ruda-optim 0.21.13

Ruda optimizer updates, gradient state, clipping and learning-rate schedules.
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

// SPDX-License-Identifier: Apache-2.0
//! CPU oracle, called explicitly. The device implementation never dispatches here.
//!
//! Arithmetic is evaluated in FP64 and persistent state is rounded to FP32 after
//! every step. Device FP32 reassociation/FMA is compared with a tolerance, not bit
//! equality. Half-precision gradients must first be rounded to their storage type
//! and then converted to FP32 by the test caller.
use super::{AdamWOptions, FusedAdamWError, StepControl};

/// Explicit host state for independent correctness tests and fixtures.
#[derive(Debug, Clone, PartialEq)]
pub struct ReferenceState {
    /// Number of completed updates.
    pub step: u64,
    /// FP32 first moments.
    pub first: Vec<f32>,
    /// FP32 second moments.
    pub second: Vec<f32>,
    /// FP32 running maximum, present iff AMSGrad is selected.
    pub maximum: Option<Vec<f32>>,
}

/// Result of the independent CPU oracle.
#[derive(Debug, Clone, PartialEq)]
pub struct ReferenceUpdate {
    /// Updated parameters, or a copy of the original for an empty/skipped call.
    pub parameters: Vec<f32>,
    /// Empty/skipped calls preserve an absent state.
    pub state: Option<ReferenceState>,
    /// Whether a mathematical update was performed.
    pub updated: bool,
}

/// Compute the reference formula without modifying any input.
pub fn adamw_reference(
    parameters: &[f32],
    gradients: &[f32],
    state: Option<&ReferenceState>,
    options: &AdamWOptions,
    control: StepControl,
) -> Result<ReferenceUpdate, FusedAdamWError> {
    options.validate()?;
    control.validate()?;
    if parameters.len() != gradients.len() {
        return Err(FusedAdamWError::ShapeMismatch("gradient"));
    }
    if let Some(s) = state {
        if s.step == 0 || s.maximum.is_some() != options.amsgrad {
            return Err(FusedAdamWError::InvalidState("step/AMSGrad mode"));
        }
        if s.first.len() != parameters.len() || s.second.len() != parameters.len()
            || s.maximum.as_ref().is_some_and(|x| x.len() != parameters.len())
        {
            return Err(FusedAdamWError::ShapeMismatch("moment"));
        }
    }
    if parameters.is_empty() || control.skip_update {
        return Ok(ReferenceUpdate { parameters: parameters.to_vec(), state: state.cloned(), updated: false });
    }
    let previous_step = state.map_or(0, |s| s.step);
    let step = previous_step.checked_add(1).ok_or(FusedAdamWError::StepOverflow)?;
    let b1 = options.beta1 as f64;
    let b2 = options.beta2 as f64;
    // Independent of the production exponentiation helper.
    let correction1 = 1.0 - b1.powf(step as f64);
    let correction2 = 1.0 - b2.powf(step as f64);
    let lr = options.learning_rate as f64;
    let decay = 1.0 - lr * options.weight_decay as f64;
    let mut result = ReferenceUpdate {
        parameters: Vec::with_capacity(parameters.len()),
        state: Some(ReferenceState {
            step, first: Vec::with_capacity(parameters.len()), second: Vec::with_capacity(parameters.len()),
            maximum: options.amsgrad.then(|| Vec::with_capacity(parameters.len())),
        }),
        updated: true,
    };
    let next = result.state.as_mut().expect("state initialized above");
    for i in 0..parameters.len() {
        let mut g = gradients[i] as f64 / control.gradient_scale as f64;
        if options.maximize { g = -g; }
        let m = (b1 * state.map_or(0.0, |s| s.first[i] as f64) + (1.0 - b1) * g) as f32;
        let v = (b2 * state.map_or(0.0, |s| s.second[i] as f64) + (1.0 - b2) * g * g) as f32;
        next.first.push(m);
        next.second.push(v);
        let used_v = if let Some(maximum) = next.maximum.as_mut() {
            let old = state.and_then(|s| s.maximum.as_ref()).map_or(0.0, |s| s[i]);
            let value = if old.is_nan() || v.is_nan() { f32::NAN } else { old.max(v) };
            maximum.push(value);
            value
        } else { v };
        let denominator = (used_v as f64 / correction2).sqrt() + options.epsilon as f64;
        result.parameters.push((parameters[i] as f64 * decay - lr * (m as f64 / correction1) / denominator) as f32);
    }
    Ok(result)
}

#[cfg(test)]
mod tests {
    use super::*;
    fn close(a: f32, b: f32) { assert!((a - b).abs() <= 3e-6 * b.abs().max(1.0), "{a} != {b}"); }

    #[test]
    fn first_step_known_values_and_epsilon_outside_sqrt() {
        let o = AdamWOptions { learning_rate: 0.1, beta1: 0.0, beta2: 0.0, epsilon: 0.5, weight_decay: 0.2, ..Default::default() };
        let r = adamw_reference(&[2.0, -1.0, 0.0], &[1.0, -2.0, 0.0], None, &o, StepControl::default()).unwrap();
        close(r.parameters[0], 2.0 * 0.98 - 0.1 / 1.5);
        close(r.parameters[1], -0.98 + 0.2 / 2.5);
        close(r.parameters[2], 0.0);
        assert_eq!(r.state.unwrap().first, [1.0, -2.0, 0.0]);
    }
    #[test]
    fn scale_and_maximize_do_not_change_decay_direction() {
        let o = AdamWOptions { maximize: true, ..Default::default() };
        let scaled = adamw_reference(&[1.0], &[128.0], None, &o, StepControl { gradient_scale: 128.0, skip_update: false }).unwrap();
        let plain = adamw_reference(&[1.0], &[-1.0], None, &AdamWOptions { maximize: false, ..o }, StepControl::default()).unwrap();
        assert_eq!(scaled, plain);
    }
    #[test]
    fn skip_is_transactional_even_for_nan_gradients() {
        let o = AdamWOptions { amsgrad: true, ..Default::default() };
        let first = adamw_reference(&[1.0, 2.0], &[0.1, -0.3], None, &o, StepControl::default()).unwrap();
        let snapshot = first.clone();
        let skipped = adamw_reference(&first.parameters, &[f32::NAN, f32::INFINITY], first.state.as_ref(), &o, StepControl { skip_update: true, ..Default::default() }).unwrap();
        assert_eq!(skipped.parameters, snapshot.parameters);
        assert_eq!(skipped.state, snapshot.state);
        assert!(!skipped.updated);
        assert_eq!(first, snapshot);
    }
    #[test]
    fn empty_and_first_skip_keep_state_absent() {
        let o = AdamWOptions::default();
        assert_eq!(adamw_reference(&[], &[], None, &o, StepControl::default()).unwrap().state, None);
        let r = adamw_reference(&[1.0], &[f32::NAN], None, &o, StepControl { skip_update: true, ..Default::default() }).unwrap();
        assert_eq!(r.parameters, [1.0]);
        assert!(r.state.is_none());
    }
    #[test]
    fn zero_learning_rate_still_updates_moments() {
        let o = AdamWOptions { learning_rate: 0.0, ..Default::default() };
        let r = adamw_reference(&[1.0], &[3.0], None, &o, StepControl::default()).unwrap();
        assert_eq!(r.parameters, [1.0]);
        let s = r.state.unwrap();
        assert_eq!(s.step, 1);
        assert!(s.first[0] > 0.0 && s.second[0] > 0.0);
    }
    #[test]
    fn amsgrad_preserves_historical_maximum() {
        let o = AdamWOptions { beta2: 0.5, amsgrad: true, ..Default::default() };
        let a = adamw_reference(&[1.0], &[4.0], None, &o, StepControl::default()).unwrap();
        let b = adamw_reference(&a.parameters, &[0.0], a.state.as_ref(), &o, StepControl::default()).unwrap();
        let s = b.state.unwrap();
        assert_eq!(s.second, [4.0]);
        assert_eq!(s.maximum.unwrap(), [8.0]);
    }
    #[test]
    fn restored_state_matches_uninterrupted_steps() {
        let o = AdamWOptions { amsgrad: true, ..Default::default() };
        let first = adamw_reference(&[0.5, -0.1], &[0.3, 0.7], None, &o, StepControl::default()).unwrap();
        let snapshot = first.clone();
        let next = adamw_reference(&first.parameters, &[-0.1, 0.2], first.state.as_ref(), &o, StepControl::default()).unwrap();
        let resumed = adamw_reference(&snapshot.parameters, &[-0.1, 0.2], snapshot.state.as_ref(), &o, StepControl::default()).unwrap();
        assert_eq!(next, resumed);
        assert_eq!(next.state.unwrap().step, 2);
    }
    #[test]
    fn invalid_state_and_lengths_reject() {
        let o = AdamWOptions::default();
        assert!(adamw_reference(&[1.0], &[], None, &o, StepControl::default()).is_err());
        let s = ReferenceState { step: 0, first: vec![0.0], second: vec![0.0], maximum: None };
        assert!(adamw_reference(&[1.0], &[0.0], Some(&s), &o, StepControl::default()).is_err());
        let s = ReferenceState { step: 1, ..s };
        assert!(adamw_reference(&[1.0], &[0.0], Some(&s), &AdamWOptions { amsgrad: true, ..o }, StepControl::default()).is_err());
        let s = ReferenceState { step: u64::MAX, ..s };
        assert_eq!(adamw_reference(&[1.0], &[0.0], Some(&s), &o, StepControl::default()).unwrap_err(), FusedAdamWError::StepOverflow);
    }
}