boostr 0.1.0

ML framework built on numr - attention, quantization, model architectures
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
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278

//! Gradient clipping utilities
//!
//! Clip gradients by global norm to prevent exploding gradients during training.

use crate::error::{Error, Result};
use numr::autograd::GradStore;
use numr::dtype::DType;
use numr::ops::{BinaryOps, ReduceOps, ScalarOps, UnaryOps, UtilityOps};
use numr::runtime::{Runtime, RuntimeClient};
use numr::tensor::TensorId;

/// Clip gradients in-place by global L2 norm.
///
/// Computes the global norm across all gradients, then scales each gradient
/// so the global norm does not exceed `max_norm`.
///
/// Returns the original global norm (before clipping).
pub fn clip_grad_norm<R, C>(client: &C, grads: &mut GradStore<R>, max_norm: f64) -> Result<f64>
where
    R: Runtime<DType = DType>,
    C: RuntimeClient<R> + ReduceOps<R> + ScalarOps<R> + UnaryOps<R> + BinaryOps<R>,
{
    if max_norm <= 0.0 {
        return Err(Error::TrainingError {
            reason: format!("max_norm must be positive, got {max_norm}"),
        });
    }

    // Compute global norm: sqrt(sum of all grad element squares)
    let ids: Vec<TensorId> = grads.keys().copied().collect();

    let mut total_norm_sq = 0.0f64;
    for &id in &ids {
        if let Some(grad) = grads.get(id) {
            // Flatten then sum all elements: sum(grad^2)
            let flat = grad.reshape(&[grad.numel()])?;
            let sq = client.mul(&flat, &flat)?;
            let sum = client.sum(&sq, &[0], false)?;
            let val: f32 = sum.item()?;
            total_norm_sq += val as f64;
        }
    }

    let total_norm = total_norm_sq.sqrt();

    if total_norm > max_norm {
        let scale = max_norm / (total_norm + 1e-6);
        for id in ids {
            if let Some(grad) = grads.get(id) {
                let clipped = client.mul_scalar(grad, scale)?;
                grads.insert(id, clipped);
            }
        }
    }

    Ok(total_norm)
}

/// Clip each gradient independently by its own L2 norm.
///
/// For each gradient tensor, if its L2 norm exceeds `max_norm`, it is scaled
/// down so its norm equals `max_norm`. Other gradients are left unchanged.
///
/// Returns a map of tensor ID → original norm for every gradient that was clipped.
pub fn clip_grad_norm_per_param<R, C>(
    client: &C,
    grads: &mut GradStore<R>,
    max_norm: f64,
) -> Result<Vec<(TensorId, f64)>>
where
    R: Runtime<DType = DType>,
    C: RuntimeClient<R> + ReduceOps<R> + ScalarOps<R> + UnaryOps<R> + BinaryOps<R>,
{
    if max_norm <= 0.0 {
        return Err(Error::TrainingError {
            reason: format!("max_norm must be positive, got {max_norm}"),
        });
    }

    let ids: Vec<TensorId> = grads.keys().copied().collect();
    let mut clipped = Vec::new();

    for id in ids {
        let grad = match grads.get(id) {
            Some(g) => g,
            None => continue,
        };

        let flat = grad.reshape(&[grad.numel()])?;
        let sq = client.mul(&flat, &flat)?;
        let sum = client.sum(&sq, &[0], false)?;
        let norm_sq: f64 = sum.item::<f32>()? as f64;
        let norm = norm_sq.sqrt();

        if norm > max_norm {
            let scale = max_norm / (norm + 1e-6);
            let scaled = client.mul_scalar(grad, scale)?;
            grads.insert(id, scaled);
            clipped.push((id, norm));
        }
    }

    Ok(clipped)
}

/// Clamp every gradient element to `[-clip_value, clip_value]`.
///
/// Unlike norm-based clipping, this operates element-wise and does not
/// preserve gradient direction.
pub fn clip_grad_value<R, C>(client: &C, grads: &mut GradStore<R>, clip_value: f64) -> Result<()>
where
    R: Runtime<DType = DType>,
    C: RuntimeClient<R> + UtilityOps<R>,
{
    if clip_value <= 0.0 {
        return Err(Error::TrainingError {
            reason: format!("clip_value must be positive, got {clip_value}"),
        });
    }

    let ids: Vec<TensorId> = grads.keys().copied().collect();

    for id in ids {
        let grad = match grads.get(id) {
            Some(g) => g,
            None => continue,
        };
        let clamped = client.clamp(grad, -clip_value, clip_value)?;
        grads.insert(id, clamped);
    }

    Ok(())
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::test_utils::cpu_setup;
    use numr::autograd::GradStore;
    use numr::runtime::cpu::CpuRuntime;
    use numr::tensor::Tensor;

    #[test]
    fn test_clip_no_op_when_under_max() {
        let (client, device) = cpu_setup();

        let id = TensorId::new();
        // grad = [1, 0] → norm = 1.0
        let t = Tensor::<CpuRuntime>::from_slice(&[1.0f32, 0.0], &[2], &device);
        let mut grads = GradStore::new();
        grads.insert(id, t);

        let norm = clip_grad_norm(&client, &mut grads, 5.0).unwrap();
        assert!((norm - 1.0).abs() < 1e-6);

        // Grads should be unchanged
        let data = grads.get(id).unwrap().to_vec::<f32>();
        assert!((data[0] - 1.0).abs() < 1e-6);
        assert!((data[1] - 0.0).abs() < 1e-6);
    }

    #[test]
    fn test_clip_scales_when_over_max() {
        let (client, device) = cpu_setup();

        let id = TensorId::new();
        // grad = [3, 4] → norm = 5.0, clip to max_norm=1.0
        let t = Tensor::<CpuRuntime>::from_slice(&[3.0f32, 4.0], &[2], &device);
        let mut grads = GradStore::new();
        grads.insert(id, t);

        let norm = clip_grad_norm(&client, &mut grads, 1.0).unwrap();
        assert!((norm - 5.0).abs() < 1e-4);

        // After clipping, norm should be ~1.0
        let data = grads.get(id).unwrap().to_vec::<f32>();
        let clipped_norm = (data[0] * data[0] + data[1] * data[1]).sqrt();
        assert!((clipped_norm - 1.0).abs() < 1e-4);
    }

    #[test]
    fn test_clip_multi_param_global_norm() {
        let (client, device) = cpu_setup();

        let id1 = TensorId::new();
        let id2 = TensorId::new();
        // grad1 = [3, 0], grad2 = [0, 4] → global norm = sqrt(9+16) = 5.0
        let t1 = Tensor::<CpuRuntime>::from_slice(&[3.0f32, 0.0], &[2], &device);
        let t2 = Tensor::<CpuRuntime>::from_slice(&[0.0f32, 4.0], &[2], &device);
        let mut grads = GradStore::new();
        grads.insert(id1, t1);
        grads.insert(id2, t2);

        let norm = clip_grad_norm(&client, &mut grads, 2.5).unwrap();
        assert!((norm - 5.0).abs() < 1e-4);

        // Both grads should be scaled by 2.5/5.0 = 0.5
        let d1 = grads.get(id1).unwrap().to_vec::<f32>();
        let d2 = grads.get(id2).unwrap().to_vec::<f32>();
        assert!((d1[0] - 1.5).abs() < 1e-4);
        assert!((d2[1] - 2.0).abs() < 1e-4);
    }

    #[test]
    fn test_clip_empty_grads() {
        let (client, _device) = cpu_setup();
        let mut grads = GradStore::<CpuRuntime>::new();

        let norm = clip_grad_norm(&client, &mut grads, 1.0).unwrap();
        assert!((norm - 0.0).abs() < 1e-6);
    }

    #[test]
    fn test_clip_rejects_non_positive_max_norm() {
        let (client, _device) = cpu_setup();
        let mut grads = GradStore::<CpuRuntime>::new();

        assert!(clip_grad_norm(&client, &mut grads, 0.0).is_err());
        assert!(clip_grad_norm(&client, &mut grads, -1.0).is_err());
    }

    #[test]
    fn test_clip_per_param_only_clips_large() {
        let (client, device) = cpu_setup();

        let id1 = TensorId::new();
        let id2 = TensorId::new();
        // grad1 = [3, 4] → norm = 5.0, should be clipped to norm 2.0
        // grad2 = [1, 0] → norm = 1.0, should NOT be clipped
        let t1 = Tensor::<CpuRuntime>::from_slice(&[3.0f32, 4.0], &[2], &device);
        let t2 = Tensor::<CpuRuntime>::from_slice(&[1.0f32, 0.0], &[2], &device);
        let mut grads = GradStore::new();
        grads.insert(id1, t1);
        grads.insert(id2, t2);

        let clipped = clip_grad_norm_per_param(&client, &mut grads, 2.0).unwrap();

        // Only id1 should have been clipped
        assert_eq!(clipped.len(), 1);
        assert!((clipped[0].1 - 5.0).abs() < 1e-4);

        // id1: norm should now be ~2.0
        let d1 = grads.get(id1).unwrap().to_vec::<f32>();
        let norm1 = (d1[0] * d1[0] + d1[1] * d1[1]).sqrt();
        assert!((norm1 - 2.0).abs() < 1e-3);

        // id2: unchanged
        let d2 = grads.get(id2).unwrap().to_vec::<f32>();
        assert!((d2[0] - 1.0).abs() < 1e-6);
    }

    #[test]
    fn test_clip_value() {
        let (client, device) = cpu_setup();

        let id = TensorId::new();
        let t = Tensor::<CpuRuntime>::from_slice(&[-5.0f32, 3.0, 0.5, -0.1], &[4], &device);
        let mut grads = GradStore::new();
        grads.insert(id, t);

        clip_grad_value(&client, &mut grads, 1.0).unwrap();

        let data = grads.get(id).unwrap().to_vec::<f32>();
        assert!((data[0] - (-1.0)).abs() < 1e-6); // clamped from -5
        assert!((data[1] - 1.0).abs() < 1e-6); // clamped from 3
        assert!((data[2] - 0.5).abs() < 1e-6); // unchanged
        assert!((data[3] - (-0.1)).abs() < 1e-6); // unchanged
    }

    #[test]
    fn test_clip_value_rejects_non_positive() {
        let (client, _device) = cpu_setup();
        let mut grads = GradStore::<CpuRuntime>::new();

        assert!(clip_grad_value(&client, &mut grads, 0.0).is_err());
        assert!(clip_grad_value(&client, &mut grads, -1.0).is_err());
    }
}