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
177
178
179

// SPDX-License-Identifier: Apache-2.0
use super::*;
use crate::{MuonMatrixLayout, MuonMomentumMode, TestAutodiffBackend, TestBackend};
use ruda_model::{module::Module, record::FullPrecisionSettings, tensor::backend::Backend};

#[derive(Module, Debug)]
struct SmallModel<B: Backend> {
    hidden: Param<Tensor<B, 2>>,
    head: Param<Tensor<B, 2>>,
    bias: Param<Tensor<B, 1>>,
}
fn model() -> SmallModel<TestAutodiffBackend> {
    let device = Default::default();
    SmallModel {
        hidden: Param::initialized(101u64.into(), Tensor::from_data([[0.1, 0.2, 0.3], [0.4, 0.5, 0.6]], &device).require_grad()),
        head: Param::initialized(102u64.into(), Tensor::from_data([[0.2, 0.1], [0.4, 0.3], [0.6, 0.5]], &device).require_grad()),
        bias: Param::initialized(103u64.into(), Tensor::from_data([0.1, -0.1], &device).require_grad()),
    }
}
fn config() -> MuonAdamWConfig {
    MuonAdamWConfig::new().with_muon(MuonConfig::new()
        .with_momentum_mode(MuonMomentumMode::Ema).with_stable_normalization(true)
        .with_matrix_layout(MuonMatrixLayout::InputOutput))
}
fn gradients(step: usize) -> GradientsParams {
    let device = Default::default();
    let mut g = GradientsParams::new();
    g.register::<TestBackend, 2>(101u64.into(), Tensor::from_data([[0.1, 0.3, -0.2], [0.7, 0.1, -0.1]], &device).mul_scalar(1.0+step as f32*0.1));
    g.register::<TestBackend, 2>(102u64.into(), Tensor::from_data([[0.25, -0.125], [0.5, 0.125], [0.75, -0.5]], &device));
    g.register::<TestBackend, 1>(103u64.into(), Tensor::from_data([0.3, -0.4], &device));
    g
}
fn same_model(a: &SmallModel<TestAutodiffBackend>, b: &SmallModel<TestAutodiffBackend>) {
    assert_eq!(a.hidden.val().to_data().to_vec::<f32>().unwrap(), b.hidden.val().to_data().to_vec::<f32>().unwrap());
    assert_eq!(a.head.val().to_data().to_vec::<f32>().unwrap(), b.head.val().to_data().to_vec::<f32>().unwrap());
    assert_eq!(a.bias.val().to_data().to_vec::<f32>().unwrap(), b.bias.val().to_data().to_vec::<f32>().unwrap());
}

#[test]
fn muon_groups_route_hidden_only_and_preserve_parameter_ids() {
    let mut mixed_model = model();
    let mut expected = mixed_model.clone();
    let config = config();
    let mut mixed = config.init(&mixed_model, &[mixed_model.hidden.id]).unwrap();
    let mut only_muon = config.muon.init();
    let mut only_adam = config.adamw.init();
    for step in 0..4 {
        let mut grads = gradients(step);
        let mut selected = GradientsParams::new();
        selected.register::<TestBackend, 2>(expected.hidden.id, grads.remove(expected.hidden.id).unwrap());
        expected = only_muon.step(0.02, expected, selected);
        expected = only_adam.step(0.0003, expected, grads);
        mixed_model = mixed.try_step_with_lrs(0.02, 0.0003, mixed_model, gradients(step)).unwrap();
        same_model(&mixed_model, &expected);
        assert_eq!(mixed_model.hidden.id.val(), 101);
        assert!(mixed_model.hidden.is_require_grad());
    }
    let record = mixed.to_record();
    assert_eq!(record.muon_state_count(), 1);
    assert_eq!(record.adamw_state_count(), 2); // Includes the 2D output head.
}

#[test]
fn muon_groups_missing_gradient_skips_decay_and_state() {
    let original = model();
    let mut optim = config().init(&original, &[original.hidden.id]).unwrap();
    let mut grads = gradients(0);
    grads.remove::<TestBackend, 2>(original.head.id);
    let next = optim.try_step_with_lrs(0.02, 0.0003, original.clone(), grads).unwrap();
    assert_eq!(next.head.val().to_data().to_vec::<f32>().unwrap(), original.head.val().to_data().to_vec::<f32>().unwrap());
    assert_eq!(optim.to_record().adamw_state_count(), 1);
}

#[test]
fn muon_groups_explicit_skip_preserves_model_and_all_state() {
    let original = model();
    let mut optim = config().init(&original, &[original.hidden.id]).unwrap();
    let next = optim.try_step_or_skip(0.02, 0.0003, original.clone(), gradients(0), true).unwrap();
    same_model(&original, &next);
    assert_eq!(optim.to_record().muon_state_count(), 0);
    assert_eq!(optim.to_record().adamw_state_count(), 0);
}

#[test]
fn muon_groups_reject_bias_unknown_duplicate_and_empty_selection() {
    let m = model();
    assert!(matches!(config().init(&m, &[m.bias.id]), Err(MuonError::ExpectedMatrix { rank: 1 })));
    assert!(matches!(config().init(&m, &[999u64.into()]), Err(MuonError::UnknownParameter(999))));
    assert!(matches!(config().init(&m, &[m.hidden.id, m.hidden.id]), Err(MuonError::DuplicateParameter(101))));
    assert!(matches!(config().init(&m, &[]), Err(MuonError::EmptyMuonGroup)));
}

#[test]
fn muon_groups_reject_frozen_hidden_matrix() {
    let mut m = model();
    m.hidden = m.hidden.set_require_grad(false);
    assert!(matches!(config().init(&m, &[m.hidden.id]), Err(MuonError::FrozenParameter(101))));
}

#[test]
fn muon_groups_validate_both_groups_before_updating() {
    let m = model();
    let mut optim = config().init(&m, &[m.hidden.id]).unwrap();
    let mut g = gradients(0);
    g.register::<TestBackend, 1>(m.bias.id, Tensor::ones([1], &Default::default()));
    assert!(matches!(optim.try_step_with_lrs(0.02, 0.0003, m, g), Err(MuonError::ShapeMismatch("gradient"))));
    assert_eq!(optim.to_record().muon_state_count(), 0);
    assert_eq!(optim.to_record().adamw_state_count(), 0);
}

#[test]
fn muon_groups_reject_unknown_gradient_and_changed_model() {
    let mut m = model();
    let mut optim = config().init(&m, &[m.hidden.id]).unwrap();
    let mut g = gradients(0);
    g.register::<TestBackend, 1>(999u64.into(), Tensor::ones([1], &Default::default()));
    assert!(matches!(optim.try_step_with_lrs(0.02, 0.0003, m.clone(), g), Err(MuonError::UnusedGradients)));
    m.head = Param::initialized(m.head.id, Tensor::ones([4, 2], &Default::default()).require_grad());
    assert!(matches!(optim.try_step_with_lrs(0.02, 0.0003, m, GradientsParams::new()), Err(MuonError::ModelChanged)));
}

#[test]
fn muon_groups_record_roundtrip_continues_exactly() {
    let mut m = model();
    let mut optim = config().init(&m, &[m.hidden.id]).unwrap();
    m = optim.step(0.02, m, gradients(0));
    let item = optim.to_record().into_item::<FullPrecisionSettings>();
    let restored = MuonAdamWRecord::<TestAutodiffBackend>::from_item::<FullPrecisionSettings>(item, &Default::default());
    let mut other = config().init(&m, &[m.hidden.id]).unwrap().try_load_record(restored).unwrap();
    let a = optim.step(0.02, m.clone(), gradients(1));
    let b = other.step(0.02, m, gradients(1));
    same_model(&a, &b);
}

#[test]
fn muon_groups_reject_record_role_and_config_changes() {
    let m = model();
    let mut optim = config().init(&m, &[m.hidden.id]).unwrap();
    let _ = optim.step(0.02, m.clone(), gradients(0));
    assert!(matches!(config().init(&m, &[m.head.id]).unwrap().try_load_record(optim.to_record()), Err(MuonError::IncompatibleRecord)));
    assert!(matches!(config().with_adamw_lr_ratio(0.1).init(&m, &[m.hidden.id]).unwrap().try_load_record(optim.to_record()), Err(MuonError::IncompatibleRecord)));
    let altered = config().with_muon(MuonConfig::new());
    assert!(matches!(altered.init(&m, &[m.hidden.id]).unwrap().try_load_record(optim.to_record()), Err(MuonError::IncompatibleRecord)));
}

#[test]
fn muon_groups_invalid_auxiliary_config_and_rates() {
    let m = model();
    assert!(config().with_adamw(AdamWConfig::new().with_beta_1(f32::NAN)).init(&m, &[m.hidden.id]).is_err());
    assert!(config().with_adamw_lr_ratio(-1.0).init(&m, &[m.hidden.id]).is_err());
    let mut optim = config().init(&m, &[m.hidden.id]).unwrap();
    assert!(optim.try_step_with_lrs(0.02, f64::NAN, m, gradients(0)).is_err());
    assert_eq!(optim.to_record().muon_state_count(), 0);
}

#[derive(Module, Debug)]
struct TiedModel<B: Backend> { left: Param<Tensor<B, 2>>, right: Param<Tensor<B, 2>> }
#[test]
fn muon_groups_tied_parameter_updated_once() {
    let m = model();
    let tied = TiedModel { left: m.hidden.clone(), right: m.hidden.clone() };
    let mut optim = config().init(&tied, &[tied.left.id]).unwrap();
    let mut g = GradientsParams::new();
    let grad = Tensor::<TestBackend, 2>::ones([2, 3], &Default::default());
    g.register(tied.left.id, grad.clone());
    let (expected, _) = config().muon.build::<TestBackend>().try_step(0.02, tied.left.val().inner(), grad, None).unwrap();
    let next = optim.step(0.02, tied, g);
    assert_eq!(next.left.val().to_data().to_vec::<f32>().unwrap(), expected.to_data().to_vec::<f32>().unwrap());
    assert_eq!(next.left.val().to_data().to_vec::<f32>().unwrap(), next.right.val().to_data().to_vec::<f32>().unwrap());
    assert_eq!(optim.to_record().muon_state_count(), 1);
}

#[test]
#[should_panic(expected = "complete synchronized matrices")]
fn muon_groups_refuse_implicit_multi_device_shards() {
    let m = model();
    let mut optim = config().init(&m, &[m.hidden.id]).unwrap();
    let _ = optim.step_multi(0.02, m, MultiGradientsParams::default());
}