ruda-nn 0.21.27

Ruda neural network layers, activation modules and losses.
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

use super::*;
use crate::{Linear, LoRALinearConfig, TestAutodiffBackend as B};
use ruda_model::{
    module::Param,
    tensor::{TensorData, Tolerance},
};

#[test]
fn precomputed_logits_preserve_causal_shift_ignore_and_gradient_semantics() {
    let device = Default::default();
    let logits = Tensor::<B, 3>::from_floats(
        [[[1.0, 2.0, -1.0], [0.0, -0.5, 3.0], [1.5, 1.0, 0.0]]],
        &device,
    )
    .require_grad();
    let labels = Tensor::<B, 2, Int>::from_data([[2, 1, -100]], &device);
    let expected = -log_softmax(logits.clone().slice([0..1, 0..1, 0..3]).reshape([1, 3]), 1)
        .slice([0..1, 1..2])
        .sum();
    let expected_data = expected.to_data();
    let expected_grads = expected.backward();
    for chunk in [1, 3] {
        let result = CausalCrossEntropyConfig::new()
            .with_token_chunk_size(chunk)
            .forward_logits(logits.clone(), labels.clone());
        assert_eq!(result.valid_tokens.clone().into_scalar(), 1);
        result
            .mean()
            .to_data()
            .assert_approx_eq::<f32>(&expected_data, Tolerance::absolute(1e-6));
        let grads = result.mean().backward();
        logits
            .grad(&grads)
            .unwrap()
            .to_data()
            .assert_approx_eq::<f32>(
                &logits.grad(&expected_grads).unwrap().to_data(),
                Tolerance::absolute(1e-6),
            );
    }
}

#[test]
fn chunked_causal_loss_and_gradients_match_full_vocabulary_reference() {
    let device = Default::default();
    let head = Linear::<B> {
        weight: Param::from_tensor(Tensor::from_floats(
            [[1.0, 0.0, -1.0], [0.5, -0.5, 0.0]],
            &device,
        )),
        bias: None,
    };
    let hidden = Tensor::<B, 3>::from_floats(
        [
            [[1.0, 2.0], [-1.0, 0.5], [2.0, -1.0]],
            [[0.0, 1.0], [0.5, -0.5], [-1.0, -2.0]],
        ],
        &device,
    )
    .require_grad();
    let labels = Tensor::<B, 2, Int>::from_data([[-100, 2, -100], [0, 1, 0]], &device);
    let full_hidden = hidden.clone().slice([0..2, 0..2, 0..2]).reshape([4, 2]);
    let logp = log_softmax(head.forward(full_hidden), 1);
    let expected = -(logp.clone().slice([0..1, 2..3]).sum()
        + logp.clone().slice([2..3, 1..2]).sum()
        + logp.slice([3..4, 0..1]).sum())
        / 3;
    let expected_grads = expected.clone().backward();
    let data = expected.to_data();
    for chunk in [1, 2, 3, 8] {
        let output = CausalCrossEntropyConfig::new()
            .with_token_chunk_size(chunk)
            .forward_hidden(hidden.clone(), labels.clone(), |rows| head.forward(rows));
        output
            .valid_tokens
            .to_data()
            .assert_eq(&TensorData::from([3_i64]), false);
        let loss = output.mean();
        loss.to_data()
            .assert_approx_eq::<f32>(&data, Tolerance::absolute(1e-6));
        let grads = loss.backward();
        hidden
            .grad(&grads)
            .unwrap()
            .to_data()
            .assert_approx_eq::<f32>(
                &hidden.grad(&expected_grads).unwrap().to_data(),
                Tolerance::absolute(1e-6),
            );
        head.weight
            .val()
            .grad(&grads)
            .unwrap()
            .to_data()
            .assert_approx_eq::<f32>(
                &head.weight.val().grad(&expected_grads).unwrap().to_data(),
                Tolerance::absolute(1e-6),
            );
    }
}

#[test]
fn ignored_short_and_unshifted_sequences_have_defined_losses() {
    let device = Default::default();
    for sequence in [0, 1, 3] {
        let hidden = Tensor::<B, 3>::ones([1, sequence, 2], &device).require_grad();
        let labels = Tensor::<B, 2, Int>::full([1, sequence], -100, &device);
        let result = CausalCrossEntropyConfig::new().forward_hidden(hidden, labels, |rows| rows);
        assert_eq!(result.mean().into_scalar(), 0.0);
        assert_eq!(result.valid_tokens.into_scalar(), 0);
    }
    let result = CausalCrossEntropyConfig::new()
        .with_shift(false)
        .with_ignore_index(-7)
        .forward_hidden(
            Tensor::<B, 3>::zeros([1, 2, 3], &device),
            Tensor::<B, 2, Int>::from_data([[1, -7]], &device),
            |rows| rows,
        );
    assert!((result.loss_sum.into_scalar() - 3.0_f32.ln()).abs() < 1e-6);
    assert_eq!(result.valid_tokens.into_scalar(), 1);
}

#[test]
fn chunked_projection_trains_lora_without_base_gradients() {
    let device = Default::default();
    let base = Linear::<B> {
        weight: Param::from_tensor(Tensor::from_floats(
            [[1.0, 0.0, -1.0], [0.5, -0.5, 0.0]],
            &device,
        )),
        bias: None,
    };
    let head = LoRALinearConfig::new(1, 1.0).init(base);
    let hidden = Tensor::<B, 3>::ones([1, 3, 2], &device).require_grad();
    let result = CausalCrossEntropyConfig::new()
        .with_token_chunk_size(1)
        .forward_hidden(
            hidden.clone(),
            Tensor::<B, 2, Int>::from_data([[0, 1, 2]], &device),
            |rows| head.forward(rows),
        );
    let gradients = result.mean().backward();
    assert!(hidden.grad(&gradients).is_some());
    assert!(head.adapter_b.weight.val().grad(&gradients).is_some());
    assert!(head.base.weight.val().grad(&gradients).is_none());
}