luminal 0.2.0

Deep learning at the speed of light.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
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
266
267
268
269
270
271
272
273
274
275
276
277
278

use std::ops::Mul;

use crate::{nn::linear::Linear, prelude::*};

// This is still single head attention because I need a runtime reshape, like the try_reshape in dfdx
pub struct MultiHeadSelfAttention<
    const DIM: usize,
    const K_DIM: usize,
    const V_DIM: usize,
    const HEADS: usize,
> {
    pub w_q: Linear<DIM, K_DIM>,
    pub w_k: Linear<DIM, K_DIM>,
    pub w_v: Linear<DIM, V_DIM>,
    pub w_o: Linear<V_DIM, DIM>,
}

impl<const DIM: usize, const K_DIM: usize, const V_DIM: usize, const HEADS: usize> InitModule
    for MultiHeadSelfAttention<DIM, K_DIM, V_DIM, HEADS>
{
    fn initialize(cx: &mut Graph) -> Self {
        Self {
            w_q: InitModule::initialize(cx),
            w_k: InitModule::initialize(cx),
            w_v: InitModule::initialize(cx),
            w_o: InitModule::initialize(cx),
        }
    }
}

impl<const DIM: usize, const K_DIM: usize, const V_DIM: usize, const HEADS: usize> SerializeModule
    for MultiHeadSelfAttention<DIM, K_DIM, V_DIM, HEADS>
{
    fn serialize(&self, s: &mut Serializer) {
        s.module("w_q", &self.w_q);
        s.module("w_k", &self.w_k);
        s.module("w_v", &self.w_v);
        s.module("w_o", &self.w_o);
    }
}

// Single
impl<
        const DIM: usize,
        const K_DIM: usize,
        const V_DIM: usize,
        const HEADS: usize,
        S: Dimension,
    > Module<GraphTensor<(S, Const<DIM>)>> for MultiHeadSelfAttention<DIM, K_DIM, V_DIM, HEADS>
{
    type Output = GraphTensor<(S, Const<DIM>)>;

    fn forward(&self, input: GraphTensor<(S, Const<DIM>)>) -> Self::Output {
        // Pass to batched forward
        <Self as Module<GraphTensor<(Const<1>, S, Const<DIM>)>>>::forward(self, input.expand())
            .reshape()
    }
}

impl<
        const DIM: usize,
        const K_DIM: usize,
        const V_DIM: usize,
        const HEADS: usize,
        S: Dimension,
        S1: Dimension,
    >
    Module<(
        GraphTensor<(S, Const<DIM>)>,
        GraphTensor<(S1, Const<DIM>)>,
        GraphTensor<(S, Const<DIM>)>,
    )> for MultiHeadSelfAttention<DIM, K_DIM, V_DIM, HEADS>
{
    type Output = GraphTensor<(S1, Const<DIM>)>;

    fn forward(
        &self,
        (k, q, v): (
            GraphTensor<(S, Const<DIM>)>,
            GraphTensor<(S1, Const<DIM>)>,
            GraphTensor<(S, Const<DIM>)>,
        ),
    ) -> Self::Output {
        // Pass to batched forward
        <Self as Module<(
            GraphTensor<(Const<1>, S, Const<DIM>)>,
            GraphTensor<(Const<1>, S1, Const<DIM>)>,
            GraphTensor<(Const<1>, S, Const<DIM>)>,
        )>>::forward(self, (k.expand(), q.expand(), v.expand()))
        .reshape()
    }
}

// Batched
impl<
        const DIM: usize,
        const K_DIM: usize,
        const V_DIM: usize,
        const HEADS: usize,
        S: Dimension,
        B: Dimension,
    > Module<GraphTensor<(B, S, Const<DIM>)>> for MultiHeadSelfAttention<DIM, K_DIM, V_DIM, HEADS>
{
    type Output = GraphTensor<(B, S, Const<DIM>)>;

    fn forward(&self, input: GraphTensor<(B, S, Const<DIM>)>) -> Self::Output {
        <Self as Module<(
            GraphTensor<(B, S, Const<DIM>)>,
            GraphTensor<(B, S, Const<DIM>)>,
            GraphTensor<(B, S, Const<DIM>)>,
        )>>::forward(self, (input, input, input))
    }
}

// Batched different key-query-value
impl<
        const DIM: usize,
        const K_DIM: usize,
        const V_DIM: usize,
        const HEADS: usize,
        S1: Dimension,
        S2: Dimension,
        B: Dimension,
    >
    Module<(
        GraphTensor<(B, S1, Const<DIM>)>,
        GraphTensor<(B, S2, Const<DIM>)>,
        GraphTensor<(B, S1, Const<DIM>)>,
    )> for MultiHeadSelfAttention<DIM, K_DIM, V_DIM, HEADS>
{
    type Output = GraphTensor<(B, S2, Const<DIM>)>;

    fn forward(
        &self,
        (keys, queries, values): (
            GraphTensor<(B, S1, Const<DIM>)>,
            GraphTensor<(B, S2, Const<DIM>)>,
            GraphTensor<(B, S1, Const<DIM>)>,
        ),
    ) -> Self::Output {
        let values = self
            .w_v
            .forward(values)
            .dyn_reshape::<(B, S1, Dyn<'-'>, Dyn<'-'>)>(vec![
                B::const_size(),
                S1::const_size(),
                HEADS.into(),
                (K_DIM / HEADS).into(),
            ])
            .permute::<_, Axes4<0, 2, 1, 3>>();
        let keys = self
            .w_k
            .forward(keys)
            .dyn_reshape::<(B, S1, Dyn<'-'>, Dyn<'-'>)>(vec![
                B::const_size(),
                S1::const_size(),
                HEADS.into(),
                (K_DIM / HEADS).into(),
            ])
            .permute::<_, Axes4<0, 2, 3, 1>>();
        let queries = self
            .w_q
            .forward(queries)
            .dyn_reshape::<(B, S2, Dyn<'-'>, Dyn<'-'>)>(vec![
                B::const_size(),
                S2::const_size(),
                HEADS.into(),
                (K_DIM / HEADS).into(),
            ])
            .permute::<_, Axes4<0, 2, 1, 3>>();

        let weights = queries
            .matmul(keys)
            .mul((1.0 / ((K_DIM / HEADS) as f64).sqrt()) as f32)
            .softmax::<3>();

        let tokens: GraphTensor<(B, S2, Const<V_DIM>)> = weights
            .matmul(values)
            .permute::<_, Axes4<0, 2, 1, 3>>()
            .reshape();
        self.w_o.forward(tokens)
    }
}

#[cfg(test)]
mod tests {
    use crate::{
        prelude::{Module, *},
        tests::assert_close,
    };
    use dfdx::prelude::{Module as DfdxModule, *};

    use super::MultiHeadSelfAttention;
    #[test]
    fn test_self_attention() {
        let mut cx = Graph::new();
        let model: MultiHeadSelfAttention<3, 3, 3, 1> = InitModule::initialize(&mut cx);
        model
            .w_k
            .weight
            .set(vec![1., 22., 3., 1., 2., 3., 1., 2., 3.]);
        model
            .w_q
            .weight
            .set(vec![3., 2., 3., 1.3, 2., 3., 3., 2., 3.]);
        model
            .w_v
            .weight
            .set(vec![-1., 12., 3., -1., 2., -3., 11., 2., 3.]);
        model
            .w_o
            .weight
            .set(vec![1., 22., 3., 1., 2., 3., 1., 2., 3.]);

        let a = cx.tensor::<(Dyn<'d'>, crate::shape::Const<3>)>();
        let e = cx.tensor::<(Dyn<'e'>, crate::shape::Const<3>)>();
        let b = model.forward((e, a, e));

        a.set_dyn(
            vec![
                0.56587636, -1.4053632, 0.8394869, 0.5916256, -1.4082357, 0.8166099,
            ],
            &[2, 3],
        );
        e.set_dyn(
            vec![-1.0, 2.0, 3.0, 3.0, 3.0, -1.0, -1.0, 2.0, 3.0],
            &[3, 3],
        );
        b.retrieve();

        cx.execute();

        let d_dev = Cpu::default();
        let mut d_model: dfdx::nn::modules::MultiHeadAttention<3, 1, 3, 3, f32, Cpu> =
            d_dev.build_module::<MultiHeadAttention<3, 1, 3, 3>, f32>();
        d_model.w_k.bias.copy_from(&[0.0, 0.0, 0.0]);
        d_model.w_v.bias.copy_from(&[0.0, 0.0, 0.0]);
        d_model.w_q.bias.copy_from(&[0.0, 0.0, 0.0]);
        d_model.w_o.bias.copy_from(&[0.0, 0.0, 0.0]);
        d_model.w_o.weight = d_dev
            .tensor_from_vec(
                vec![1., 22., 3., 1., 2., 3., 1., 2., 3.],
                (dfdx::shapes::Const::<3>, dfdx::shapes::Const::<3>),
            )
            .permute();
        d_model.w_k.weight = d_dev
            .tensor_from_vec(
                vec![1., 22., 3., 1., 2., 3., 1., 2., 3.],
                (dfdx::shapes::Const::<3>, dfdx::shapes::Const::<3>),
            )
            .permute();
        d_model.w_q.weight = d_dev
            .tensor_from_vec(
                vec![3., 2., 3., 1.3, 2., 3., 3., 2., 3.],
                (dfdx::shapes::Const::<3>, dfdx::shapes::Const::<3>),
            )
            .permute();
        d_model.w_v.weight = d_dev
            .tensor_from_vec(
                vec![-1., 12., 3., -1., 2., -3., 11., 2., 3.],
                (dfdx::shapes::Const::<3>, dfdx::shapes::Const::<3>),
            )
            .permute();
        let d_a = d_dev.tensor_from_vec(
            vec![
                0.56587636, -1.4053632, 0.8394869, 0.5916256, -1.4082357, 0.8166099,
            ],
            (dfdx::shapes::Const::<2>, dfdx::shapes::Const::<3>),
        );
        let d_e = d_dev.tensor_from_vec(
            vec![-1.0, 2.0, 3.0, 3.0, 3.0, -1.0, -1.0, 2.0, 3.0],
            (dfdx::shapes::Const::<3>, dfdx::shapes::Const::<3>),
        );
        let d_b = d_model.forward((d_a, d_e.clone(), d_e));

        assert_close(&b.data(), &d_b.as_vec());
    }
}