"LICENSE" did not exist on "1ae241092f47a7bf78857a8545f84790e70bf1aa"
model_text.go 8.65 KB
Newer Older
Michael Yang's avatar
Michael Yang committed
1
2
3
4
5
6
package mllama

import (
	"math"
	"slices"

Jesse Gross's avatar
Jesse Gross committed
7
	"github.com/ollama/ollama/kvcache"
Michael Yang's avatar
Michael Yang committed
8
9
10
11
12
13
14
15
16
17
18
	"github.com/ollama/ollama/ml"
	"github.com/ollama/ollama/ml/nn"
)

type TextSelfAttention struct {
	Query  *nn.Linear `gguf:"attn_q"`
	Key    *nn.Linear `gguf:"attn_k"`
	Value  *nn.Linear `gguf:"attn_v"`
	Output *nn.Linear `gguf:"attn_output"`
}

Jesse Gross's avatar
Jesse Gross committed
19
func (sa *TextSelfAttention) Forward(ctx ml.Context, hiddenState, positions, _ ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
20
21
22
23
24
25
26
27
28
29
30
31
32
33
	batchSize := hiddenState.Dim(1)
	headDim := opts.hiddenSize / opts.numHeads

	query := sa.Query.Forward(ctx, hiddenState)
	query = query.Reshape(ctx, headDim, opts.numHeads, batchSize)
	query = query.RoPE(ctx, positions, opts.RopeFactors, opts.ropeDim, opts.ropeBase, opts.ropeScale)

	key := sa.Key.Forward(ctx, hiddenState)
	key = key.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
	key = key.RoPE(ctx, positions, opts.RopeFactors, opts.ropeDim, opts.ropeBase, opts.ropeScale)

	value := sa.Value.Forward(ctx, hiddenState)
	value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)

34
	scaleFactor := 1.0 / math.Sqrt(float64(headDim))
35
	attention := nn.Attention(ctx, query, key, value, scaleFactor, cache)
Michael Yang's avatar
Michael Yang committed
36
37
38
39
40
	attention = attention.Reshape(ctx, opts.hiddenSize, batchSize)

	return sa.Output.Forward(ctx, attention)
}

Jesse Gross's avatar
Jesse Gross committed
41
func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
42
	// This will only get called for layers in the causal cache, which are just the self attention layers
Jesse Gross's avatar
Jesse Gross committed
43
44
45
	return key.RoPE(ctx, shift, m.RopeFactors, m.ropeDim, m.ropeBase, m.ropeScale), nil
}

Michael Yang's avatar
Michael Yang committed
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
type TextMLP struct {
	Up   *nn.Linear `gguf:"ffn_up"`
	Down *nn.Linear `gguf:"ffn_down"`
	Gate *nn.Linear `gguf:"ffn_gate"`
}

func (mlp *TextMLP) Forward(ctx ml.Context, hiddenState ml.Tensor, opts *TextModelOptions) ml.Tensor {
	hiddenState = mlp.Gate.Forward(ctx, hiddenState).SILU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenState))
	return mlp.Down.Forward(ctx, hiddenState)
}

type TextSelfAttentionDecoderLayer struct {
	AttentionNorm *nn.RMSNorm `gguf:"attn_norm"`
	SelfAttention *TextSelfAttention

	MLPNorm *nn.RMSNorm `gguf:"ffn_norm"`
	MLP     *TextMLP
}

65
func (d *TextSelfAttentionDecoderLayer) Forward(ctx ml.Context, hiddenState, positions, outputs, mask, _, _ ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
66
67
68
69
	residual := hiddenState

	hiddenState = d.AttentionNorm.Forward(ctx, hiddenState, opts.eps)
	hiddenState = d.SelfAttention.Forward(ctx, hiddenState, positions, mask, cache, opts)
70
71
72
73
74
75
76
77

	// In the final layer (outputs != nil), optimize by pruning to just the token positions
	// we need logits for.
	if outputs != nil {
		hiddenState = hiddenState.Rows(ctx, outputs)
		residual = residual.Rows(ctx, outputs)
	}

Michael Yang's avatar
Michael Yang committed
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
	hiddenState = hiddenState.Add(ctx, residual)
	residual = hiddenState

	hiddenState = d.MLPNorm.Forward(ctx, hiddenState, opts.eps)
	hiddenState = d.MLP.Forward(ctx, hiddenState, opts)
	return hiddenState.Add(ctx, residual)
}

type TextCrossAttention struct {
	QueryNorm *nn.RMSNorm `gguf:"cross_attn_q_norm"`
	Query     *nn.Linear  `gguf:"cross_attn_q_proj"`
	KeyNorm   *nn.RMSNorm `gguf:"cross_attn_k_norm"`
	Key       *nn.Linear  `gguf:"cross_attn_k_proj"`
	Value     *nn.Linear  `gguf:"cross_attn_v_proj"`
	Output    *nn.Linear  `gguf:"cross_attn_o_proj"`
}

Jesse Gross's avatar
Jesse Gross committed
95
func (ca *TextCrossAttention) Forward(ctx ml.Context, hiddenState, crossAttentionStates ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
96
97
98
99
100
101
102
	batchSize := hiddenState.Dim(1)
	headDim := opts.hiddenSize / opts.numHeads

	query := ca.Query.Forward(ctx, hiddenState)
	query = query.Reshape(ctx, headDim, opts.numHeads, batchSize)
	query = ca.QueryNorm.Forward(ctx, query, opts.eps)

103
	var key, value ml.Tensor
Jesse Gross's avatar
Jesse Gross committed
104
105
	if crossAttentionStates != nil {
		numVisionTokens, numTiles := crossAttentionStates.Dim(1), crossAttentionStates.Dim(2)
Michael Yang's avatar
Michael Yang committed
106

Jesse Gross's avatar
Jesse Gross committed
107
108
109
		key = ca.Key.Forward(ctx, crossAttentionStates)
		key = key.Reshape(ctx, headDim, opts.numKVHeads, numVisionTokens*numTiles)
		key = ca.KeyNorm.Forward(ctx, key, opts.eps)
Michael Yang's avatar
Michael Yang committed
110

Jesse Gross's avatar
Jesse Gross committed
111
112
113
114
115
		value = ca.Value.Forward(ctx, crossAttentionStates)
		value = value.Reshape(ctx, headDim, opts.numKVHeads, numVisionTokens*numTiles)

		cache.Put(ctx, key, value)
	}
Michael Yang's avatar
Michael Yang committed
116

117
	key, value, _ = cache.Get(ctx)
Michael Yang's avatar
Michael Yang committed
118

119
	scaleFactor := 1.0 / math.Sqrt(float64(headDim))
120
121
122
123
124
125
126
127
128
129
130
131

	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)

	kq := key.MulmatFullPrec(ctx, query)

	kq = kq.Scale(ctx, scaleFactor)
	kq = kq.Softmax(ctx)

	kqv := value.Mulmat(ctx, kq)
	attention := kqv.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
Michael Yang's avatar
Michael Yang committed
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
	attention = attention.Reshape(ctx, opts.hiddenSize, batchSize)

	return ca.Output.Forward(ctx, attention)
}

type TextCrossAttentionDecoderLayer struct {
	AttentionNorm  *nn.RMSNorm `gguf:"attn_norm"`
	CrossAttention *TextCrossAttention
	AttentionGate  ml.Tensor `gguf:"cross_attn_attn_gate"`

	MLPNorm *nn.RMSNorm `gguf:"ffn_norm"`
	MLP     *TextMLP
	MLPGate ml.Tensor `gguf:"cross_attn_mlp_gate"`
}

147
func (d *TextCrossAttentionDecoderLayer) Forward(ctx ml.Context, hiddenState, _, _, _, crossAttentionStates, crossAttentionMask ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
	residual := hiddenState

	hiddenState = d.AttentionNorm.Forward(ctx, hiddenState, opts.eps)
	hiddenState = d.CrossAttention.Forward(ctx, hiddenState, crossAttentionStates, cache, opts)
	hiddenState = hiddenState.Mul(ctx, d.AttentionGate.Tanh(ctx))
	hiddenState = hiddenState.Add(ctx, residual)
	residual = hiddenState

	hiddenState = d.MLPNorm.Forward(ctx, hiddenState, opts.eps)
	hiddenState = d.MLP.Forward(ctx, hiddenState, opts)
	hiddenState = hiddenState.Mul(ctx, d.MLPGate.Tanh(ctx))
	return hiddenState.Add(ctx, residual)
}

type TextDecoderLayer interface {
163
	Forward(ctx ml.Context, hiddenState, positionIDs, outputs, mask, crossAttentionStates, crossAttentionMask ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor
Michael Yang's avatar
Michael Yang committed
164
165
166
167
168
169
}

type TextDecoder struct {
	Layers []TextDecoderLayer
}

170
func (d *TextDecoder) Forward(ctx ml.Context, hiddenState, positionIDs, outputs, mask, crossAttentionStates, crossAttentionMask ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
171
	for i, layer := range d.Layers {
Jesse Gross's avatar
Jesse Gross committed
172
173
174
175
176
177
178
179
180
		layerType := selfAttentionLayer
		if slices.Contains(opts.crossAttentionLayers, uint32(i)) {
			layerType = crossAttentionLayer
		}

		cache.SetLayer(i)
		cache.SetLayerType(layerType)

		if layerType == selfAttentionLayer || crossAttentionStates != nil || cache.UnderlyingCache().(*kvcache.EncoderCache).EncoderCached() {
181
182
183
184
185
186
			var lastLayerOutputs ml.Tensor
			if i == len(d.Layers)-1 {
				lastLayerOutputs = outputs
			}

			hiddenState = layer.Forward(ctx, hiddenState, positionIDs, lastLayerOutputs, mask, crossAttentionStates, crossAttentionMask, cache, opts)
Michael Yang's avatar
Michael Yang committed
187
188
189
190
191
192
193
194
195
		}
	}

	return hiddenState
}

type TextModelOptions struct {
	RopeFactors ml.Tensor `gguf:"rope_freqs.weight"`

196
	hiddenSize, numHeads, numKVHeads int
Michael Yang's avatar
Michael Yang committed
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
	eps, ropeBase, ropeScale         float32
	ropeDim                          uint32

	crossAttentionLayers []uint32
}

type TextModel struct {
	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
	Transformer    *TextDecoder  `gguf:"blk"`
	OutputNorm     *nn.RMSNorm   `gguf:"output_norm"`
	Output         *nn.Linear    `gguf:"output"`

	*TextModelOptions
}

212
func (m *TextModel) Forward(ctx ml.Context, inputIDs, positionIDs, outputs, mask, crossAttentionStates, crossAttentionMask ml.Tensor, cache *kvcache.WrapperCache) ml.Tensor {
Michael Yang's avatar
Michael Yang committed
213
	hiddenState := m.TokenEmbedding.Forward(ctx, inputIDs)
214
	hiddenState = m.Transformer.Forward(ctx, hiddenState, positionIDs, outputs, mask, crossAttentionStates, crossAttentionMask, cache, m.TextModelOptions)
Michael Yang's avatar
Michael Yang committed
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
	hiddenState = m.OutputNorm.Forward(ctx, hiddenState, m.eps)
	return m.Output.Forward(ctx, hiddenState)
}

func newTextModel(c ml.Config) *TextModel {
	var decoderLayers []TextDecoderLayer
	for i := range c.Uint("block_count") {
		var textDecoderLayer TextDecoderLayer
		if slices.Contains(c.Uints("attention.cross_attention_layers"), i) {
			textDecoderLayer = &TextCrossAttentionDecoderLayer{}
		} else {
			textDecoderLayer = &TextSelfAttentionDecoderLayer{}
		}

		decoderLayers = append(decoderLayers, textDecoderLayer)
	}

	return &TextModel{
		Transformer: &TextDecoder{Layers: decoderLayers},
		TextModelOptions: &TextModelOptions{
235
236
237
			hiddenSize:           int(c.Uint("embedding_length")),
			numHeads:             int(c.Uint("attention.head_count")),
			numKVHeads:           int(c.Uint("attention.head_count_kv")),
Michael Yang's avatar
Michael Yang committed
238
239
240
241
242
243
244
245
			eps:                  c.Float("attention.layer_norm_rms_epsilon"),
			ropeBase:             c.Float("rope.freq_base"),
			ropeScale:            c.Float("rope.freq_scale", 1),
			ropeDim:              c.Uint("rope.dimension_count"),
			crossAttentionLayers: c.Uints("attention.cross_attention_layers"),
		},
	}
}