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

//! Training configuration and metrics

use crate::error::Result;
use crate::optimizer::grad_scaler::GradScaler;
use numr::dtype::DType;

/// Training configuration
#[derive(Debug, Clone)]
pub struct TrainingConfig {
    pub learning_rate: f64,
    pub weight_decay: f64,
    pub max_grad_norm: Option<f64>,
    pub grad_accum_steps: usize,
}

impl Default for TrainingConfig {
    fn default() -> Self {
        Self {
            learning_rate: 1e-3,
            weight_decay: 0.01,
            max_grad_norm: Some(1.0),
            grad_accum_steps: 1,
        }
    }
}

impl TrainingConfig {
    pub fn with_lr(mut self, lr: f64) -> Self {
        self.learning_rate = lr;
        self
    }

    pub fn with_weight_decay(mut self, wd: f64) -> Self {
        self.weight_decay = wd;
        self
    }

    pub fn with_max_grad_norm(mut self, norm: Option<f64>) -> Self {
        self.max_grad_norm = norm;
        self
    }

    pub fn with_grad_accum_steps(mut self, steps: usize) -> Self {
        self.grad_accum_steps = steps;
        self
    }
}

/// Metrics from a training step
#[derive(Debug, Clone)]
pub struct TrainingMetrics {
    pub step: u64,
    pub loss: f64,
    pub grad_norm: Option<f64>,
    pub lr: f64,
}

/// Mixed precision training configuration
#[derive(Debug, Clone)]
pub struct MixedPrecisionConfig {
    /// Dtype for forward/backward computation (BF16 or F16)
    pub compute_dtype: DType,
    /// Dtype for master weights (always F32)
    pub master_dtype: DType,
    /// Loss scaling strategy
    pub loss_scale: LossScaleStrategy,
}

impl MixedPrecisionConfig {
    /// BF16 mixed precision (recommended).
    ///
    /// No loss scaling needed — BF16 has the same exponent range as F32.
    pub fn bf16() -> Self {
        Self {
            compute_dtype: DType::BF16,
            master_dtype: DType::F32,
            loss_scale: LossScaleStrategy::None,
        }
    }

    /// FP16 mixed precision with dynamic loss scaling.
    ///
    /// Requires loss scaling because FP16 has a narrower exponent range.
    pub fn fp16() -> Self {
        Self {
            compute_dtype: DType::F16,
            master_dtype: DType::F32,
            loss_scale: LossScaleStrategy::Dynamic {
                initial_scale: 65536.0,
                growth_factor: 2.0,
                backoff_factor: 0.5,
                growth_interval: 2000,
            },
        }
    }
}

/// Loss scaling strategy for mixed precision training
#[derive(Debug, Clone)]
pub enum LossScaleStrategy {
    /// No loss scaling (for BF16)
    None,
    /// Fixed loss scale factor
    Fixed(f64),
    /// Dynamic loss scaling (for FP16)
    Dynamic {
        initial_scale: f64,
        growth_factor: f64,
        backoff_factor: f64,
        growth_interval: u64,
    },
}

impl LossScaleStrategy {
    /// Convert to an optional GradScaler.
    pub(crate) fn to_grad_scaler(&self) -> Result<Option<GradScaler>> {
        match self {
            LossScaleStrategy::None => Ok(None),
            LossScaleStrategy::Fixed(scale) => {
                // Fixed scale: use huge growth interval so it never changes
                Ok(Some(GradScaler::new(*scale, 2.0, 0.5, u64::MAX)?))
            }
            LossScaleStrategy::Dynamic {
                initial_scale,
                growth_factor,
                backoff_factor,
                growth_interval,
            } => Ok(Some(GradScaler::new(
                *initial_scale,
                *growth_factor,
                *backoff_factor,
                *growth_interval,
            )?)),
        }
    }
}

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

    #[test]
    fn test_default_config() {
        let config = TrainingConfig::default();
        assert_eq!(config.learning_rate, 1e-3);
        assert_eq!(config.weight_decay, 0.01);
        assert_eq!(config.max_grad_norm, Some(1.0));
        assert_eq!(config.grad_accum_steps, 1);
    }

    #[test]
    fn test_builder() {
        let config = TrainingConfig::default()
            .with_lr(0.01)
            .with_weight_decay(0.1)
            .with_max_grad_norm(None)
            .with_grad_accum_steps(4);
        assert_eq!(config.learning_rate, 0.01);
        assert_eq!(config.weight_decay, 0.1);
        assert_eq!(config.max_grad_norm, None);
        assert_eq!(config.grad_accum_steps, 4);
    }
}