fastnn 0.3.3

A robust deep learning library in Rust with CUDA kernels
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

//! Cached decoding against the plain forward, plus sampling and beam search.

use fastnn::prelude::*;

/// Eval mode, so dropout is identity on both paths.
fn stack() -> TransformerStack {
    manual_seed(7);
    let stack = TransformerStack::causal(32, 4, 64, 2, 0.1);
    stack.set_training(false);
    stack
}

const TOLERANCE: f32 = 1e-3;

fn assert_close(a: &Tensor, b: &Tensor, what: &str) {
    let (a, b) = (a.to_vec(), b.to_vec());
    assert_eq!(a.len(), b.len(), "{what}: lengths differ");
    for (i, (x, y)) in a.iter().zip(&b).enumerate() {
        assert!(
            (x - y).abs() < TOLERANCE,
            "{what}: element {i} differs, {x} vs {y}"
        );
    }
}

#[test]
fn cached_decoding_matches_full_forward() {
    let stack = stack();
    let input = Tensor::randn(&[1, 6, 32]);

    let full = no_grad(|| stack.forward(&input));

    let mut cache = stack.new_cache();
    for t in 0..6 {
        let step = no_grad(|| stack.forward_cached(&input.narrow(1, t, 1), &mut cache));
        assert_close(&step, &full.narrow(1, t, 1), &format!("position {t}"));
    }
    assert_eq!(cache.len(), 6);
}

/// Prefill then single steps, the way a prompt is actually consumed.
#[test]
fn prefill_then_single_steps_matches_full_forward() {
    let stack = stack();
    let input = Tensor::randn(&[2, 5, 32]);

    let full = no_grad(|| stack.forward(&input));

    let mut cache = stack.new_cache();
    let prefill = no_grad(|| stack.forward_cached(&input.narrow(1, 0, 3), &mut cache));
    assert_close(&prefill, &full.narrow(1, 0, 3), "prefill");

    for t in 3..5 {
        let step = no_grad(|| stack.forward_cached(&input.narrow(1, t, 1), &mut cache));
        assert_close(&step, &full.narrow(1, t, 1), &format!("position {t}"));
    }
}

#[test]
fn cleared_cache_starts_over() {
    let stack = stack();
    let input = Tensor::randn(&[1, 4, 32]);

    let mut cache = stack.new_cache();
    let first = no_grad(|| stack.forward_cached(&input, &mut cache));

    cache.clear();
    assert!(cache.is_empty());
    let second = no_grad(|| stack.forward_cached(&input, &mut cache));
    assert_close(&first, &second, "after clear");
}

#[test]
fn greedy_picks_the_largest_logit() {
    let logits = Tensor::from_vec(vec![0.1, 3.0, -2.0, 1.5], &[4]);
    assert_eq!(Sampler::greedy().sample(&logits), 1);
    assert_eq!(Sampler::new().temperature(0.0).sample(&logits), 1);
}

#[test]
fn top_k_of_one_is_greedy_at_any_temperature() {
    let logits = Tensor::from_vec(vec![0.1, 3.0, -2.0, 1.5], &[4]);
    let sampler = Sampler::new().temperature(5.0).top_k(1);
    for _ in 0..20 {
        assert_eq!(sampler.sample(&logits), 1);
    }
}

#[test]
fn tiny_top_p_is_greedy() {
    // The top token alone already exceeds p.
    let logits = Tensor::from_vec(vec![0.1, 3.0, -2.0, 1.5], &[4]);
    let sampler = Sampler::new().top_p(0.01);
    for _ in 0..20 {
        assert_eq!(sampler.sample(&logits), 1);
    }
}

#[test]
fn sampling_is_reproducible_under_manual_seed() {
    let logits = Tensor::from_vec(vec![1.0, 1.1, 0.9, 1.05], &[4]);
    let sampler = Sampler::new().temperature(0.9).top_k(3);

    manual_seed(42);
    let first: Vec<usize> = (0..10).map(|_| sampler.sample(&logits)).collect();
    manual_seed(42);
    let second: Vec<usize> = (0..10).map(|_| sampler.sample(&logits)).collect();
    assert_eq!(first, second);
}

#[test]
fn sampling_respects_the_distribution() {
    // Anything else is about 2e-9 likely.
    let logits = Tensor::from_vec(vec![10.0, -10.0, -10.0], &[3]);
    let sampler = Sampler::new();
    for _ in 0..100 {
        assert_eq!(sampler.sample(&logits), 0);
    }
}

/// Changing a padded position must not change the unpadded outputs.
#[test]
fn padding_mask_blocks_padded_positions() {
    let encoder = TransformerStack::encoder(16, 2, 32, 1, 0.1);
    encoder.set_training(false);
    let mask = Tensor::from_vec(vec![1.0, 1.0, 1.0, 0.0], &[1, 4]);

    let a = Tensor::randn(&[1, 4, 16]);
    let mut data = a.to_vec();
    for slot in data.iter_mut().skip(3 * 16) {
        *slot += 5.0;
    }
    let b = Tensor::from_vec(data, &[1, 4, 16]);

    let out_a = no_grad(|| encoder.forward_masked(&a, Some(&mask)));
    let out_b = no_grad(|| encoder.forward_masked(&b, Some(&mask)));
    assert_close(
        &out_a.narrow(1, 0, 3),
        &out_b.narrow(1, 0, 3),
        "padded position leaked into real ones",
    );
}

/// Same through the decoder's cross-attention.
#[test]
fn memory_mask_blocks_padded_source_positions() {
    let decoder = TransformerDecoder::new(16, 2, 32, 1, 0.1);
    decoder.set_training(false);
    let mask = Tensor::from_vec(vec![1.0, 1.0, 0.0], &[1, 3]);
    let target = Tensor::randn(&[1, 4, 16]);

    let memory_a = Tensor::randn(&[1, 3, 16]);
    let mut data = memory_a.to_vec();
    for slot in data.iter_mut().skip(2 * 16) {
        *slot -= 7.0;
    }
    let memory_b = Tensor::from_vec(data, &[1, 3, 16]);

    let out_a = no_grad(|| decoder.decode_masked(&target, &memory_a, Some(&mask)));
    let out_b = no_grad(|| decoder.decode_masked(&target, &memory_b, Some(&mask)));
    assert_close(
        &out_a,
        &out_b,
        "masked memory leaked through cross-attention",
    );
}

/// The best path can start with the second-best token.
#[test]
fn beam_search_finds_what_greedy_misses() {
    let step = |seq: &[usize]| match seq.last() {
        // Token 0 is slightly likelier than token 1 at the start.
        Some(9) => vec![0.55f32.ln(), 0.45f32.ln(), f32::NEG_INFINITY],
        // After the greedy pick everything is mediocre.
        Some(0) => vec![0.5f32.ln(), 0.5f32.ln(), f32::NEG_INFINITY],
        // After the runner-up one continuation is near-certain.
        _ => vec![0.99f32.ln(), 0.01f32.ln(), f32::NEG_INFINITY],
    };

    let greedy = BeamSearch::new(1).decode(&[9], 2, step);
    let beamed = BeamSearch::new(2).decode(&[9], 2, step);
    assert_eq!(greedy[0], 0, "width 1 should behave greedily");
    assert_eq!(
        beamed,
        vec![1, 0],
        "width 2 should find the higher-probability path"
    );
}

#[test]
fn beam_search_stops_at_eos() {
    // Token 2 is eos and always the most likely.
    let step = |_: &[usize]| vec![0.0, 1.0, 5.0];
    let out = BeamSearch::new(2).eos(2).decode(&[0], 10, step);
    assert!(
        out.is_empty(),
        "eos as first token should end generation, got {out:?}"
    );
}

#[test]
fn repetition_penalty_discourages_used_tokens() {
    let logits = Tensor::from_vec(vec![1.0, 1.2], &[2]);
    let sampler = Sampler::greedy().repetition_penalty(2.0);

    assert_eq!(sampler.sample(&logits), 1);
    assert_eq!(sampler.sample_with_history(&logits, &[1]), 0);
}