tang 0.2.0

Math library for physical reality — geometry, spatial algebra, tensor, training, GPU compute, and 3D gaussian splatting
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

//! Free-standing tensor operations generic over ComputeDevice.

use super::device::ComputeDevice;
use super::tensor::ComputeTensor;
use crate::Scalar;
use crate::expr::node::ExprId;

/// Element-wise addition of two tensors.
pub fn add_tensors<D: ComputeDevice>(
    dev: &D,
    a: &ComputeTensor<D::Buffer>,
    b: &ComputeTensor<D::Buffer>,
) -> ComputeTensor<D::Buffer> {
    assert_eq!(a.numel(), b.numel(), "add_tensors: numel mismatch");
    let buf = dev.elementwise(
        &[&a.buffer, &b.buffer],
        a.numel(),
        &|ids: &[ExprId]| ids[0] + ids[1],
    );
    ComputeTensor::from_buffer(buf, a.shape().to_vec())
}

/// Broadcast bias addition: out[i] = matrix[i] + bias[i % dim].
///
/// matrix: [n_rows, dim], bias: [dim] → out: [n_rows, dim].
pub fn bias_add<D: ComputeDevice>(
    dev: &D,
    matrix: &ComputeTensor<D::Buffer>,
    bias: &ComputeTensor<D::Buffer>,
) -> ComputeTensor<D::Buffer> {
    let numel = matrix.numel();
    let dim = bias.numel();
    assert_eq!(numel % dim, 0, "bias_add: matrix numel not divisible by bias dim");
    let buf = dev.bias_add(&matrix.buffer, &bias.buffer, numel, dim);
    ComputeTensor::from_buffer(buf, matrix.shape().to_vec())
}

/// SwiGLU activation: silu(gate) * up, where silu(x) = x / (1 + exp(-x)).
pub fn swiglu_fused<D: ComputeDevice>(
    dev: &D,
    gate: &ComputeTensor<D::Buffer>,
    up: &ComputeTensor<D::Buffer>,
) -> ComputeTensor<D::Buffer> {
    assert_eq!(gate.numel(), up.numel(), "swiglu_fused: numel mismatch");
    let buf = dev.elementwise(
        &[&gate.buffer, &up.buffer],
        gate.numel(),
        &|ids: &[ExprId]| {
            // silu(gate) = gate * sigmoid(gate) = gate / (1 + exp(-gate))
            let one = ExprId::from_f64(1.0);
            let neg_gate = -ids[0];
            let exp_neg = Scalar::exp(neg_gate);
            let denom = one + exp_neg;
            let sigmoid = one / denom;
            let silu = ids[0] * sigmoid;
            silu * ids[1]
        },
    );
    ComputeTensor::from_buffer(buf, gate.shape().to_vec())
}

/// SwiGLU backward: given grad_output, gate, and up from the forward pass,
/// computes grad_gate and grad_up.
///
/// Forward: out = silu(gate) * up, where silu(x) = x * sigmoid(x).
/// Backward:
///   grad_up = grad * silu(gate)
///   grad_gate = grad * up * dsilu(gate), where dsilu(x) = sigmoid(x) * (1 + x*(1-sigmoid(x)))
pub fn swiglu_backward<D: ComputeDevice>(
    dev: &D,
    grad_output: &ComputeTensor<D::Buffer>,
    gate: &ComputeTensor<D::Buffer>,
    up: &ComputeTensor<D::Buffer>,
) -> (ComputeTensor<D::Buffer>, ComputeTensor<D::Buffer>) {
    let numel = grad_output.numel();

    // grad_up = grad * silu(gate)
    let grad_up_buf = dev.elementwise(
        &[&grad_output.buffer, &gate.buffer],
        numel,
        &|ids: &[ExprId]| {
            let one = ExprId::from_f64(1.0);
            let neg_gate = -ids[1];
            let exp_neg = Scalar::exp(neg_gate);
            let sigmoid = one / (one + exp_neg);
            let silu = ids[1] * sigmoid;
            ids[0] * silu
        },
    );

    // grad_gate = grad * up * dsilu(gate)
    // dsilu(x) = sigmoid(x) + x * sigmoid(x) * (1 - sigmoid(x))
    //          = sigmoid(x) * (1 + x * (1 - sigmoid(x)))
    let grad_gate_buf = dev.elementwise(
        &[&grad_output.buffer, &gate.buffer, &up.buffer],
        numel,
        &|ids: &[ExprId]| {
            let one = ExprId::from_f64(1.0);
            let neg_gate = -ids[1];
            let exp_neg = Scalar::exp(neg_gate);
            let sigmoid = one / (one + exp_neg);
            let dsilu = sigmoid * (one + ids[1] * (one - sigmoid));
            ids[0] * ids[2] * dsilu
        },
    );

    (
        ComputeTensor::from_buffer(grad_gate_buf, grad_output.shape().to_vec()),
        ComputeTensor::from_buffer(grad_up_buf, grad_output.shape().to_vec()),
    )
}

/// Causal self-attention backward.
///
/// Recomputes attention scores from Q,K,V, then computes gradients.
/// Q, grad_output: `[seq_len, n_heads * head_dim]`
/// K, V: `[seq_len, n_kv_heads * head_dim]`
/// For GQA: n_kv_heads < n_heads, with n_heads/n_kv_heads heads per KV group.
///
/// Returns (grad_Q, grad_K, grad_V) with same shapes as inputs.
pub fn causal_attention_backward<D: ComputeDevice>(
    dev: &D,
    grad_output: &ComputeTensor<D::Buffer>,
    q: &ComputeTensor<D::Buffer>,
    k: &ComputeTensor<D::Buffer>,
    v: &ComputeTensor<D::Buffer>,
    seq_len: usize,
    n_heads: usize,
    n_kv_heads: usize,
    head_dim: usize,
) -> (ComputeTensor<D::Buffer>, ComputeTensor<D::Buffer>, ComputeTensor<D::Buffer>) {
    let total_dim = n_heads * head_dim;
    let kv_dim = n_kv_heads * head_dim;

    let (gq, gk, gv) = dev.causal_attention_backward(
        &grad_output.buffer,
        &q.buffer,
        &k.buffer,
        &v.buffer,
        seq_len,
        n_heads,
        n_kv_heads,
        head_dim,
    );

    (
        ComputeTensor::from_buffer(gq, vec![seq_len, total_dim]),
        ComputeTensor::from_buffer(gk, vec![seq_len, kv_dim]),
        ComputeTensor::from_buffer(gv, vec![seq_len, kv_dim]),
    )
}

#[cfg(test)]
mod tests {
    use super::*;
    use super::CpuDevice;

    #[test]
    fn test_add_tensors() {
        let dev = CpuDevice::new();
        let a = ComputeTensor::from_data(&dev, &[1.0, 2.0, 3.0], &[3]);
        let b = ComputeTensor::from_data(&dev, &[4.0, 5.0, 6.0], &[3]);
        let c = add_tensors(&dev, &a, &b);
        let out = c.to_vec();
        assert!((out[0] - 5.0).abs() < 1e-5);
        assert!((out[1] - 7.0).abs() < 1e-5);
        assert!((out[2] - 9.0).abs() < 1e-5);
    }

    #[test]
    fn test_bias_add() {
        let dev = CpuDevice::new();
        // 2x3 matrix + 3-element bias
        let mat = ComputeTensor::from_data(&dev, &[1.0, 2.0, 3.0, 4.0, 5.0, 6.0], &[2, 3]);
        let bias = ComputeTensor::from_data(&dev, &[0.1, 0.2, 0.3], &[3]);
        let out = bias_add(&dev, &mat, &bias);
        let v = out.to_vec();
        assert!((v[0] - 1.1).abs() < 1e-5);
        assert!((v[1] - 2.2).abs() < 1e-5);
        assert!((v[2] - 3.3).abs() < 1e-5);
        assert!((v[3] - 4.1).abs() < 1e-5);
        assert!((v[4] - 5.2).abs() < 1e-5);
        assert!((v[5] - 6.3).abs() < 1e-5);
    }

    #[test]
    fn test_swiglu_backward() {
        let dev = CpuDevice::new();
        let gate = ComputeTensor::from_data(&dev, &[0.0, 1.0, -1.0], &[3]);
        let up = ComputeTensor::from_data(&dev, &[1.0, 1.0, 1.0], &[3]);
        let grad = ComputeTensor::from_data(&dev, &[1.0, 1.0, 1.0], &[3]);
        let (grad_gate, grad_up) = swiglu_backward(&dev, &grad, &gate, &up);
        let gg = grad_gate.to_vec();
        let gu = grad_up.to_vec();
        // grad_up[i] = grad[i] * silu(gate[i])
        // At gate=0: silu(0)=0, so grad_up[0] should be 0
        assert!(gu[0].abs() < 1e-5);
        // At gate=1: silu(1)≈0.731, so grad_up[1]≈0.731
        assert!((gu[1] - 0.7311).abs() < 1e-3);
        // grad_gate should be finite
        for v in &gg {
            assert!(v.is_finite());
        }
    }

    #[test]
    fn test_causal_attention_backward() {
        let dev = CpuDevice::new();
        // seq_len=2, n_heads=1, n_kv_heads=1, head_dim=2
        let q = ComputeTensor::from_data(&dev, &[1.0, 0.0, 0.0, 1.0], &[2, 2]);
        let k = ComputeTensor::from_data(&dev, &[1.0, 0.0, 0.0, 1.0], &[2, 2]);
        let v = ComputeTensor::from_data(&dev, &[1.0, 2.0, 3.0, 4.0], &[2, 2]);
        let grad_out = ComputeTensor::from_data(&dev, &[1.0, 0.0, 0.0, 1.0], &[2, 2]);

        let (gq, gk, gv) = causal_attention_backward(
            &dev, &grad_out, &q, &k, &v, 2, 1, 1, 2,
        );
        assert_eq!(gq.shape(), &[2, 2]);
        assert_eq!(gk.shape(), &[2, 2]);
        assert_eq!(gv.shape(), &[2, 2]);
        // All gradients should be finite
        for v in gq.to_vec().iter().chain(gk.to_vec().iter()).chain(gv.to_vec().iter()) {
            assert!(v.is_finite(), "gradient should be finite");
        }
    }

    #[test]
    fn test_swiglu_fused() {
        let dev = CpuDevice::new();
        let gate = ComputeTensor::from_data(&dev, &[0.0, 1.0, -1.0], &[3]);
        let up = ComputeTensor::from_data(&dev, &[1.0, 1.0, 1.0], &[3]);
        let out = swiglu_fused(&dev, &gate, &up);
        let v = out.to_vec();
        // silu(0) * 1 = 0
        assert!(v[0].abs() < 1e-5);
        // silu(1) * 1 = 1 * sigmoid(1) ≈ 0.7311
        assert!((v[1] - 0.7311).abs() < 1e-3);
        // silu(-1) * 1 = -1 * sigmoid(-1) ≈ -0.2689
        assert!((v[2] - (-0.2689)).abs() < 1e-3);
    }
}