ruda-tensor 0.21.49

Ruda tensor operations, backend contracts, and framework-facing primitives.
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

use crate::{Backend, DType, FloatDType, TensorMetadata, tensor::FloatTensor};
use ruda_core::tensor::Shape;
use super::{LayerNormBackward, LayerNormOutput, RmsNormBackward, RmsNormOutput};

/// Last-axis RMSNorm retaining FP32/FP64 row statistics on the current backend.
pub fn rms_norm_with_stats<B: Backend>(input: FloatTensor<B>, gamma: FloatTensor<B>, epsilon: f64) -> RmsNormOutput<B> {
    let shape = input.shape();
    let width = *shape.last().expect("RMSNorm requires an axis");
    assert!(width > 0, "RMSNorm final axis must be nonempty");
    assert_eq!(gamma.shape(), Shape::new([width]), "RMSNorm weight shape differs");
    assert!(epsilon.is_finite() && epsilon > 0.0, "RMSNorm epsilon must be finite and positive");
    let rows = shape.num_elements() / width;
    let storage: FloatDType = input.dtype().into();
    let compute = if input.dtype() == DType::F64 { FloatDType::F64 } else { FloatDType::F32 };
    if rows == 0 {
        let device = B::float_device(&input);
        return RmsNormOutput { output: input, rstd: B::float_zeros(Shape::new([0]), &device, compute) };
    }
    let input = B::float_reshape(B::float_cast(input, compute), Shape::new([rows, width]));
    let square_mean = B::float_mean_dim(B::float_mul(input.clone(), input.clone()), 1);
    let rstd = B::float_recip(B::float_sqrt(B::float_add_scalar(square_mean, epsilon.into())));
    let gamma = B::float_reshape(B::float_cast(gamma, compute), Shape::new([1, width]));
    let output = B::float_mul(B::float_mul(input, rstd.clone()), gamma);
    RmsNormOutput { output: B::float_reshape(B::float_cast(output, storage), shape),
        rstd: B::float_reshape(rstd, Shape::new([rows])) }
}

/// RMSNorm derivatives from saved reciprocal norms and the original working values.
pub fn rms_norm_backward<B: Backend>(input: FloatTensor<B>, gamma: FloatTensor<B>, grad: FloatTensor<B>,
    rstd: FloatTensor<B>) -> RmsNormBackward<B> {
    let [input, weight] = rms_norm_backward_select::<B>(input, gamma, grad, rstd, [true; 2]);
    RmsNormBackward { input: input.expect("requested RMSNorm input gradient"),
        weight: weight.expect("requested RMSNorm weight gradient") }
}

/// Compute only requested input and weight derivatives on the current backend.
pub fn rms_norm_backward_select<B: Backend>(input: FloatTensor<B>, gamma: FloatTensor<B>, grad: FloatTensor<B>,
    rstd: FloatTensor<B>, mask: [bool; 2]) -> [Option<FloatTensor<B>>; 2] {
    if mask == [false; 2] { return [None, None]; }
    let shape = input.shape();
    let width = *shape.last().expect("RMSNorm requires an axis");
    assert!(width > 0, "RMSNorm final axis must be nonempty");
    let rows = shape.num_elements() / width;
    assert_eq!(gamma.shape(), Shape::new([width]), "RMSNorm weight shape differs");
    assert_eq!(grad.shape(), shape, "RMSNorm gradient shape differs");
    assert_eq!(rstd.shape(), Shape::new([rows]), "RMSNorm reciprocal norm shape differs");
    let storage: FloatDType = input.dtype().into();
    let weight_storage: FloatDType = gamma.dtype().into();
    let forward_compute = if input.dtype() == DType::F64 { FloatDType::F64 } else { FloatDType::F32 };
    let compute = if [&input, &gamma, &grad, &rstd].iter().any(|value| value.dtype() == DType::F64) {
        FloatDType::F64
    } else { FloatDType::F32 };
    if rows == 0 {
        let device = B::float_device(&input);
        return [mask[0].then_some(input), mask[1].then(|| B::float_zeros(Shape::new([width]), &device, weight_storage))];
    }
    let input = B::float_reshape(B::float_cast(input, forward_compute), Shape::new([rows, width]));
    let rstd = B::float_reshape(B::float_cast(rstd, forward_compute), Shape::new([rows, 1]));
    let normalized = B::float_cast(B::float_mul(input, rstd.clone()), compute);
    let grad = B::float_reshape(B::float_cast(grad, compute), Shape::new([rows, width]));
    let input_grad = mask[0].then(|| {
        let rstd = B::float_cast(rstd, compute);
        let gamma = B::float_reshape(B::float_cast(B::float_cast(gamma, forward_compute), compute), Shape::new([1, width]));
        let scaled_grad = B::float_mul(grad.clone(), gamma);
        let correction = B::float_mean_dim(B::float_mul(scaled_grad.clone(), normalized.clone()), 1);
        let value = B::float_mul(rstd, B::float_sub(scaled_grad, B::float_mul(normalized.clone(), correction)));
        B::float_reshape(B::float_cast(value, storage), shape)
    });
    let weight_grad = mask[1].then(|| {
        let value = B::float_sum_dim(B::float_mul(grad, normalized), 0);
        B::float_reshape(B::float_cast(value, weight_storage), Shape::new([width]))
    });
    [input_grad, weight_grad]
}

/// Last-axis normalization on the current backend, retaining FP32/FP64 row statistics.
pub fn layer_norm_with_stats<B: Backend>(
    input: FloatTensor<B>, gamma: FloatTensor<B>, beta: Option<FloatTensor<B>>, epsilon: f64,
) -> LayerNormOutput<B> {
    let shape = input.shape();
    let width = *shape.last().expect("LayerNorm input must have an axis");
    assert!(width > 0, "LayerNorm final axis must be nonempty");
    assert_eq!(gamma.shape(), Shape::new([width]), "LayerNorm weight shape differs");
    if let Some(beta) = &beta { assert_eq!(beta.shape(), gamma.shape(), "LayerNorm bias shape differs"); }
    assert!(epsilon.is_finite() && epsilon > 0.0, "LayerNorm epsilon must be finite and positive");
    let rows = shape.num_elements() / width;
    let storage: FloatDType = input.dtype().into();
    let compute = if input.dtype() == DType::F64 { FloatDType::F64 } else { FloatDType::F32 };
    if rows == 0 {
        let device = B::float_device(&input);
        return LayerNormOutput { output: input,
            mean: B::float_zeros(Shape::new([0]), &device, compute),
            rstd: B::float_zeros(Shape::new([0]), &device, compute) };
    }
    let input = B::float_reshape(B::float_cast(input, compute), Shape::new([rows, width]));
    let mean = B::float_mean_dim(input.clone(), 1);
    let centered = B::float_sub(input, mean.clone());
    let variance = B::float_mean_dim(B::float_mul(centered.clone(), centered.clone()), 1);
    let rstd = B::float_recip(B::float_sqrt(B::float_add_scalar(variance, epsilon.into())));
    let gamma = B::float_reshape(B::float_cast(gamma, compute), Shape::new([1, width]));
    let mut output = B::float_mul(B::float_mul(centered, rstd.clone()), gamma);
    if let Some(beta) = beta {
        output = B::float_add(output, B::float_reshape(B::float_cast(beta, compute), Shape::new([1, width])));
    }
    LayerNormOutput {
        output: B::float_reshape(B::float_cast(output, storage), shape),
        mean: B::float_reshape(mean, Shape::new([rows])),
        rstd: B::float_reshape(rstd, Shape::new([rows])),
    }
}

/// Complete first-order derivatives from saved statistics, without repeating forward.
pub fn layer_norm_backward<B: Backend>(
    input: FloatTensor<B>, gamma: FloatTensor<B>, grad: FloatTensor<B>,
    mean: FloatTensor<B>, rstd: FloatTensor<B>,
) -> LayerNormBackward<B> {
    let [input, weight, bias] = layer_norm_backward_select::<B>(input, gamma, grad, mean, rstd, [true; 3]);
    LayerNormBackward { input: input.expect("requested LayerNorm input gradient"),
        weight: weight.expect("requested LayerNorm weight gradient"), bias: bias.expect("requested LayerNorm bias gradient") }
}

/// Compute requested input, weight and bias derivatives without building unrequested graphs.
pub fn layer_norm_backward_select<B: Backend>(input: FloatTensor<B>, gamma: FloatTensor<B>, grad: FloatTensor<B>,
    mean: FloatTensor<B>, rstd: FloatTensor<B>, mask: [bool; 3]) -> [Option<FloatTensor<B>>; 3] {
    if mask == [false; 3] { return [None, None, None]; }
    let shape = input.shape();
    let width = *shape.last().expect("LayerNorm input must have an axis");
    assert!(width > 0, "LayerNorm final axis must be nonempty");
    let rows = shape.num_elements() / width;
    assert_eq!(gamma.shape(), Shape::new([width]), "LayerNorm weight shape differs");
    assert_eq!(grad.shape(), shape, "LayerNorm gradient shape differs");
    assert_eq!(mean.shape(), Shape::new([rows]), "LayerNorm mean shape differs");
    assert_eq!(rstd.shape(), mean.shape(), "LayerNorm reciprocal deviation shape differs");
    let input_storage: FloatDType = input.dtype().into();
    let weight_storage: FloatDType = gamma.dtype().into();
    let forward_compute = if input.dtype() == DType::F64 { FloatDType::F64 } else { FloatDType::F32 };
    let compute = if [&input, &gamma, &grad, &mean, &rstd].iter().any(|value| value.dtype() == DType::F64) {
        FloatDType::F64
    } else { FloatDType::F32 };
    if rows == 0 {
        let device = B::float_device(&input);
        return [mask[0].then_some(input),
            mask[1].then(|| B::float_zeros(Shape::new([width]), &device, weight_storage)),
            mask[2].then(|| B::float_zeros(Shape::new([width]), &device, compute))];
    }
    let grad = B::float_reshape(B::float_cast(grad, compute), Shape::new([rows, width]));
    if !mask[0] && !mask[1] {
        let bias = B::float_reshape(B::float_sum_dim(grad, 0), Shape::new([width]));
        return [None, None, Some(bias)];
    }
    let input = B::float_reshape(B::float_cast(input, forward_compute), Shape::new([rows, width]));
    let mean = B::float_reshape(B::float_cast(mean, forward_compute), Shape::new([rows, 1]));
    let rstd = B::float_reshape(B::float_cast(rstd, forward_compute), Shape::new([rows, 1]));
    let normalized = B::float_cast(B::float_mul(B::float_sub(input, mean), rstd.clone()), compute);
    let input_grad = mask[0].then(|| {
        let rstd = B::float_cast(rstd, compute);
        let gamma = B::float_reshape(B::float_cast(B::float_cast(gamma, forward_compute), compute), Shape::new([1, width]));
        let scaled_grad = B::float_mul(grad.clone(), gamma);
        let average_grad = B::float_mean_dim(scaled_grad.clone(), 1);
        let average_product = B::float_mean_dim(B::float_mul(scaled_grad.clone(), normalized.clone()), 1);
        let value = B::float_mul(rstd, B::float_sub(B::float_sub(scaled_grad, average_grad),
            B::float_mul(normalized.clone(), average_product)));
        B::float_reshape(B::float_cast(value, input_storage), shape)
    });
    let weight_grad = mask[1].then(|| {
        let value = B::float_sum_dim(B::float_mul(grad.clone(), normalized), 0);
        B::float_reshape(B::float_cast(value, weight_storage), Shape::new([width]))
    });
    let bias_grad = mask[2].then(|| B::float_reshape(B::float_sum_dim(grad, 0), Shape::new([width])));
    [input_grad, weight_grad, bias_grad]
}