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
//! Turning a row of logits into the next token id.
//!
//! The model produces logits; choosing from them is a separate decision with
//! its own knobs, and every one of them is a way of throwing away part of the
//! distribution. Greedy decoding keeps only the argmax, temperature reshapes
//! what is left, and the two truncations cut the tail that a model with a
//! large vocabulary spreads real probability mass across.
use rand::rngs::StdRng;
use rand::{Rng, SeedableRng};
/// How the next token is drawn from a row of logits.
///
/// Built by [`Sampler::greedy`] or [`Sampler::temperature`] and narrowed with
/// the chained methods:
///
/// ```
/// # use rusting_brain::Sampler;
/// let mut sampler = Sampler::temperature(0.8, Some(42))
/// .top_k(40)
/// .top_p(0.95)
/// .repetition_penalty(1.1);
/// let next = sampler.pick(&[0.1, 3.0, 0.2], &[]);
/// ```
#[derive(Clone, Debug)]
pub struct Sampler {
temperature: f32,
top_k: usize,
top_p: f32,
min_p: f32,
repetition_penalty: f32,
presence_penalty: f32,
frequency_penalty: f32,
rng: StdRng,
}
impl Sampler {
/// Always takes the highest-scoring token. Reproducible by construction,
/// and the right default for anything being measured rather than read.
pub fn greedy() -> Self {
Self {
temperature: 0.0,
top_k: 0,
top_p: 1.0,
min_p: 0.0,
repetition_penalty: 1.0,
presence_penalty: 0.0,
frequency_penalty: 0.0,
rng: StdRng::seed_from_u64(0),
}
}
/// Draws from the distribution after dividing the logits by `temperature`:
/// below one sharpens it, above one flattens it, and zero is
/// [`Sampler::greedy`].
///
/// `seed` of `None` takes one from the operating system, so two runs of the
/// same prompt differ.
pub fn temperature(temperature: f32, seed: Option<u64>) -> Self {
Self {
temperature,
rng: seed.map_or_else(StdRng::from_entropy, StdRng::seed_from_u64),
..Self::greedy()
}
}
/// Keeps only the `k` highest-scoring tokens. Zero keeps all of them.
pub fn top_k(mut self, k: usize) -> Self {
self.top_k = k;
self
}
/// Keeps the smallest set of tokens whose probabilities reach `p`, so the
/// number kept follows how confident the model is. One keeps all of them.
pub fn top_p(mut self, p: f32) -> Self {
self.top_p = p;
self
}
/// Keeps only the tokens at least `p` times as likely as the most likely
/// one, after the temperature, so the cut follows how confident the model
/// is from the top down. Zero keeps all of them.
pub fn min_p(mut self, p: f32) -> Self {
self.min_p = p;
self
}
/// Subtracts `penalty` once from the logit of every token already in the
/// history, however often it appears there.
pub fn presence_penalty(mut self, penalty: f32) -> Self {
self.presence_penalty = penalty;
self
}
/// Subtracts `penalty` from a token's logit for every time it appears in
/// the history.
pub fn frequency_penalty(mut self, penalty: f32) -> Self {
self.frequency_penalty = penalty;
self
}
/// Divides the logit of every token already in the history by `penalty`,
/// which is how a small model is stopped from repeating one phrase for as
/// long as it is allowed to run. One leaves the logits alone.
pub fn repetition_penalty(mut self, penalty: f32) -> Self {
self.repetition_penalty = penalty;
self
}
/// The next token id, given one row of logits and the ids generated so far.
///
/// `history` is read only by [`Sampler::repetition_penalty`] and can be
/// empty otherwise.
///
/// # Panics
///
/// If `logits` is empty: there is no token to return.
pub fn pick(&mut self, logits: &[f32], history: &[u32]) -> u32 {
let (ranked, weights) = self.kept(logits, history);
if weights.len() == 1 {
return ranked[0].0 as u32;
}
let total: f32 = weights.iter().sum();
let mut threshold = self.rng.gen_range(0.0..total);
for (&(id, _), &weight) in ranked.iter().zip(&weights) {
threshold -= weight;
if threshold <= 0.0 {
return id as u32;
}
}
ranked[0].0 as u32
}
/// The distribution [`pick`](Sampler::pick) draws from, over the whole
/// vocabulary: zero outside what the truncations keep, and all of it on
/// the argmax for a greedy sampler.
pub(crate) fn probabilities(&self, logits: &[f32], history: &[u32]) -> Vec<f32> {
let (ranked, weights) = self.kept(logits, history);
let total: f32 = weights.iter().sum();
let mut probabilities = vec![0.0; logits.len()];
for (&(id, _), &weight) in ranked.iter().zip(&weights) {
probabilities[id] = weight / total;
}
probabilities
}
/// An id drawn in proportion to `weights`, which need not sum to one.
pub(crate) fn draw(&mut self, weights: &[f32]) -> u32 {
let total: f32 = weights.iter().sum();
let mut threshold = self.rng.gen_range(0.0..total);
for (id, &weight) in weights.iter().enumerate() {
threshold -= weight;
if weight > 0.0 && threshold <= 0.0 {
return id as u32;
}
}
weights
.iter()
.rposition(|&weight| weight > 0.0)
.unwrap_or(0) as u32
}
/// A uniform draw from `[0, 1)`.
pub(crate) fn uniform(&mut self) -> f32 {
self.rng.gen_range(0.0..1.0)
}
/// The candidates in descending order and their unnormalised weights.
/// A single candidate means there is nothing to draw.
fn kept(&self, logits: &[f32], history: &[u32]) -> (Vec<(usize, f32)>, Vec<f32>) {
assert!(!logits.is_empty(), "cannot sample from an empty vocabulary");
let mut scores: Vec<f32> = logits.to_vec();
if self.repetition_penalty != 1.0 {
for &id in history {
if let Some(score) = scores.get_mut(id as usize) {
// A negative logit is made *more* negative, so the penalty
// pushes a token down whichever side of zero it sits on.
*score = if *score > 0.0 {
*score / self.repetition_penalty
} else {
*score * self.repetition_penalty
};
}
}
}
if self.presence_penalty != 0.0 || self.frequency_penalty != 0.0 {
let mut counts = vec![0u32; scores.len()];
for &id in history {
if let Some(count) = counts.get_mut(id as usize) {
*count += 1;
}
}
for (score, &count) in scores.iter_mut().zip(&counts) {
if count > 0 {
*score -= self.presence_penalty + self.frequency_penalty * count as f32;
}
}
}
let mut ranked: Vec<(usize, f32)> = scores.into_iter().enumerate().collect();
// `total_cmp` rather than `partial_cmp`: a NaN logit means the model has
// diverged, and an ordering that panics or silently shuffles makes that
// harder to see than a token chosen from a defined order.
let descending = |a: &(usize, f32), b: &(usize, f32)| b.1.total_cmp(&a.1);
if self.temperature <= 0.0 {
let mut best = 0;
for (index, &(_, score)) in ranked.iter().enumerate() {
if score.total_cmp(&ranked[best].1).is_gt() {
best = index;
}
}
return (vec![ranked[best]], vec![1.0]);
}
// Sorting a 32000-token vocabulary costs more than some models spend
// producing it. With a top-k, only those k entries have to be in order,
// and a linear partition finds them.
if self.top_k > 0 && self.top_k < ranked.len() {
ranked.select_nth_unstable_by(self.top_k - 1, descending);
ranked.truncate(self.top_k);
}
ranked.sort_unstable_by(descending);
// Subtracting the maximum before `exp` is not a nicety: logits reach 30
// or more, `exp(30)` overflows to infinity, and every weight becomes
// NaN.
let max = ranked[0].1;
let mut weights: Vec<f32> = ranked
.iter()
.map(|&(_, score)| ((score - max) / self.temperature).exp())
.collect();
// The leading weight is `exp(0)`, one, so a weight is already its
// token's probability relative to the top. The argmax always stays.
if self.min_p > 0.0 {
let kept = weights
.iter()
.position(|&weight| weight < self.min_p)
.map_or(weights.len(), |index| index.max(1));
weights.truncate(kept);
}
let total: f32 = weights.iter().sum();
if self.top_p < 1.0 && total > 0.0 {
let mut cumulative = 0.0;
// The token that crosses the threshold is kept, so a distribution
// whose first token already exceeds `top_p` still leaves one
// candidate rather than none.
let kept = weights
.iter()
.position(|weight| {
cumulative += weight / total;
cumulative >= self.top_p
})
.map_or(weights.len(), |index| index + 1);
weights.truncate(kept);
}
let total: f32 = weights.iter().sum();
// NaN as well as zero: a diverged model gives `gen_range` an empty or
// undefined range, which panics.
if total <= 0.0 || !total.is_finite() {
return (vec![ranked[0]], vec![1.0]);
}
ranked.truncate(weights.len());
(ranked, weights)
}
}
#[cfg(test)]
mod tests {
use super::*;
const LOGITS: [f32; 5] = [0.5, 3.0, 1.0, -2.0, 0.25];
#[test]
fn greedy_takes_the_argmax() {
assert_eq!(Sampler::greedy().pick(&LOGITS, &[]), 1);
}
#[test]
fn top_k_of_one_is_greedy() {
let mut sampler = Sampler::temperature(5.0, Some(7)).top_k(1);
for _ in 0..32 {
assert_eq!(sampler.pick(&LOGITS, &[]), 1);
}
}
#[test]
fn top_p_keeps_the_token_that_crosses_the_threshold() {
// Any `top_p` at or below the leading token's own probability has to
// leave exactly that token, never an empty candidate set.
let mut sampler = Sampler::temperature(1.0, Some(7)).top_p(0.01);
for _ in 0..32 {
assert_eq!(sampler.pick(&LOGITS, &[]), 1);
}
}
#[test]
fn the_same_seed_draws_the_same_tokens() {
let draw = || {
let mut sampler = Sampler::temperature(1.5, Some(99));
(0..64)
.map(|_| sampler.pick(&LOGITS, &[]))
.collect::<Vec<_>>()
};
assert_eq!(draw(), draw());
}
#[test]
fn temperature_widens_the_spread() {
let distinct = |temperature: f32| {
let mut sampler = Sampler::temperature(temperature, Some(5));
let mut seen = (0..256)
.map(|_| sampler.pick(&LOGITS, &[]))
.collect::<Vec<_>>();
seen.sort_unstable();
seen.dedup();
seen.len()
};
assert!(distinct(0.1) < distinct(5.0));
}
#[test]
fn the_repetition_penalty_pushes_a_repeated_token_down() {
// The penalty has to beat a 2.0 logit gap, so greedy moves off token 1.
let mut sampler = Sampler::greedy().repetition_penalty(4.0);
assert_eq!(sampler.pick(&LOGITS, &[1]), 2);
// A negative logit is pushed further down, never up past its peers.
let mut sampler = Sampler::greedy().repetition_penalty(4.0);
assert_eq!(sampler.pick(&LOGITS, &[3]), 1);
}
#[test]
fn min_p_keeps_the_tokens_at_or_above_its_share_of_the_top() {
// Relative to token 1, token 2 is at exp(-2) = 0.135, token 0 at
// 0.082, token 4 at 0.064 and token 3 at 0.0067.
let sampler = Sampler::temperature(1.0, Some(1)).min_p(0.1);
let kept: Vec<bool> = sampler
.probabilities(&LOGITS, &[])
.iter()
.map(|&p| p > 0.0)
.collect();
assert_eq!(kept, [false, true, true, false, false]);
// A threshold above every ratio still leaves the argmax.
let sampler = Sampler::temperature(1.0, Some(1)).min_p(2.0);
assert_eq!(sampler.probabilities(&LOGITS, &[])[1], 1.0);
}
#[test]
fn presence_counts_once_and_frequency_per_occurrence() {
// Token 1 leads token 2 by 2.0. Seen twice, a presence penalty of 1.5
// leaves it ahead; a frequency penalty of 1.5 takes 3.0 and does not.
let history = [1, 1];
let mut sampler = Sampler::greedy().presence_penalty(1.5);
assert_eq!(sampler.pick(&LOGITS, &history), 1);
let mut sampler = Sampler::greedy().frequency_penalty(1.5);
assert_eq!(sampler.pick(&LOGITS, &history), 2);
let mut sampler = Sampler::greedy().presence_penalty(1.5);
assert_eq!(sampler.pick(&LOGITS, &[1, 2]), 1);
}
#[test]
fn a_diverged_row_returns_a_token_rather_than_panicking() {
let mut sampler = Sampler::temperature(1.0, Some(3)).top_k(2);
let id = sampler.pick(&[f32::NAN, 1.0, f32::NAN], &[]);
assert!(id < 3);
}
#[test]
fn an_out_of_range_history_id_is_ignored() {
let mut sampler = Sampler::greedy().repetition_penalty(4.0);
assert_eq!(sampler.pick(&LOGITS, &[900]), 1);
}
}