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

//! CPU implementation of FusedFp8TrainingOps
//!
//! Single-pass gradient unscale + clip with inf/nan detection.

use crate::error::{Error, Result};
use crate::ops::impl_generic::training::dynamic_loss_scale_update_impl;
use crate::ops::traits::FusedFp8TrainingOps;
use numr::dtype::DType;
use numr::runtime::cpu::{CpuClient, CpuRuntime};
use numr::tensor::Tensor;

impl FusedFp8TrainingOps<CpuRuntime> for CpuClient {
    fn fused_grad_unscale_clip(
        &self,
        grad: &Tensor<CpuRuntime>,
        max_norm: f64,
        loss_scale: f64,
    ) -> Result<(Tensor<CpuRuntime>, f64, bool)> {
        match grad.dtype() {
            DType::F32 => fused_grad_unscale_clip_f32(grad, max_norm, loss_scale),
            DType::F64 => fused_grad_unscale_clip_f64(grad, max_norm, loss_scale),
            dt => Err(Error::InvalidArgument {
                arg: "dtype",
                reason: format!("fused_grad_unscale_clip: unsupported dtype {:?}", dt),
            }),
        }
    }

    fn dynamic_loss_scale_update(
        &self,
        found_inf: bool,
        loss_scale: f64,
        growth_tracker: i32,
        growth_interval: i32,
        backoff_factor: f64,
    ) -> Result<(f64, i32)> {
        dynamic_loss_scale_update_impl(
            found_inf,
            loss_scale,
            growth_tracker,
            growth_interval,
            backoff_factor,
        )
    }
}

fn fused_grad_unscale_clip_f32(
    grad: &Tensor<CpuRuntime>,
    max_norm: f64,
    loss_scale: f64,
) -> Result<(Tensor<CpuRuntime>, f64, bool)> {
    let n: usize = grad.shape().iter().product();
    let g = grad.to_vec::<f32>();
    let inv_scale = (1.0 / loss_scale) as f32;
    let max_norm_f = max_norm as f32;

    // Pass 1: check inf/nan + unscale + accumulate norm²
    let mut unscaled = vec![0.0f32; n];
    let mut norm_sq: f64 = 0.0;
    let mut found_inf = false;

    for i in 0..n {
        let gi = g[i];
        if gi.is_infinite() || gi.is_nan() {
            found_inf = true;
            break;
        }
        let u = gi * inv_scale;
        unscaled[i] = u;
        norm_sq += (u as f64) * (u as f64);
    }

    if found_inf {
        // Return zeros — caller should skip this step
        let zeros = Tensor::<CpuRuntime>::zeros(grad.shape(), DType::F32, grad.device());
        return Ok((zeros, 0.0, true));
    }

    let norm = (norm_sq as f32).sqrt();

    // Pass 2: clip if norm > max_norm
    if norm > max_norm_f && max_norm_f > 0.0 {
        let clip_coef = max_norm_f / (norm + 1e-6);
        for u in &mut unscaled {
            *u *= clip_coef;
        }
    }

    let result = Tensor::<CpuRuntime>::from_slice(&unscaled, grad.shape(), grad.device());
    Ok((result, norm as f64, false))
}

fn fused_grad_unscale_clip_f64(
    grad: &Tensor<CpuRuntime>,
    max_norm: f64,
    loss_scale: f64,
) -> Result<(Tensor<CpuRuntime>, f64, bool)> {
    let n: usize = grad.shape().iter().product();
    let g = grad.to_vec::<f64>();
    let inv_scale = 1.0 / loss_scale;

    // Pass 1: check inf/nan + unscale + accumulate norm²
    let mut unscaled = vec![0.0f64; n];
    let mut norm_sq: f64 = 0.0;
    let mut found_inf = false;

    for i in 0..n {
        let gi = g[i];
        if gi.is_infinite() || gi.is_nan() {
            found_inf = true;
            break;
        }
        let u = gi * inv_scale;
        unscaled[i] = u;
        norm_sq += u * u;
    }

    if found_inf {
        let zeros = Tensor::<CpuRuntime>::zeros(grad.shape(), DType::F64, grad.device());
        return Ok((zeros, 0.0, true));
    }

    let norm = norm_sq.sqrt();

    // Pass 2: clip if norm > max_norm
    if norm > max_norm && max_norm > 0.0 {
        let clip_coef = max_norm / (norm + 1e-6);
        for u in &mut unscaled {
            *u *= clip_coef;
        }
    }

    let result = Tensor::<CpuRuntime>::from_slice(&unscaled, grad.shape(), grad.device());
    Ok((result, norm, false))
}

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

    #[test]
    fn test_fused_grad_unscale_clip_basic() {
        let (client, device) = cpu_setup();
        let grad = Tensor::<CpuRuntime>::from_slice(&[2.0f32, 4.0, 6.0, 8.0], &[4], &device);
        let loss_scale = 2.0;
        let max_norm = 10.0;

        let (clipped, norm, found_inf) = client
            .fused_grad_unscale_clip(&grad, max_norm, loss_scale)
            .unwrap();

        assert!(!found_inf);
        // unscaled = [1, 2, 3, 4], norm = sqrt(1+4+9+16) = sqrt(30) ≈ 5.477
        let data = clipped.to_vec::<f32>();
        assert!((data[0] - 1.0).abs() < 1e-5);
        assert!((data[1] - 2.0).abs() < 1e-5);
        assert!((norm - 30.0f64.sqrt()).abs() < 1e-4);
    }

    #[test]
    fn test_fused_grad_unscale_clip_clips() {
        let (client, device) = cpu_setup();
        let grad = Tensor::<CpuRuntime>::from_slice(&[20.0f32, 40.0, 60.0, 80.0], &[4], &device);
        let loss_scale = 2.0;
        let max_norm = 1.0;

        let (clipped, _norm, found_inf) = client
            .fused_grad_unscale_clip(&grad, max_norm, loss_scale)
            .unwrap();

        assert!(!found_inf);
        // unscaled = [10, 20, 30, 40], norm ≈ 54.77, should be clipped to max_norm=1.0
        let data = clipped.to_vec::<f32>();
        let clipped_norm: f32 = data.iter().map(|x| x * x).sum::<f32>().sqrt();
        assert!((clipped_norm - 1.0).abs() < 1e-3);
    }

    #[test]
    fn test_fused_grad_unscale_clip_inf() {
        let (client, device) = cpu_setup();
        let grad =
            Tensor::<CpuRuntime>::from_slice(&[1.0f32, f32::INFINITY, 3.0, 4.0], &[4], &device);

        let (_clipped, _norm, found_inf) = client.fused_grad_unscale_clip(&grad, 1.0, 1.0).unwrap();

        assert!(found_inf);
    }

    #[test]
    fn test_fused_grad_unscale_clip_nan() {
        let (client, device) = cpu_setup();
        let grad = Tensor::<CpuRuntime>::from_slice(&[1.0f32, f32::NAN, 3.0, 4.0], &[4], &device);

        let (_clipped, _norm, found_inf) = client.fused_grad_unscale_clip(&grad, 1.0, 1.0).unwrap();

        assert!(found_inf);
    }

    #[test]
    fn test_dynamic_loss_scale_growth() {
        let (client, _device) = cpu_setup();
        let (scale, tracker) = client
            .dynamic_loss_scale_update(false, 1024.0, 499, 500, 0.5)
            .unwrap();
        assert!((scale - 2048.0).abs() < 1e-10);
        assert_eq!(tracker, 0);
    }

    #[test]
    fn test_dynamic_loss_scale_backoff() {
        let (client, _device) = cpu_setup();
        let (scale, tracker) = client
            .dynamic_loss_scale_update(true, 1024.0, 100, 500, 0.5)
            .unwrap();
        assert!((scale - 512.0).abs() < 1e-10);
        assert_eq!(tracker, 0);
    }
}