNodes() int
104
	Close()
105
106
107
108
109
110
111
112
113

	// Input returns a context appropriate for creating input tensors
	Input() Context

	// Output returns a context appropriate for creating output tensors
	Output() Context

	// Layer returns a context appropriate for creating intermediate tensors
	Layer(int) Context
Michael Yang's avatar
Michael Yang committed
114
115
116
}

type Tensor interface {
117
118
	Dim(n int) int
	Stride(n int) int
Michael Yang's avatar
Michael Yang committed
119

120
	Shape() []int
Michael Yang's avatar
Michael Yang committed
121
122
123
124
125
126
127
128
	DType() DType

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

	Add(ctx Context, t2 Tensor) Tensor
	Mul(ctx Context, t2 Tensor) Tensor
	Mulmat(ctx Context, t2 Tensor) Tensor
129
	MulmatFullPrec(ctx Context, t2 Tensor) Tensor
Michael Yang's avatar
Michael Yang committed
130
131
132
133
134
135
136
137
138
139
140
141
142

	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

	Conv2D(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
	RoPE(ctx Context, positionIDs, ropeFactors Tensor, dim uint32, base, scale float32) Tensor

	Tanh(ctx Context) Tensor
	GELU(ctx Context) Tensor
	SILU(ctx Context) Tensor

143
	Reshape(ctx Context, shape ...int) Tensor
Michael Yang's avatar
Michael Yang committed
144
145
146
147
	View(ctx Context, offset int, shape ...int) Tensor
	Permute(ctx Context, shape ...int) Tensor
	Contiguous(ctx Context) Tensor

148
149
	Pad(ctx Context, shape ...int) Tensor
	Unpad(ctx Context, shape ...int) Tensor
Michael Yang's avatar
Michael Yang committed
150
151
152
153
154
155
156

	Stack(ctx Context, dim int, s ...Tensor) Tensor
	Concat(ctx Context, t2 Tensor, dim int) Tensor
	Rows(ctx Context, t2 Tensor) Tensor
	Copy(ctx Context, t2 Tensor) Tensor
}

157
158
159
160
// ScaledDotProductAttention implements a fused attention
// operation equivalent to following code on a tensor named
// query:
//
161
162
163
164
// 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)
//
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
// 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
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
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.
199
	Items int
Michael Yang's avatar
Michael Yang committed
200
201
202
203
204

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

Jesse Gross's avatar
Jesse Gross committed
205
func Dump(ctx Context, t Tensor, opts ...DumpOptions) string {
Michael Yang's avatar
Michael Yang committed
206
207
208
209
210
211
212
213
214
	if len(opts) < 1 {
		opts = append(opts, DumpOptions{
			Items:     3,
			Precision: 4,
		})
	}

	switch t.DType() {
	case DTypeF32:
Jesse Gross's avatar
Jesse Gross committed
215
216
217
218
		return dump[[]float32](ctx, t, opts[0].Items, func(f float32) string {
			return strconv.FormatFloat(float64(f), 'f', opts[0].Precision, 32)
		})
	case DTypeF16:
219
		f32 := ctx.Empty(DTypeF32, t.Shape()...)
Jesse Gross's avatar
Jesse Gross committed
220
221
		f32 = t.Copy(ctx, f32)
		return dump[[]float32](ctx, f32, opts[0].Items, func(f float32) string {
Michael Yang's avatar
Michael Yang committed
222
223
224
			return strconv.FormatFloat(float64(f), 'f', opts[0].Precision, 32)
		})
	case DTypeI32:
Jesse Gross's avatar
Jesse Gross committed
225
		return dump[[]int32](ctx, t, opts[0].Items, func(i int32) string {
Michael Yang's avatar
Michael Yang committed
226
227
228
229
230
231
232
			return strconv.FormatInt(int64(i), 10)
		})
	default:
		return "<unsupported>"
	}
}

Jesse Gross's avatar
Jesse Gross committed
233
234
func dump[S ~[]E, E number](ctx Context, t Tensor, items int, fn func(E) string) string {
	if t.Bytes() == nil {
235
		ctx.Forward(t).Compute(t)
Michael Yang's avatar
Michael Yang committed
236
237
238
239
240
241
242
243
244
245
	}

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

	shape := t.Shape()

	var sb strings.Builder
246
247
	var f func([]int, int)
	f = func(dims []int, stride int) {
Michael Yang's avatar
Michael Yang committed
248
249
250
		prefix := strings.Repeat(" ", len(shape)-len(dims)+1)
		fmt.Fprint(&sb, "[")
		defer func() { fmt.Fprint(&sb, "]") }()
251
		for i := 0; i < dims[0]; i++ {
Michael Yang's avatar
Michael Yang committed
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
			if i >= items && i < dims[0]-items {
				fmt.Fprint(&sb, "..., ")
				// 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 {
				fmt.Fprint(&sb, fn(s[stride+i]))
				if i < dims[0]-1 {
					fmt.Fprint(&sb, ", ")
				}
			}
		}
	}
	f(shape, 0)

	return sb.String()
}

type DType int

const (
Jesse Gross's avatar
Jesse Gross committed
283
284
285
	DTypeOther DType = iota
	DTypeF32
	DTypeF16
Michael Yang's avatar
Michael Yang committed
286
287
	DTypeI32
)