wasmicro 0.3.1

Tiny transformer inference for the web. BERT, GPT-2 and T5 in a 199 KB WASM bundle.
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

//! Weight-only quantized linear operations.

use crate::quant::{QuantizedTensorI8, QuantizedTensorQ4, QuantizedTensorU8};
use crate::tensor::Tensor;

/// Computes `C = A @ B^T` where `B` is signed 8-bit quantized.
///
/// - `a` shape: `[m, k]`
/// - `b` shape: `[n, k]`
/// - returned tensor shape: `[m, n]`
pub fn matmul_t_b_i8(a: &Tensor, b: &QuantizedTensorI8) -> Tensor {
    let a_shape = a.shape().as_slice();
    let b_shape = b.shape().as_slice();
    validate_matmul_shapes(a_shape, b_shape, "matmul_t_b_i8");

    let m = a_shape[0];
    let k = a_shape[1];
    let n = b_shape[0];
    let a_data = a.data();
    let b_data = b.data();
    let mut out = vec![0.0f32; m * n];

    for i in 0..m {
        let a_row = &a_data[i * k..(i + 1) * k];
        let out_row = &mut out[i * n..(i + 1) * n];
        for j in 0..n {
            let b_row = &b_data[j * k..(j + 1) * k];
            let scale = b.scale_for_row(j);
            let mut acc = 0.0f32;
            for kk in 0..k {
                acc += a_row[kk] * b_row[kk] as f32 * scale;
            }
            out_row[j] = acc;
        }
    }

    Tensor::from_vec(out, &[m, n])
}

/// Applies `y = x @ W^T + b` where `W` is signed 8-bit quantized.
pub fn linear_i8(x: &Tensor, weight: &QuantizedTensorI8, bias: Option<&Tensor>) -> Tensor {
    let mut y = matmul_t_b_i8(x, weight);
    add_optional_bias(&mut y, bias, "linear_i8");
    y
}

/// Computes `C = A @ B^T` where `B` is unsigned affine 8-bit quantized.
///
/// - `a` shape: `[m, k]`
/// - `b` shape: `[n, k]`
/// - returned tensor shape: `[m, n]`
pub fn matmul_t_b_u8(a: &Tensor, b: &QuantizedTensorU8) -> Tensor {
    let a_shape = a.shape().as_slice();
    let b_shape = b.shape().as_slice();
    validate_matmul_shapes(a_shape, b_shape, "matmul_t_b_u8");

    let m = a_shape[0];
    let k = a_shape[1];
    let n = b_shape[0];
    let a_data = a.data();
    let b_data = b.data();
    let mut out = vec![0.0f32; m * n];

    for i in 0..m {
        let a_row = &a_data[i * k..(i + 1) * k];
        let out_row = &mut out[i * n..(i + 1) * n];
        for (j, out_cell) in out_row.iter_mut().enumerate() {
            let b_row = &b_data[j * k..(j + 1) * k];
            let scale = b.scale_for_row(j);
            let zero_point = b.zero_point_for_row(j) as i32;
            let mut acc = 0.0f32;
            for (&a_val, &b_val) in a_row.iter().zip(b_row) {
                let deq = (b_val as i32 - zero_point) as f32 * scale;
                acc += a_val * deq;
            }
            *out_cell = acc;
        }
    }

    Tensor::from_vec(out, &[m, n])
}

/// Applies `y = x @ W^T + b` where `W` is unsigned affine 8-bit quantized.
pub fn linear_u8(x: &Tensor, weight: &QuantizedTensorU8, bias: Option<&Tensor>) -> Tensor {
    let mut y = matmul_t_b_u8(x, weight);
    add_optional_bias(&mut y, bias, "linear_u8");
    y
}

/// Computes `C = A @ B^T` where `B` is packed signed 4-bit quantized.
///
/// - `a` shape: `[m, k]`
/// - `b` shape: `[n, k]`
/// - returned tensor shape: `[m, n]`
pub fn matmul_t_b_q4(a: &Tensor, b: &QuantizedTensorQ4) -> Tensor {
    let a_shape = a.shape().as_slice();
    let b_shape = b.shape().as_slice();
    validate_matmul_shapes(a_shape, b_shape, "matmul_t_b_q4");

    let m = a_shape[0];
    let k = a_shape[1];
    let n = b_shape[0];
    let a_data = a.data();
    let mut out = vec![0.0f32; m * n];

    for i in 0..m {
        let a_row = &a_data[i * k..(i + 1) * k];
        let out_row = &mut out[i * n..(i + 1) * n];
        for (j, out_cell) in out_row.iter_mut().enumerate() {
            let scale = b.scale_for_row(j);
            let mut acc = 0.0f32;
            for (kk, &a_val) in a_row.iter().enumerate() {
                let q = b.get(j * k + kk) as f32;
                acc += a_val * q * scale;
            }
            *out_cell = acc;
        }
    }

    Tensor::from_vec(out, &[m, n])
}

/// Applies `y = x @ W^T + b` where `W` is packed signed 4-bit quantized.
pub fn linear_q4(x: &Tensor, weight: &QuantizedTensorQ4, bias: Option<&Tensor>) -> Tensor {
    let mut y = matmul_t_b_q4(x, weight);
    add_optional_bias(&mut y, bias, "linear_q4");
    y
}

fn validate_matmul_shapes(a_shape: &[usize], b_shape: &[usize], op: &str) {
    assert_eq!(a_shape.len(), 2, "{}: `a` must be 2D", op);
    assert_eq!(b_shape.len(), 2, "{}: `b` must be 2D", op);
    assert_eq!(
        a_shape[1], b_shape[1],
        "{}: inner dimensions must match: {:?} @ {:?}^T",
        op, a_shape, b_shape
    );
}

fn add_optional_bias(y: &mut Tensor, bias: Option<&Tensor>, op: &str) {
    let Some(bias) = bias else {
        return;
    };

    let y_shape = y.shape().as_slice();
    let bias_shape = bias.shape().as_slice();
    assert_eq!(bias_shape.len(), 1, "{}: bias must be 1D", op);
    assert_eq!(
        bias_shape[0], y_shape[1],
        "{}: bias length must match output columns",
        op
    );

    let n = y_shape[1];
    let bias_data = bias.data();
    for row in y.data_mut().chunks_mut(n) {
        for (row_val, &b_val) in row.iter_mut().zip(bias_data) {
            *row_val += b_val;
        }
    }
}

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

    fn assert_close(actual: &[f32], expected: &[f32]) {
        assert_eq!(actual.len(), expected.len());
        for (a, e) in actual.iter().zip(expected) {
            assert!((a - e).abs() < 1e-5, "actual {a}, expected {e}");
        }
    }

    #[test]
    fn linear_i8_matches_known_values() {
        let x = Tensor::from_vec(vec![3.0, 5.0], &[1, 2]);
        let w = QuantizedTensorI8::from_vec(vec![2, -4, 1, 2], &[2, 2], vec![0.5, 2.0]);
        let b = Tensor::from_vec(vec![1.0, -6.0], &[2]);
        let y = linear_i8(&x, &w, Some(&b));
        assert_eq!(y.shape().as_slice(), &[1, 2]);
        assert_close(y.data(), &[-6.0, 20.0]);
    }

    #[test]
    fn linear_u8_matches_known_values() {
        let x = Tensor::from_vec(vec![3.0, 5.0], &[1, 2]);
        let w =
            QuantizedTensorU8::from_vec(vec![12, 6, 11, 12], &[2, 2], vec![0.5, 2.0], vec![10, 10]);
        let b = Tensor::from_vec(vec![1.0, -6.0], &[2]);
        let y = linear_u8(&x, &w, Some(&b));
        assert_eq!(y.shape().as_slice(), &[1, 2]);
        assert_close(y.data(), &[-6.0, 20.0]);
    }

    #[test]
    fn linear_q4_matches_known_values() {
        let x = Tensor::from_vec(vec![3.0, 5.0], &[1, 2]);
        let w = QuantizedTensorQ4::from_i4_values(&[2, -4, 1, 2], &[2, 2], vec![0.5, 2.0]);
        let b = Tensor::from_vec(vec![1.0, -6.0], &[2]);
        let y = linear_q4(&x, &w, Some(&b));
        assert_eq!(y.shape().as_slice(), &[1, 2]);
        assert_close(y.data(), &[-6.0, 20.0]);
    }
}