moshi 0.6.4

moshi, a real-time voice AI
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
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328

// Copyright (c) Kyutai, all rights reserved.
// This source code is licensed under the license found in the
// LICENSE file in the root directory of this source tree.

use candle::{DType, Device, Result, Tensor};
use candle_nn::kv_cache::RotatingKvCache;

#[derive(Debug, Clone)]
pub struct IndicesAndMask {
    indices: Tensor,
    mask: Tensor,
}

impl IndicesAndMask {
    pub fn mask(&self) -> &Tensor {
        &self.mask
    }
}

#[derive(Debug, Clone)]
pub struct ScatteredKvCache {
    k: Tensor,
    v: Tensor,
    context: usize,
}

impl ScatteredKvCache {
    pub fn append(
        &mut self,
        k: &Tensor,
        v: &Tensor,
        iam: &IndicesAndMask,
    ) -> Result<(Tensor, Tensor)> {
        if self.context <= k.dim(2)? {
            return Ok((k.clone(), v.clone()));
        }
        let indices = iam.indices.unsqueeze(2)?.unsqueeze(1)?;
        let indices = indices.broadcast_as(k.shape())?.contiguous()?;
        self.k.scatter_set(&indices, k, 2)?;
        self.v.scatter_set(&indices, v, 2)?;
        Ok((self.k.clone(), self.v.clone()))
    }

    pub fn k(&self) -> &Tensor {
        &self.k
    }

    pub fn v(&self) -> &Tensor {
        &self.v
    }
}

#[derive(Debug, Clone)]
pub struct ScatteredCacheBuilder {
    context: usize,
    // The current position in the stream, this can be larger than context.
    positions: Vec<usize>,
    // The index where the next element will be stored.
    indices: Vec<usize>,
    dtype: DType,
    device: Device,
}

impl ScatteredCacheBuilder {
    pub fn new(batch_size: usize, context: usize, dtype: DType, device: &Device) -> Result<Self> {
        let positions = vec![0; batch_size];
        let indices = vec![0; batch_size];
        Ok(Self { positions, indices, context, dtype, device: device.clone() })
    }

    pub fn make_cache(&self, num_heads: usize, head_dim: usize) -> Result<ScatteredKvCache> {
        let batch_size = self.batch_size();
        let shape = (batch_size, num_heads, self.context, head_dim);
        let k = Tensor::zeros(shape, self.dtype, self.device())?;
        let v = Tensor::zeros(shape, self.dtype, self.device())?;
        Ok(ScatteredKvCache { k, v, context: self.context })
    }

    pub fn positions(&self) -> &[usize] {
        &self.positions
    }

    pub fn reset(&mut self) {
        self.positions.fill(0);
        self.indices.fill(0);
    }

    pub fn batch_size(&self) -> usize {
        self.positions.len()
    }

    pub fn reset_batch_index(&mut self, batch_index: usize) {
        self.positions[batch_index] = 0;
        self.indices[batch_index] = 0;
    }

    #[allow(clippy::needless_range_loop)]
    pub fn indices_and_mask(
        &mut self,
        seq_len: usize,
        batch_mask: &[bool],
    ) -> Result<IndicesAndMask> {
        // mask shape is (b, h, t, k)
        let context = self.context;
        if self.context <= seq_len {
            return self.indices_and_mask_abs(seq_len, batch_mask);
        }
        let mut attention_masks = Vec::with_capacity(self.batch_size());
        let mut cache_indices = Vec::with_capacity(self.batch_size());
        for (batch_i, &batch_mask) in batch_mask.iter().enumerate() {
            if !batch_mask {
                let masks: Vec<Vec<f32>> = vec![vec![0.0; context]; seq_len];
                let indices = vec![self.indices[batch_i] as u32; seq_len];
                attention_masks.push(masks);
                cache_indices.push(indices);
            } else {
                let start_index = self.indices[batch_i];
                let start_pos = self.positions[batch_i];
                let mut masks: Vec<Vec<f32>> = Vec::with_capacity(seq_len);
                let mut indices = Vec::with_capacity(seq_len);
                let mut all_pos = vec![usize::MAX; context];
                if start_pos < context {
                    for i in 0..start_pos {
                        all_pos[i] = i;
                    }
                } else {
                    let offset = start_pos - start_index;
                    for i in 0..context {
                        all_pos[i] =
                            if i < start_index { i + offset } else { i + offset - context };
                    }
                }
                for seq_i in 0..seq_len {
                    let index = self.indices[batch_i];
                    all_pos[index] = seq_i + start_pos;
                    indices.push(index as u32);
                    self.indices[batch_i] += 1;
                    self.positions[batch_i] += 1;
                    if self.indices[batch_i] >= self.context {
                        self.indices[batch_i] = 0;
                    }
                }

                for seq_i in 0..seq_len {
                    let my_pos = seq_i + start_pos;
                    let mask = all_pos
                        .iter()
                        .map(|&pos| if pos <= my_pos { 0.0 } else { f32::NEG_INFINITY })
                        .collect::<Vec<f32>>();
                    masks.push(mask);
                }

                attention_masks.push(masks);
                cache_indices.push(indices);
            }
        }
        // Flattening the attention mask then using Tensor::from_vec rather using Tensor::new ends
        // up being almost 10x faster with candle 0.9.0. The slowness seems to be on the CPU
        // copies, to be further investigated.
        let attention_masks =
            attention_masks.into_iter().flat_map(|m| m.into_iter().flatten()).collect::<Vec<f32>>();
        let mask = Tensor::from_vec(attention_masks, ((), 1, seq_len, context), self.device())?
            .to_dtype(self.dtype)?;
        let indices = Tensor::new(cache_indices, self.device())?;
        Ok(IndicesAndMask { indices, mask })
    }

    pub fn device(&self) -> &Device {
        &self.device
    }

    #[allow(clippy::needless_range_loop)]
    fn indices_and_mask_abs(
        &mut self,
        seq_len: usize,
        batch_mask: &[bool],
    ) -> Result<IndicesAndMask> {
        let mask = self.get_mask_abs(seq_len, seq_len)?;
        let mut cache_indices = Vec::with_capacity(self.batch_size());
        for (batch_i, &batch_mask) in batch_mask.iter().enumerate() {
            if !batch_mask {
                let indices = vec![self.indices[batch_i] as u32; seq_len];
                cache_indices.push(indices);
            } else {
                let mut indices = Vec::with_capacity(seq_len);
                for _ in 0..seq_len {
                    let index = self.indices[batch_i];
                    indices.push(index as u32);
                    self.indices[batch_i] += 1;
                    self.positions[batch_i] += 1;
                    if self.indices[batch_i] >= self.context {
                        self.indices[batch_i] = 0;
                    }
                }
                cache_indices.push(indices);
            }
        }
        let indices = Tensor::new(cache_indices, self.device())?;
        Ok(IndicesAndMask { indices, mask })
    }

    fn get_mask_abs(&self, size1: usize, size2: usize) -> Result<Tensor> {
        let context = self.context;
        let mask: Vec<_> = (0..size1)
            .flat_map(|i| {
                (0..size2).map(move |j| {
                    if size1 + j > size2 + i || size1 + j + context < size2 + i {
                        f32::NEG_INFINITY
                    } else {
                        0.0
                    }
                })
            })
            .collect();
        Tensor::from_slice(&mask, (size1, size2), self.device())
    }
}

#[derive(Debug, Clone)]
pub enum KvCache {
    Rotating(RotatingKvCache),
}

impl KvCache {
    pub fn new(dim: usize, max_seq_len: usize) -> Self {
        let cache = RotatingKvCache::new(dim, max_seq_len);
        Self::Rotating(cache)
    }

    pub fn current_seq_len(&self) -> usize {
        match self {
            KvCache::Rotating(cache) => cache.current_seq_len(),
        }
    }

    pub fn reset(&mut self) {
        match self {
            KvCache::Rotating(cache) => cache.reset(),
        }
    }

    pub fn append(&mut self, key: &Tensor, value: &Tensor) -> Result<(Tensor, Tensor)> {
        match self {
            KvCache::Rotating(cache) => cache.append(key, value),
        }
    }

    pub fn positions(&self, seq_len: usize) -> Vec<usize> {
        match self {
            KvCache::Rotating(cache) => cache.positions(seq_len),
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use candle::IndexOp;

    #[test]
    fn test_scattered_kv_cache() -> Result<()> {
        let device = Device::Cpu;
        let mut cache = ScatteredCacheBuilder::new(2, 5, DType::F32, &device)?;
        let inf = f32::INFINITY;

        let iam = cache.indices_and_mask(1, &[true, false])?;
        let mask = iam.mask.i((.., 0))?.to_vec3::<f32>()?;
        assert_eq!(iam.indices.to_vec2::<u32>()?, [[0], [0]]);
        assert_eq!(mask, [[[0.0, -inf, -inf, -inf, -inf]], [[0.0, 0.0, 0.0, 0.0, 0.0]]]);

        let iam = cache.indices_and_mask(1, &[true, false])?;
        let mask = iam.mask.i((.., 0))?.to_vec3::<f32>()?;
        assert_eq!(iam.indices.to_vec2::<u32>()?, [[1], [0]]);
        assert_eq!(mask, [[[0.0, 0.0, -inf, -inf, -inf]], [[0.0, 0.0, 0.0, 0.0, 0.0]]]);

        let iam = cache.indices_and_mask(3, &[false, true])?;
        let mask = iam.mask.i((.., 0))?.to_vec3::<f32>()?;
        assert_eq!(iam.indices.to_vec2::<u32>()?, [[2, 2, 2], [0, 1, 2]]);
        assert_eq!(
            mask,
            [
                [[0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0]],
                [
                    [0.0, -inf, -inf, -inf, -inf],
                    [0.0, 0.0, -inf, -inf, -inf],
                    [0.0, 0.0, 0.0, -inf, -inf]
                ]
            ]
        );

        let iam = cache.indices_and_mask(3, &[true, true])?;
        let mask = iam.mask.i((.., 0))?.to_vec3::<f32>()?;
        assert_eq!(iam.indices.to_vec2::<u32>()?, [[2, 3, 4], [3, 4, 0]]);
        assert_eq!(
            mask,
            [
                [
                    [0.0, 0.0, 0.0, -inf, -inf],
                    [0.0, 0.0, 0.0, 0.0, -inf],
                    [0.0, 0.0, 0.0, 0.0, 0.0]
                ],
                [
                    [-inf, 0.0, 0.0, 0.0, -inf],
                    [-inf, 0.0, 0.0, 0.0, 0.0],
                    [0.0, 0.0, 0.0, 0.0, 0.0]
                ]
            ]
        );

        let iam = cache.indices_and_mask(1, &[true, false])?;
        let mask = iam.mask.i((.., 0))?.to_vec3::<f32>()?;
        assert_eq!(iam.indices.to_vec2::<u32>()?, [[0], [1]]);
        assert_eq!(mask, [[[0.0, 0.0, 0.0, 0.0, 0.0]], [[0.0, 0.0, 0.0, 0.0, 0.0]]]);

        let iam = cache.indices_and_mask(2, &[true, false])?;
        let mask = iam.mask.i((.., 0))?.to_vec3::<f32>()?;
        assert_eq!(iam.indices.to_vec2::<u32>()?, [[1, 2], [1, 1]]);
        assert_eq!(
            mask,
            [
                [[0.0, 0.0, -inf, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0]],
                [[0.0, 0.0, 0.0, 0.0, 0.0], [0.0, 0.0, 0.0, 0.0, 0.0]]
            ]
        );

        Ok(())
    }
}