bunsen 0.31.0

bunsen is a batteries included common library for burn
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
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
//! # Beam search, as upstream does it.
//!
//! Rows are laid out `row = audio_idx * k + beam_idx`, beam varying fastest.
//! Every step, each beam proposes its `k + 1` best next tokens; the
//! `k * (k + 1)` candidates of one audio are deduplicated **by full
//! sequence** (at the first step every beam is the same prompt, and without
//! this the beam would quietly collapse to width one while still paying for
//! `k`), ranked by cumulative log probability, and the best `k` that did not
//! end survive, with the self-attention cache permuted to follow them. A
//! candidate that ended joins the audio's finished set, which `patience`
//! caps at `round(k * patience)`; the search completes when every audio's
//! set is full.
//!
//! Log probabilities accumulate in `f32`, as upstream's do, so that a
//! near-tie ranks the same way.

use std::collections::HashMap;

use burn::{
    Tensor,
    prelude::Backend,
    tensor::activation::log_softmax,
};

use crate::kits::speech::whisper::decode::TokenDecoder;

/// A finished set: insertion-ordered, unique by sequence, as a Python dict.
#[derive(Debug, Clone, Default)]
struct Finished {
    entries: Vec<(Vec<i64>, f32)>,
}

impl Finished {
    fn len(&self) -> usize {
        self.entries.len()
    }

    /// Mark `{ sequence => score }`.
    ///
    /// If the sequence already had a score, overwrite it.
    fn score_sequence(
        &mut self,
        sequence: Vec<i64>,
        score: f32,
    ) {
        match self.entries.iter_mut().find(|(s, _)| *s == sequence) {
            // TODO: should we take the max?
            Some(entry) => entry.1 = score,
            None => self.entries.push((sequence, score)),
        }
    }

    /// Clip the tails (`[offset..<EOT>]`) of sequences.
    fn clip_tails(
        &self,
        offset: usize,
        stop_token: i64,
    ) -> Vec<(Vec<i64>, f32)> {
        self.entries
            .iter()
            .map(|(sequence, score)| {
                let end = sequence[offset..]
                    .iter()
                    .position(|&t| t == stop_token)
                    .map_or(sequence.len(), |i| offset + i);
                (sequence[offset..end].to_vec(), *score)
            })
            .collect()
    }
}

/// Upstream's `BeamSearchDecoder`.
#[derive(Debug, Clone)]
pub struct WhisperBeamSearchDecoder {
    k: usize,
    stop_token: i64,
    max_candidates: usize,
    /// One per audio, once the first update has said how many there are.
    finished: Option<Vec<Finished>>,
}

impl WhisperBeamSearchDecoder {
    /// # Arguments
    /// * `beam_size` - beams per audio; at least 1.
    /// * `eot` - the stop token.
    /// * `patience` - how many finished candidates to collect before stopping,
    ///   as a multiple of the beam size; `None` is 1.
    ///
    /// # Panics
    /// If `round(beam_size * patience)` is zero.
    pub fn new(
        beam_size: usize,
        stop_token: i64,
        patience: Option<f64>,
    ) -> Self {
        assert!(beam_size >= 1, "a beam search needs at least one beam");
        let patience = patience.unwrap_or(1.0);
        let max_candidates = (beam_size as f64 * patience).round() as usize;
        assert!(
            max_candidates > 0,
            "invalid beam size ({beam_size}) or patience ({patience})"
        );
        Self {
            k: beam_size,
            stop_token,
            max_candidates,
            finished: None,
        }
    }

    /// Beams per audio.
    pub fn beam_size(&self) -> usize {
        self.k
    }

    /// Finished candidates collected per audio before the search stops.
    pub fn max_candidates(&self) -> usize {
        self.max_candidates
    }
}

impl<B: Backend> TokenDecoder<B> for WhisperBeamSearchDecoder {
    fn group_size(&self) -> usize {
        self.k
    }

    fn reset(&mut self) {
        self.finished = None;
    }

    fn update(
        &mut self,
        tokens: &mut Vec<Vec<i64>>,
        logits: Tensor<B, 2>,
        sum_logprobs: &mut [f32],
        reorder: &mut dyn FnMut(&[usize]),
    ) -> (Vec<i64>, bool) {
        let k = self.k;
        let rows = tokens.len();
        assert_eq!(
            rows % k,
            0,
            "{rows} rows is not a whole number of {k}-beam groups"
        );
        let n_audio = rows / k;
        let finished = self
            .finished
            .get_or_insert_with(|| vec![Finished::default(); n_audio]);

        // Each row's k + 1 best next tokens, with their log probabilities.
        let (values, indices) = log_softmax(logits, 1).topk_with_indices(k + 1, 1);
        let values: Vec<f32> = values.into_data().convert::<f32>().to_vec().unwrap();
        let indices: Vec<i64> = indices.into_data().convert::<i64>().to_vec().unwrap();

        let mut next: Vec<Vec<i64>> = Vec::with_capacity(rows);
        let mut next_sums: Vec<f32> = Vec::with_capacity(rows);
        let mut sources: Vec<usize> = Vec::with_capacity(rows);
        let mut newly_finished: Vec<Finished> = Vec::with_capacity(n_audio);

        for audio in 0..n_audio {
            // STEP 1: every candidate, deduplicated by full sequence. A
            // repeat keeps its first position and takes the later score and
            // source, as assigning into a dict does.
            let mut candidates: Vec<(Vec<i64>, f32, usize)> = Vec::with_capacity(k * (k + 1));
            let mut position: HashMap<Vec<i64>, usize> = HashMap::with_capacity(k * (k + 1));
            for beam in 0..k {
                let row = audio * k + beam;
                for c in 0..=k {
                    let token = indices[row * (k + 1) + c];
                    let score = sum_logprobs[row] + values[row * (k + 1) + c];
                    let mut sequence = tokens[row].clone();
                    sequence.push(token);
                    match position.get(&sequence) {
                        Some(&at) => {
                            candidates[at].1 = score;
                            candidates[at].2 = row;
                        }
                        None => {
                            position.insert(sequence.clone(), candidates.len());
                            candidates.push((sequence, score, row));
                        }
                    }
                }
            }

            // STEP 2: rank, and keep the best k that did not end. Stable, so
            // equals keep their first-seen order, as a sorted dict does.
            let mut order: Vec<usize> = (0..candidates.len()).collect();
            order.sort_by(|&a, &b| candidates[b].1.total_cmp(&candidates[a].1));

            let mut done = Finished::default();
            let mut saved = 0;
            for &c in &order {
                let (sequence, score, source) = &candidates[c];
                if sequence.last() == Some(&self.stop_token) {
                    done.score_sequence(sequence.clone(), *score);
                } else {
                    next_sums.push(*score);
                    next.push(sequence.clone());
                    sources.push(*source);
                    saved += 1;
                    if saved == k {
                        break;
                    }
                }
            }
            assert_eq!(saved, k, "fewer than {k} live candidates for audio {audio}");
            newly_finished.push(done);
        }

        reorder(&sources);
        *tokens = next;
        sum_logprobs.copy_from_slice(&next_sums);

        // Newly finished sequences join the audio's set, best first, until
        // it is full.
        for (previously, newly) in finished.iter_mut().zip(newly_finished) {
            let mut entries = newly.entries;
            entries.sort_by(|a, b| b.1.total_cmp(&a.1));
            for (sequence, score) in entries {
                if previously.len() >= self.max_candidates {
                    break;
                }
                previously.score_sequence(sequence, score);
            }
        }

        let completed = finished.iter().all(|f| f.len() >= self.max_candidates);
        let feed = tokens
            .iter()
            .map(|s| *s.last().expect("non-empty"))
            .collect();
        (feed, completed)
    }

    fn finalize(
        &mut self,
        tokens: Vec<Vec<i64>>,
        sum_logprobs: Vec<f32>,
        prompt_len: usize,
    ) -> Vec<Vec<(Vec<i64>, f32)>> {
        let k = self.k;
        let n_audio = tokens.len() / k;
        let mut finished = self
            .finished
            .take()
            .unwrap_or_else(|| vec![Finished::default(); n_audio]);

        for (audio, set) in finished.iter_mut().enumerate() {
            if set.len() < k {
                // Not enough finished: take the live beams, best first,
                // ended with the stop token.
                let mut order: Vec<usize> = (0..k).collect();
                order.sort_by(|&a, &b| {
                    sum_logprobs[audio * k + b].total_cmp(&sum_logprobs[audio * k + a])
                });
                for beam in order {
                    let row = audio * k + beam;
                    let mut sequence = tokens[row].clone();
                    sequence.push(self.stop_token);
                    set.score_sequence(sequence, sum_logprobs[row]);
                    if set.len() >= k {
                        break;
                    }
                }
            }
        }

        finished
            .into_iter()
            .map(|f| f.clip_tails(prompt_len, self.stop_token))
            .collect()
    }
}

#[cfg(test)]
mod tests {
    use burn::prelude::TensorData;

    use super::*;
    use crate::support::testing::PerformanceBackend;

    type B = PerformanceBackend;

    fn logits<B: Backend>(
        rows: &[&[f32]],
        device: &B::Device,
    ) -> Tensor<B, 2> {
        let vocab = rows[0].len();
        let flat: Vec<f32> = rows.iter().flat_map(|r| r.iter().copied()).collect();
        Tensor::from_data(TensorData::new(flat, [rows.len(), vocab]), device)
    }

    const EOT: i64 = 0;

    /// At the first step every beam is the same prompt, so the candidates
    /// must be deduplicated by sequence: three beams become the three best
    /// distinct tokens, not three copies of the best one, and the cache is
    /// told which rows they came from.
    #[test]
    fn test_first_step_deduplicates() {
        let device = Default::default();

        let mut decoder = WhisperBeamSearchDecoder::new(3, EOT, None);
        let mut tokens = vec![vec![7]; 3];
        let mut sums = vec![0.0; 3];
        let mut sources = Vec::new();
        let mut reorder = |s: &[usize]| sources = s.to_vec();

        // Vocabulary of 5; token 4 best, then 3, then 2.
        let row: &[f32] = &[-9.0, -9.0, 1.0, 2.0, 3.0];
        let (feed, done) = TokenDecoder::<B>::update(
            &mut decoder,
            &mut tokens,
            logits::<B>(&[row, row, row], &device),
            &mut sums,
            &mut reorder,
        );
        assert_eq!(tokens, vec![vec![7, 4], vec![7, 3], vec![7, 2]]);
        assert_eq!(feed, vec![4, 3, 2]);
        assert!(!done);
        // A repeated candidate takes the last source that proposed it.
        assert_eq!(sources, vec![2, 2, 2]);
        assert!(sums[0] > sums[1] && sums[1] > sums[2]);
    }

    /// A candidate that ends goes to the finished set rather than the beam;
    /// the beam refills from the runner-up; the search completes when the
    /// set holds `round(k * patience)` sequences.
    #[test]
    fn test_finished_set_and_patience() {
        let device = Default::default();
        let mut decoder = WhisperBeamSearchDecoder::new(2, EOT, None);
        assert_eq!(decoder.beam_size(), 2);
        assert_eq!(decoder.max_candidates(), 2);
        let mut tokens = vec![vec![7]; 2];
        let mut sums = vec![0.0; 2];
        let mut reorder = |_: &[usize]| {};

        // Best is the stop token, then 3, then 2.
        let row: &[f32] = &[3.0, -9.0, 1.0, 2.0];
        let (feed, done) = TokenDecoder::<B>::update(
            &mut decoder,
            &mut tokens,
            logits::<B>(&[row, row], &device),
            &mut sums,
            &mut reorder,
        );
        assert_eq!(tokens, vec![vec![7, 3], vec![7, 2]]);
        assert_eq!(feed, vec![3, 2]);
        assert!(!done, "one finished, two wanted");

        // Both live beams end now: the set fills and the search completes.
        let (_, done) = TokenDecoder::<B>::update(
            &mut decoder,
            &mut tokens,
            logits::<B>(&[row, row], &device),
            &mut sums,
            &mut reorder,
        );
        assert!(done);

        // Finalize strips the prompt and the stop token; the best finished
        // candidate is the one that ended first.
        let out = TokenDecoder::<B>::finalize(&mut decoder, tokens, sums, 1);
        assert_eq!(out.len(), 1);
        assert_eq!(out[0].len(), 2);
        assert_eq!(out[0][0].0, Vec::<i64>::new());
        assert_eq!(out[0][1].0, vec![3]);
        assert!(out[0][0].1 > out[0][1].1);
    }

    /// With more patience the set is larger, and finalize fills a short
    /// set from the live beams, best first.
    #[test]
    fn test_patience_and_finalize_fill() {
        let device = Default::default();
        let mut decoder = WhisperBeamSearchDecoder::new(2, EOT, Some(2.0));
        assert_eq!(decoder.max_candidates(), 4);
        let mut tokens = vec![vec![7]; 2];
        let mut sums = vec![0.0; 2];
        let mut reorder = |_: &[usize]| {};

        // Nothing ends.
        let row: &[f32] = &[-9.0, -9.0, 1.0, 2.0];
        let (_, done) = TokenDecoder::<B>::update(
            &mut decoder,
            &mut tokens,
            logits::<B>(&[row, row], &device),
            &mut sums,
            &mut reorder,
        );
        assert!(!done);
        assert_eq!(tokens, vec![vec![7, 3], vec![7, 2]]);

        let out = TokenDecoder::<B>::finalize(&mut decoder, tokens, sums, 1);
        assert_eq!(
            out[0].len(),
            2,
            "filled up to the beam size from live beams"
        );
        assert_eq!(out[0][0].0, vec![3], "the better live beam first");
        assert_eq!(out[0][1].0, vec![2]);
    }

    /// Two audios are independent groups: sources index the whole batch,
    /// and each group refills from its own candidates.
    #[test]
    fn test_groups_are_independent() {
        let device = Default::default();
        let mut decoder = WhisperBeamSearchDecoder::new(2, EOT, None);
        let mut tokens = vec![vec![7]; 4];
        let mut sums = vec![0.0; 4];
        let mut sources = Vec::new();
        let mut reorder = |s: &[usize]| sources = s.to_vec();

        let a: &[f32] = &[-9.0, -9.0, 1.0, 2.0];
        let b: &[f32] = &[-9.0, 2.0, 1.0, -9.0];
        let (feed, _) = TokenDecoder::<B>::update(
            &mut decoder,
            &mut tokens,
            logits::<B>(&[a, a, b, b], &device),
            &mut sums,
            &mut reorder,
        );
        assert_eq!(feed, vec![3, 2, 1, 2]);
        assert_eq!(sources, vec![1, 1, 3, 3]);
    }

    #[test]
    #[should_panic(expected = "invalid beam size")]
    fn test_rejects_zero_candidates() {
        let _ = WhisperBeamSearchDecoder::new(1, EOT, Some(0.1));
    }
}