oxidelm 0.2.0

Bare-metal pure-Rust Transformer engine with reverse-mode autograd, BPE tokenization, and wgpu acceleration
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
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
#[cfg(feature = "gpu")]
use crate::gpu::AdamWParams;
use crate::gpu::{GpuContext, GpuTensor};
use crate::Tensor;
use half::f16;
use serde::{Deserialize, Serialize};
use std::error::Error;
use std::fmt;
use std::fs;
use std::io::{BufReader, BufWriter};
use std::path::PathBuf;
use std::sync::atomic::{AtomicU64, Ordering};

fn default_true() -> bool {
    true
}

fn default_one() -> f32 {
    1.0
}

/// Embedding parameter names excluded from weight decay by default.
pub fn default_no_decay_embeddings() -> Vec<String> {
    vec![
        "tok_embeddings.weight".into(),
        "lm_head.weight".into(),
    ]
}

/// Which optimizer algorithm a config/step should use.
#[derive(Clone, Copy, Debug, Default, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "lowercase")]
pub enum OptimizerKind {
    #[default]
    AdamW,
    Sgd,
    Lion,
    Adafactor,
}

/// Precision used to store AdamW/Lion moment buffers in memory (and on disk when offloaded).
#[derive(Clone, Copy, Debug, Default, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "lowercase")]
pub enum MomentDtype {
    #[default]
    Fp32,
    Fp16,
}

#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
pub struct AdamWConfig {
    pub beta1: f32,
    pub beta2: f32,
    pub epsilon: f32,
    pub weight_decay: f32,
    /// Exclude RMSNorm scale parameters from weight decay.
    #[serde(default = "default_true")]
    pub no_decay_rms_norm: bool,
    /// Embedding parameter names that receive zero weight decay.
    #[serde(default = "default_no_decay_embeddings")]
    pub no_decay_embeddings: Vec<String>,
    /// Learning-rate multiplier for RMSNorm parameters relative to the schedule.
    #[serde(default = "default_one")]
    pub rms_norm_lr_scale: f32,
    /// Learning-rate multiplier for embedding parameters relative to the schedule.
    #[serde(default = "default_one")]
    pub embedding_lr_scale: f32,
    /// Optimizer algorithm to run. Defaults to AdamW for checkpoint compatibility.
    #[serde(default)]
    pub kind: OptimizerKind,
    /// Storage precision for AdamW/Lion moment buffers. Ignored for SGD/Adafactor.
    #[serde(default)]
    pub moment_dtype: MomentDtype,
    /// When true, moment buffers are written to a temp file and freed from memory between steps.
    #[serde(default)]
    pub offload_moments: bool,
    /// Momentum coefficient used only when `kind == OptimizerKind::Sgd`.
    #[serde(default)]
    pub momentum: f32,
}

impl Default for AdamWConfig {
    fn default() -> Self {
        Self {
            beta1: 0.9,
            beta2: 0.999,
            epsilon: 1e-8,
            weight_decay: 0.01,
            no_decay_rms_norm: true,
            no_decay_embeddings: default_no_decay_embeddings(),
            rms_norm_lr_scale: 1.0,
            embedding_lr_scale: 1.0,
            kind: OptimizerKind::default(),
            moment_dtype: MomentDtype::default(),
            offload_moments: false,
            momentum: 0.0,
        }
    }
}

impl AdamWConfig {
    pub fn validate(&self) -> Result<(), OptimizerError> {
        validate_config(self)
    }

    /// Resolve per-parameter weight decay and learning-rate scale from a name.
    pub fn group_for(&self, name: &str) -> ParameterGroup {
        let kind = parameter_kind(name);
        let weight_decay = match kind {
            ParameterKind::RmsNorm if self.no_decay_rms_norm => 0.0,
            ParameterKind::Embedding
                if self
                    .no_decay_embeddings
                    .iter()
                    .any(|candidate| candidate == name) =>
            {
                0.0
            }
            _ => self.weight_decay,
        };
        let lr_scale = match kind {
            ParameterKind::RmsNorm => self.rms_norm_lr_scale,
            ParameterKind::Embedding => self.embedding_lr_scale,
            ParameterKind::Other => 1.0,
        };
        ParameterGroup {
            weight_decay,
            lr_scale,
        }
    }
}

#[derive(Clone, Copy, Debug, PartialEq)]
enum ParameterKind {
    RmsNorm,
    Embedding,
    Other,
}

fn parameter_kind(name: &str) -> ParameterKind {
    if name == "norm.weight"
        || name.ends_with(".attention_norm.weight")
        || name.ends_with(".ffn_norm.weight")
    {
        return ParameterKind::RmsNorm;
    }
    if name == "tok_embeddings.weight"
        || name == "lm_head.weight"
        || name == "lm_head.bias"
        || name.starts_with("tok_embeddings.")
        || name.starts_with("lm_head.")
    {
        return ParameterKind::Embedding;
    }
    ParameterKind::Other
}

/// Per-parameter AdamW hyperparameters applied during a step.
#[derive(Clone, Copy, Debug, PartialEq, Serialize, Deserialize)]
pub struct ParameterGroup {
    pub weight_decay: f32,
    pub lr_scale: f32,
}

impl ParameterGroup {
    pub fn uniform(weight_decay: f32) -> Self {
        Self {
            weight_decay,
            lr_scale: 1.0,
        }
    }
}

/// Backing storage for a single moment buffer.
///
/// `Fp32` stores full-precision values. `Fp16` stores IEEE-754 half-precision bit
/// patterns (via `half::f16`), halving memory usage at the cost of update precision;
/// values are widened to `f32` for arithmetic and narrowed back to `f16` on write.
#[derive(Clone, Debug, Serialize, Deserialize)]
pub enum MomentStorage {
    Fp32(Vec<f32>),
    Fp16(Vec<u16>),
}

impl MomentStorage {
    fn zeros(len: usize, dtype: MomentDtype) -> Self {
        match dtype {
            MomentDtype::Fp32 => MomentStorage::Fp32(vec![0.0; len]),
            MomentDtype::Fp16 => MomentStorage::Fp16(vec![0u16; len]),
        }
    }

    pub fn len(&self) -> usize {
        match self {
            MomentStorage::Fp32(values) => values.len(),
            MomentStorage::Fp16(values) => values.len(),
        }
    }

    pub fn is_empty(&self) -> bool {
        self.len() == 0
    }

    pub fn dtype(&self) -> MomentDtype {
        match self {
            MomentStorage::Fp32(_) => MomentDtype::Fp32,
            MomentStorage::Fp16(_) => MomentDtype::Fp16,
        }
    }

    /// Read element `index`, widening from FP16 to FP32 when necessary.
    #[inline]
    pub fn get(&self, index: usize) -> f32 {
        match self {
            MomentStorage::Fp32(values) => values[index],
            MomentStorage::Fp16(values) => f16::from_bits(values[index]).to_f32(),
        }
    }

    /// Write element `index`, narrowing from FP32 to FP16 when necessary.
    #[inline]
    pub fn set(&mut self, index: usize, value: f32) {
        match self {
            MomentStorage::Fp32(values) => values[index] = value,
            MomentStorage::Fp16(values) => values[index] = f16::from_f32(value).to_bits(),
        }
    }

    pub fn to_f32_vec(&self) -> Vec<f32> {
        match self {
            MomentStorage::Fp32(values) => values.clone(),
            MomentStorage::Fp16(values) => values
                .iter()
                .map(|&bits| f16::from_bits(bits).to_f32())
                .collect(),
        }
    }

    /// Borrow the underlying FP32 buffer, or `None` if this storage is FP16.
    pub fn as_fp32(&self) -> Option<&[f32]> {
        match self {
            MomentStorage::Fp32(values) => Some(values),
            MomentStorage::Fp16(_) => None,
        }
    }
}

/// Optimizer moment buffers for a single parameter.
///
/// The meaning and length of `first`/`second` depend on [`OptimizerKind`]:
/// - `AdamW`: `first` = exp_avg (len = numel), `second` = exp_avg_sq (len = numel).
/// - `Sgd` with `momentum > 0`: `first` = velocity (len = numel), `second` = empty.
/// - `Sgd` with `momentum == 0`: both `first` and `second` are empty (stateless).
/// - `Lion`: `first` = momentum buffer (len = numel), `second` = empty.
/// - `Adafactor`, 2D parameter of shape `[rows, cols]`: `first` = row factor
///   `exp_avg_sq_row` (len = `rows`), `second` = col factor `exp_avg_sq_col`
///   (len = `cols`). The full second moment is reconstructed as
///   `outer(first, second) / mean(first)`.
/// - `Adafactor`, non-2D parameter (e.g. 1D bias/norm): `first` is empty,
///   `second` holds a full (non-factored) second moment of length `numel`,
///   i.e. plain RMSprop-style accumulation.
#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct ParameterMoments {
    pub first: MomentStorage,
    pub second: MomentStorage,
}

#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct AdamWState {
    pub step: usize,
    pub moments: Vec<ParameterMoments>,
}

#[derive(Debug)]
pub struct OptimizerError(pub String);

impl fmt::Display for OptimizerError {
    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
        write!(formatter, "AdamW error: {}", self.0)
    }
}

impl Error for OptimizerError {}

pub struct AdamW {
    config: AdamWConfig,
    state: AdamWState,
    groups: Vec<ParameterGroup>,
    gpu_moments: Option<Vec<(GpuTensor, GpuTensor)>>,
    /// Path to the on-disk moment snapshot once at least one page-out has occurred.
    offload_path: Option<PathBuf>,
    /// Unique id used to derive a per-instance offload filename.
    offload_id: u64,
    /// True when `state.moments` has been cleared and persisted to `offload_path`.
    moments_paged_out: bool,
}

fn next_offload_id() -> u64 {
    static COUNTER: AtomicU64 = AtomicU64::new(0);
    COUNTER.fetch_add(1, Ordering::Relaxed)
}

impl AdamW {
    /// Create an optimizer with a single uniform parameter group.
    pub fn new(parameters: &[Tensor], config: AdamWConfig) -> Result<Self, OptimizerError> {
        let groups = vec![ParameterGroup::uniform(config.weight_decay); parameters.len()];
        Self::from_groups(parameters, groups, config)
    }

    /// Create an optimizer with groups derived from named parameters.
    pub fn new_grouped(
        named_parameters: &[(String, Tensor)],
        config: AdamWConfig,
    ) -> Result<Self, OptimizerError> {
        let groups = named_parameters
            .iter()
            .map(|(name, _)| config.group_for(name))
            .collect();
        let parameters: Vec<Tensor> = named_parameters
            .iter()
            .map(|(_, parameter)| parameter.clone())
            .collect();
        Self::from_groups(&parameters, groups, config)
    }

    pub fn from_state(
        parameters: &[Tensor],
        config: AdamWConfig,
        state: AdamWState,
    ) -> Result<Self, OptimizerError> {
        let groups = vec![ParameterGroup::uniform(config.weight_decay); parameters.len()];
        Self::from_groups_and_state(parameters, groups, config, state)
    }

    /// Restore optimizer state and re-derive groups from parameter names.
    pub fn from_state_grouped(
        named_parameters: &[(String, Tensor)],
        config: AdamWConfig,
        state: AdamWState,
    ) -> Result<Self, OptimizerError> {
        let groups = named_parameters
            .iter()
            .map(|(name, _)| config.group_for(name))
            .collect();
        let parameters: Vec<Tensor> = named_parameters
            .iter()
            .map(|(_, parameter)| parameter.clone())
            .collect();
        Self::from_groups_and_state(&parameters, groups, config, state)
    }

    fn from_groups(
        parameters: &[Tensor],
        groups: Vec<ParameterGroup>,
        config: AdamWConfig,
    ) -> Result<Self, OptimizerError> {
        validate_config(&config)?;
        validate_groups(parameters.len(), &groups)?;
        let moments = parameters
            .iter()
            .map(|parameter| moments_for(parameter, &config))
            .collect();
        Ok(Self {
            config,
            state: AdamWState { step: 0, moments },
            groups,
            gpu_moments: None,
            offload_path: None,
            offload_id: next_offload_id(),
            moments_paged_out: false,
        })
    }

    fn from_groups_and_state(
        parameters: &[Tensor],
        groups: Vec<ParameterGroup>,
        config: AdamWConfig,
        state: AdamWState,
    ) -> Result<Self, OptimizerError> {
        validate_config(&config)?;
        validate_groups(parameters.len(), &groups)?;
        validate_state(parameters, &config, &state)?;
        Ok(Self {
            config,
            state,
            groups,
            gpu_moments: None,
            offload_path: None,
            offload_id: next_offload_id(),
            moments_paged_out: false,
        })
    }

    pub fn step(
        &mut self,
        parameters: &[Tensor],
        learning_rate: f32,
    ) -> Result<(), OptimizerError> {
        if !learning_rate.is_finite() || learning_rate < 0.0 {
            return Err(OptimizerError(
                "learning rate must be finite and non-negative".into(),
            ));
        }
        self.page_in_moments()?;
        validate_state(parameters, &self.config, &self.state)?;
        validate_groups(parameters.len(), &self.groups)?;
        self.state.step += 1;

        match self.config.kind {
            OptimizerKind::AdamW => self.step_adamw(parameters, learning_rate),
            OptimizerKind::Sgd => self.step_sgd(parameters, learning_rate),
            OptimizerKind::Lion => self.step_lion(parameters, learning_rate),
            OptimizerKind::Adafactor => self.step_adafactor(parameters, learning_rate),
        }

        if self.config.offload_moments {
            self.page_out_moments()?;
        }
        Ok(())
    }

    fn step_adamw(&mut self, parameters: &[Tensor], learning_rate: f32) {
        let step = self.state.step as f32;
        let first_correction = 1.0 - self.config.beta1.powf(step);
        let second_correction = 1.0 - self.config.beta2.powf(step);
        let beta1 = self.config.beta1;
        let beta2 = self.config.beta2;
        let epsilon = self.config.epsilon;

        for (index, (parameter, moments)) in
            parameters.iter().zip(self.state.moments.iter_mut()).enumerate()
        {
            let Some(gradient) = parameter.grad() else {
                continue;
            };
            let group = self.groups[index];
            let effective_lr = learning_rate * group.lr_scale;
            parameter.with_mut_data(|values| {
                for i in 0..values.len() {
                    let first_moment = beta1 * moments.first.get(i) + (1.0 - beta1) * gradient[i];
                    let second_moment =
                        beta2 * moments.second.get(i) + (1.0 - beta2) * gradient[i] * gradient[i];
                    moments.first.set(i, first_moment);
                    moments.second.set(i, second_moment);
                    let first_hat = first_moment / first_correction;
                    let second_hat = second_moment / second_correction;
                    let adam_update = first_hat / (second_hat.sqrt() + epsilon);
                    values[i] -= effective_lr * (adam_update + group.weight_decay * values[i]);
                }
            });
        }
    }

    /// `p -= lr * (g + wd * p)`, or with momentum: `v = m*v + g; p -= lr*(v + wd*p)`.
    fn step_sgd(&mut self, parameters: &[Tensor], learning_rate: f32) {
        let momentum = self.config.momentum;

        for (index, (parameter, moments)) in
            parameters.iter().zip(self.state.moments.iter_mut()).enumerate()
        {
            let Some(gradient) = parameter.grad() else {
                continue;
            };
            let group = self.groups[index];
            let effective_lr = learning_rate * group.lr_scale;
            parameter.with_mut_data(|values| {
                for i in 0..values.len() {
                    let update = if momentum > 0.0 {
                        let velocity = momentum * moments.first.get(i) + gradient[i];
                        moments.first.set(i, velocity);
                        velocity
                    } else {
                        gradient[i]
                    };
                    values[i] -= effective_lr * (update + group.weight_decay * values[i]);
                }
            });
        }
    }

    /// Lion (Chen et al.): `update = sign(beta1*m + (1-beta1)*g)`,
    /// `m = beta2*m + (1-beta2)*g`, `p -= lr * (update + wd*p)`.
    fn step_lion(&mut self, parameters: &[Tensor], learning_rate: f32) {
        let beta1 = self.config.beta1;
        let beta2 = self.config.beta2;

        for (index, (parameter, moments)) in
            parameters.iter().zip(self.state.moments.iter_mut()).enumerate()
        {
            let Some(gradient) = parameter.grad() else {
                continue;
            };
            let group = self.groups[index];
            let effective_lr = learning_rate * group.lr_scale;
            parameter.with_mut_data(|values| {
                for i in 0..values.len() {
                    let momentum = moments.first.get(i);
                    let update = (beta1 * momentum + (1.0 - beta1) * gradient[i]).signum();
                    let new_momentum = beta2 * momentum + (1.0 - beta2) * gradient[i];
                    moments.first.set(i, new_momentum);
                    values[i] -= effective_lr * (update + group.weight_decay * values[i]);
                }
            });
        }
    }

    /// Simplified factored Adafactor. For 2D parameters of shape `[rows, cols]`,
    /// maintains row/col second-moment factors and reconstructs the full second
    /// moment as `outer(row, col) / mean(row)`. For non-2D parameters, falls back
    /// to a plain (unfactored) RMSprop-style second moment.
    fn step_adafactor(&mut self, parameters: &[Tensor], learning_rate: f32) {
        let beta2 = self.config.beta2;
        let epsilon = self.config.epsilon;

        for (index, (parameter, moments)) in
            parameters.iter().zip(self.state.moments.iter_mut()).enumerate()
        {
            let Some(gradient) = parameter.grad() else {
                continue;
            };
            let group = self.groups[index];
            let effective_lr = learning_rate * group.lr_scale;
            let shape = parameter.shape();

            if shape.len() == 2 && shape[0] > 0 && shape[1] > 0 {
                let rows = shape[0];
                let cols = shape[1];
                let mut row_means = vec![0.0f32; rows];
                let mut col_means = vec![0.0f32; cols];
                for r in 0..rows {
                    for c in 0..cols {
                        let g = gradient[r * cols + c];
                        let g2 = g * g;
                        row_means[r] += g2;
                        col_means[c] += g2;
                    }
                }
                for value in row_means.iter_mut() {
                    *value /= cols as f32;
                }
                for value in col_means.iter_mut() {
                    *value /= rows as f32;
                }
                for r in 0..rows {
                    let vr = beta2 * moments.first.get(r) + (1.0 - beta2) * row_means[r];
                    moments.first.set(r, vr);
                }
                for c in 0..cols {
                    let vc = beta2 * moments.second.get(c) + (1.0 - beta2) * col_means[c];
                    moments.second.set(c, vc);
                }
                let row_factor_mean: f32 =
                    (0..rows).map(|r| moments.first.get(r)).sum::<f32>() / rows as f32;
                let normalizer = row_factor_mean.max(epsilon);
                parameter.with_mut_data(|values| {
                    for r in 0..rows {
                        let vr = moments.first.get(r);
                        for c in 0..cols {
                            let vc = moments.second.get(c);
                            let flat = r * cols + c;
                            let v = (vr * vc / normalizer).max(0.0);
                            let update = gradient[flat] / (v.sqrt() + epsilon);
                            values[flat] -=
                                effective_lr * (update + group.weight_decay * values[flat]);
                        }
                    }
                });
            } else {
                parameter.with_mut_data(|values| {
                    for i in 0..values.len() {
                        let g = gradient[i];
                        let second_moment = beta2 * moments.second.get(i) + (1.0 - beta2) * g * g;
                        moments.second.set(i, second_moment);
                        let update = g / (second_moment.sqrt() + epsilon);
                        values[i] -= effective_lr * (update + group.weight_decay * values[i]);
                    }
                });
            }
        }
    }

    /// GPU-accelerated AdamW step. Only supports `OptimizerKind::AdamW` with
    /// `MomentDtype::Fp32`; SGD, Lion, and Adafactor are CPU-only (see [`AdamW::step`]),
    /// as is FP16 moment storage, since neither has a GPU kernel implementation yet.
    #[cfg(feature = "gpu")]
    pub fn step_gpu(
        &mut self,
        parameters: &[Tensor],
        learning_rate: f32,
        gradient_scale: f32,
        gpu: &GpuContext,
    ) -> Result<(), OptimizerError> {
        if self.config.kind != OptimizerKind::AdamW {
            return Err(OptimizerError(format!(
                "GPU optimizer step only supports AdamW (got {:?}); SGD/Lion/Adafactor are CPU-only, use step() instead",
                self.config.kind
            )));
        }
        if self.config.moment_dtype != MomentDtype::Fp32 {
            return Err(OptimizerError(
                "GPU optimizer step only supports FP32 moments; FP16 moments are CPU-only, use step() instead"
                    .into(),
            ));
        }
        if !learning_rate.is_finite() || learning_rate < 0.0 {
            return Err(OptimizerError(
                "learning rate must be finite and non-negative".into(),
            ));
        }
        self.page_in_moments()?;
        validate_state(parameters, &self.config, &self.state)?;
        validate_groups(parameters.len(), &self.groups)?;
        let needs_upload = self.gpu_moments.as_ref().is_none_or(|moments| {
            moments
                .first()
                .is_some_and(|(first, _)| !first.belongs_to(gpu))
        });
        if needs_upload {
            let mut uploaded = Vec::with_capacity(self.state.moments.len());
            for moments in &self.state.moments {
                let first_slice = moments.first.as_fp32().ok_or_else(|| {
                    OptimizerError("GPU optimizer step requires FP32 moment storage".into())
                })?;
                let second_slice = moments.second.as_fp32().ok_or_else(|| {
                    OptimizerError("GPU optimizer step requires FP32 moment storage".into())
                })?;
                let first = gpu
                    .upload(first_slice, vec![first_slice.len()])
                    .map_err(|error| OptimizerError(error.to_string()))?;
                let second = gpu
                    .upload(second_slice, vec![second_slice.len()])
                    .map_err(|error| OptimizerError(error.to_string()))?;
                uploaded.push((first, second));
            }
            self.gpu_moments = Some(uploaded);
        }

        self.state.step += 1;
        let step = self.state.step as f32;
        let first_correction = 1.0 - self.config.beta1.powf(step);
        let second_correction = 1.0 - self.config.beta2.powf(step);
        let gpu_moments = self.gpu_moments.as_ref().unwrap();
        for (index, (parameter, (first, second))) in parameters.iter().zip(gpu_moments).enumerate()
        {
            let Some(gradient) = parameter.gpu_grad() else {
                continue;
            };
            let group = self.groups[index];
            let parameter_storage = parameter
                .gpu_storage(gpu)
                .map_err(|error| OptimizerError(error.to_string()))?;
            gpu.adamw(
                &parameter_storage,
                &gradient,
                first,
                second,
                AdamWParams {
                    length: u32::try_from(parameter.numel()).map_err(|_| {
                        OptimizerError(format!("parameter {index} exceeds GPU limits"))
                    })?,
                    padding_0: 0,
                    padding_1: 0,
                    padding_2: 0,
                    learning_rate: learning_rate * group.lr_scale,
                    beta1: self.config.beta1,
                    beta2: self.config.beta2,
                    epsilon: self.config.epsilon,
                    weight_decay: group.weight_decay,
                    first_correction,
                    second_correction,
                    gradient_scale,
                },
            )
            .map_err(|error| OptimizerError(error.to_string()))?;
            parameter.mark_gpu_authoritative();
        }
        Ok(())
    }

    #[cfg(not(feature = "gpu"))]
    pub fn step_gpu(
        &mut self,
        _parameters: &[Tensor],
        _learning_rate: f32,
        _gradient_scale: f32,
        _gpu: &GpuContext,
    ) -> Result<(), OptimizerError> {
        Err(OptimizerError(
            "GPU support is not compiled in; rebuild with `--features gpu`".into(),
        ))
    }

    /// Write moment buffers to a temp file under
    /// `std::env::temp_dir()/oxidelm-moments-{pid}/` and free the in-memory copy.
    /// No-op if moments are already paged out.
    pub fn page_out_moments(&mut self) -> Result<(), OptimizerError> {
        if self.moments_paged_out {
            return Ok(());
        }
        let path = self.ensure_offload_path()?;
        let file = fs::File::create(&path).map_err(|error| {
            OptimizerError(format!(
                "failed to create moment offload file {}: {error}",
                path.display()
            ))
        })?;
        serde_json::to_writer(BufWriter::new(file), &self.state.moments).map_err(|error| {
            OptimizerError(format!("failed to write moments to offload file: {error}"))
        })?;
        self.state.moments = Vec::new();
        self.moments_paged_out = true;
        Ok(())
    }

    /// Reload moment buffers from the offload file written by [`AdamW::page_out_moments`].
    /// No-op if moments are already in memory.
    pub fn page_in_moments(&mut self) -> Result<(), OptimizerError> {
        if !self.moments_paged_out {
            return Ok(());
        }
        let path = self.offload_path.clone().ok_or_else(|| {
            OptimizerError("moments are marked paged out but no offload path was recorded".into())
        })?;
        let file = fs::File::open(&path).map_err(|error| {
            OptimizerError(format!(
                "failed to open moment offload file {}: {error}",
                path.display()
            ))
        })?;
        let moments: Vec<ParameterMoments> = serde_json::from_reader(BufReader::new(file))
            .map_err(|error| {
                OptimizerError(format!("failed to read moment offload file: {error}"))
            })?;
        self.state.moments = moments;
        self.moments_paged_out = false;
        Ok(())
    }

    fn ensure_offload_path(&mut self) -> Result<PathBuf, OptimizerError> {
        if let Some(path) = &self.offload_path {
            return Ok(path.clone());
        }
        let directory =
            std::env::temp_dir().join(format!("oxidelm-moments-{}", std::process::id()));
        fs::create_dir_all(&directory).map_err(|error| {
            OptimizerError(format!(
                "failed to create moment offload directory {}: {error}",
                directory.display()
            ))
        })?;
        let path = directory.join(format!("optimizer-{}.json", self.offload_id));
        self.offload_path = Some(path.clone());
        Ok(path)
    }

    pub fn checkpoint_state(&self) -> Result<AdamWState, OptimizerError> {
        if let Some(gpu_moments) = &self.gpu_moments {
            let moments = gpu_moments
                .iter()
                .map(|(first, second)| {
                    Ok(ParameterMoments {
                        first: MomentStorage::Fp32(
                            first
                                .to_vec()
                                .map_err(|error| OptimizerError(error.to_string()))?,
                        ),
                        second: MomentStorage::Fp32(
                            second
                                .to_vec()
                                .map_err(|error| OptimizerError(error.to_string()))?,
                        ),
                    })
                })
                .collect::<Result<Vec<_>, OptimizerError>>()?;
            return Ok(AdamWState {
                step: self.state.step,
                moments,
            });
        }
        if self.moments_paged_out {
            let path = self.offload_path.as_ref().ok_or_else(|| {
                OptimizerError(
                    "moments are marked paged out but no offload path was recorded".into(),
                )
            })?;
            let file = fs::File::open(path).map_err(|error| {
                OptimizerError(format!(
                    "failed to open moment offload file {}: {error}",
                    path.display()
                ))
            })?;
            let moments: Vec<ParameterMoments> = serde_json::from_reader(BufReader::new(file))
                .map_err(|error| {
                    OptimizerError(format!("failed to read moment offload file: {error}"))
                })?;
            return Ok(AdamWState {
                step: self.state.step,
                moments,
            });
        }
        Ok(self.state.clone())
    }

    pub fn config(&self) -> &AdamWConfig {
        &self.config
    }

    pub fn state(&self) -> &AdamWState {
        &self.state
    }

    pub fn groups(&self) -> &[ParameterGroup] {
        &self.groups
    }
}

impl Drop for AdamW {
    fn drop(&mut self) {
        if let Some(path) = self.offload_path.take() {
            let _ = fs::remove_file(path);
        }
    }
}

/// Expected `(first.len(), second.len())` for a parameter's moments, given its
/// shape/numel and the optimizer config. See [`ParameterMoments`] for the
/// per-kind layout this encodes.
fn moment_layout(shape: &[usize], numel: usize, config: &AdamWConfig) -> (usize, usize) {
    match config.kind {
        OptimizerKind::AdamW => (numel, numel),
        OptimizerKind::Sgd => {
            if config.momentum > 0.0 {
                (numel, 0)
            } else {
                (0, 0)
            }
        }
        OptimizerKind::Lion => (numel, 0),
        OptimizerKind::Adafactor => {
            if shape.len() == 2 {
                (shape[0], shape[1])
            } else {
                (0, numel)
            }
        }
    }
}

fn moments_for(parameter: &Tensor, config: &AdamWConfig) -> ParameterMoments {
    let (first_len, second_len) = moment_layout(&parameter.shape(), parameter.numel(), config);
    let dtype = match config.kind {
        OptimizerKind::AdamW | OptimizerKind::Lion => config.moment_dtype,
        OptimizerKind::Sgd | OptimizerKind::Adafactor => MomentDtype::Fp32,
    };
    ParameterMoments {
        first: MomentStorage::zeros(first_len, dtype),
        second: MomentStorage::zeros(second_len, dtype),
    }
}

fn validate_config(config: &AdamWConfig) -> Result<(), OptimizerError> {
    if !(0.0..1.0).contains(&config.beta1) || !(0.0..1.0).contains(&config.beta2) {
        return Err(OptimizerError("betas must be in [0, 1)".into()));
    }
    if config.epsilon <= 0.0 || config.weight_decay < 0.0 {
        return Err(OptimizerError(
            "epsilon must be positive and weight decay non-negative".into(),
        ));
    }
    if !config.rms_norm_lr_scale.is_finite()
        || config.rms_norm_lr_scale < 0.0
        || !config.embedding_lr_scale.is_finite()
        || config.embedding_lr_scale < 0.0
    {
        return Err(OptimizerError(
            "parameter-group learning-rate scales must be finite and non-negative".into(),
        ));
    }
    if !config.momentum.is_finite() || !(0.0..1.0).contains(&config.momentum) {
        return Err(OptimizerError(
            "momentum must be finite and in [0, 1)".into(),
        ));
    }
    if config.moment_dtype == MomentDtype::Fp16
        && !matches!(config.kind, OptimizerKind::AdamW | OptimizerKind::Lion)
    {
        return Err(OptimizerError(
            "FP16 moment storage is only supported for AdamW and Lion; Adafactor always uses \
             factored FP32 moments and SGD state is FP32-only"
                .into(),
        ));
    }
    Ok(())
}

fn validate_groups(parameter_count: usize, groups: &[ParameterGroup]) -> Result<(), OptimizerError> {
    if groups.len() != parameter_count {
        return Err(OptimizerError(
            "parameter count does not match optimizer groups".into(),
        ));
    }
    for (index, group) in groups.iter().enumerate() {
        if !group.weight_decay.is_finite()
            || group.weight_decay < 0.0
            || !group.lr_scale.is_finite()
            || group.lr_scale < 0.0
        {
            return Err(OptimizerError(format!(
                "optimizer group {index} has invalid weight decay or learning-rate scale"
            )));
        }
    }
    Ok(())
}

fn validate_state(
    parameters: &[Tensor],
    config: &AdamWConfig,
    state: &AdamWState,
) -> Result<(), OptimizerError> {
    if parameters.len() != state.moments.len() {
        return Err(OptimizerError(
            "parameter count does not match optimizer state".into(),
        ));
    }
    for (index, (parameter, moments)) in parameters.iter().zip(state.moments.iter()).enumerate() {
        let (expected_first, expected_second) =
            moment_layout(&parameter.shape(), parameter.numel(), config);
        if moments.first.len() != expected_first || moments.second.len() != expected_second {
            return Err(OptimizerError(format!(
                "optimizer moment length mismatch for parameter {index}"
            )));
        }
    }
    Ok(())
}

#[cfg(test)]
mod tests {
    use super::{
        default_no_decay_embeddings, AdamW, AdamWConfig, MomentDtype, OptimizerKind,
        ParameterGroup,
    };
    use crate::Tensor;

    #[test]
    fn adamw_reduces_simple_quadratic_parameter() {
        let parameter = Tensor::scalar(2.0, true);
        let mut optimizer = AdamW::new(
            &[parameter.clone()],
            AdamWConfig {
                weight_decay: 0.0,
                ..Default::default()
            },
        )
        .unwrap();

        parameter.powf(2.0).backward();
        optimizer.step(&[parameter.clone()], 0.1).unwrap();

        assert!(parameter.item() < 2.0);
        assert!((parameter.item() - 1.9).abs() < 1e-4);
    }

    #[test]
    fn grouped_config_excludes_norms_and_selected_embeddings_from_decay() {
        let config = AdamWConfig {
            weight_decay: 0.1,
            no_decay_rms_norm: true,
            no_decay_embeddings: vec!["tok_embeddings.weight".into()],
            rms_norm_lr_scale: 0.5,
            embedding_lr_scale: 0.25,
            ..Default::default()
        };

        assert_eq!(
            config.group_for("layers.0.attention.q_proj.weight"),
            ParameterGroup {
                weight_decay: 0.1,
                lr_scale: 1.0
            }
        );
        assert_eq!(
            config.group_for("layers.0.attention_norm.weight"),
            ParameterGroup {
                weight_decay: 0.0,
                lr_scale: 0.5
            }
        );
        assert_eq!(
            config.group_for("norm.weight"),
            ParameterGroup {
                weight_decay: 0.0,
                lr_scale: 0.5
            }
        );
        assert_eq!(
            config.group_for("tok_embeddings.weight"),
            ParameterGroup {
                weight_decay: 0.0,
                lr_scale: 0.25
            }
        );
        assert_eq!(
            config.group_for("lm_head.weight"),
            ParameterGroup {
                weight_decay: 0.1,
                lr_scale: 0.25
            }
        );
    }

    #[test]
    fn new_grouped_applies_zero_weight_decay_to_embeddings() {
        let embedding = Tensor::new(vec![1.0, -1.0], vec![2], true);
        let projection = Tensor::new(vec![1.0, -1.0], vec![2], true);
        let named = vec![
            ("tok_embeddings.weight".into(), embedding.clone()),
            ("layers.0.attention.q_proj.weight".into(), projection.clone()),
        ];
        let mut optimizer = AdamW::new_grouped(
            &named,
            AdamWConfig {
                weight_decay: 0.5,
                no_decay_embeddings: default_no_decay_embeddings(),
                ..Default::default()
            },
        )
        .unwrap();

        let zero = Tensor::scalar(0.0, false);
        embedding
            .mul(&zero)
            .add(&projection.mul(&zero))
            .sum()
            .backward();
        let parameters = vec![embedding.clone(), projection.clone()];
        optimizer.step(&parameters, 1.0).unwrap();

        assert_eq!(embedding.data(), vec![1.0, -1.0]);
        assert_eq!(projection.data(), vec![0.5, -0.5]);
    }

    #[test]
    fn new_grouped_scales_learning_rate_per_group() {
        let embedding = Tensor::scalar(1.0, true);
        let projection = Tensor::scalar(1.0, true);
        let named = vec![
            ("tok_embeddings.weight".into(), embedding.clone()),
            ("layers.0.attention.q_proj.weight".into(), projection.clone()),
        ];
        let mut optimizer = AdamW::new_grouped(
            &named,
            AdamWConfig {
                weight_decay: 0.0,
                embedding_lr_scale: 0.5,
                ..Default::default()
            },
        )
        .unwrap();

        embedding.add(&projection).backward();
        let parameters = vec![embedding.clone(), projection.clone()];
        optimizer.step(&parameters, 0.1).unwrap();

        let embedding_delta = (1.0 - embedding.item()).abs();
        let projection_delta = (1.0 - projection.item()).abs();
        assert!((embedding_delta * 2.0 - projection_delta).abs() < 1e-5);
    }

    #[test]
    fn parameter_kind_classifies_norm_and_embedding_names() {
        assert!(matches!(
            super::parameter_kind("layers.3.ffn_norm.weight"),
            super::ParameterKind::RmsNorm
        ));
        assert!(matches!(
            super::parameter_kind("tok_embeddings.weight"),
            super::ParameterKind::Embedding
        ));
        assert!(matches!(
            super::parameter_kind("layers.0.attention.q_proj.weight"),
            super::ParameterKind::Other
        ));
    }

    #[test]
    fn fp16_adamw_moments_closely_track_fp32() {
        let initial = vec![2.0f32, -3.0, 0.5, 1.25];
        let fp32_param = Tensor::new(initial.clone(), vec![4], true);
        let fp16_param = Tensor::new(initial, vec![4], true);

        let mut fp32_optimizer = AdamW::new(
            &[fp32_param.clone()],
            AdamWConfig {
                weight_decay: 0.0,
                ..Default::default()
            },
        )
        .unwrap();
        let mut fp16_optimizer = AdamW::new(
            &[fp16_param.clone()],
            AdamWConfig {
                weight_decay: 0.0,
                moment_dtype: MomentDtype::Fp16,
                ..Default::default()
            },
        )
        .unwrap();

        for _ in 0..8 {
            fp32_param.zero_grad();
            fp32_param.powf(2.0).sum().backward();
            fp32_optimizer.step(&[fp32_param.clone()], 0.05).unwrap();

            fp16_param.zero_grad();
            fp16_param.powf(2.0).sum().backward();
            fp16_optimizer.step(&[fp16_param.clone()], 0.05).unwrap();
        }

        for (fp32_value, fp16_value) in fp32_param.data().iter().zip(fp16_param.data().iter()) {
            let diff = (fp32_value - fp16_value).abs();
            let tolerance = 1e-2 * fp32_value.abs().max(1.0);
            assert!(
                diff < tolerance,
                "fp32={fp32_value} fp16={fp16_value} diff={diff} tolerance={tolerance}"
            );
        }
    }

    #[test]
    fn sgd_reduces_simple_quadratic() {
        let parameter = Tensor::scalar(4.0, true);
        let mut optimizer = AdamW::new(
            &[parameter.clone()],
            AdamWConfig {
                kind: OptimizerKind::Sgd,
                weight_decay: 0.0,
                momentum: 0.9,
                ..Default::default()
            },
        )
        .unwrap();

        for _ in 0..20 {
            parameter.zero_grad();
            parameter.powf(2.0).backward();
            optimizer.step(&[parameter.clone()], 0.05).unwrap();
        }

        assert!(parameter.item().abs() < 4.0);
        assert!(parameter.item().is_finite());
    }

    #[test]
    fn lion_reduces_simple_quadratic() {
        let parameter = Tensor::scalar(3.0, true);
        let mut optimizer = AdamW::new(
            &[parameter.clone()],
            AdamWConfig {
                kind: OptimizerKind::Lion,
                weight_decay: 0.0,
                ..Default::default()
            },
        )
        .unwrap();

        for _ in 0..20 {
            parameter.zero_grad();
            parameter.powf(2.0).backward();
            optimizer.step(&[parameter.clone()], 0.05).unwrap();
        }

        assert!(parameter.item().abs() < 3.0);
        assert!(parameter.item().is_finite());
    }

    #[test]
    fn adafactor_reduces_2d_quadratic() {
        let values = vec![2.0f32, -1.5, 0.8, -0.3, 1.2, -2.1];
        let parameter = Tensor::new(values, vec![2, 3], true);
        let mut optimizer = AdamW::new(
            &[parameter.clone()],
            AdamWConfig {
                kind: OptimizerKind::Adafactor,
                weight_decay: 0.0,
                ..Default::default()
            },
        )
        .unwrap();

        let initial_loss: f32 = parameter.data().iter().map(|v| v * v).sum();

        for _ in 0..30 {
            parameter.zero_grad();
            parameter.powf(2.0).sum().backward();
            optimizer.step(&[parameter.clone()], 0.1).unwrap();
        }

        let final_loss: f32 = parameter.data().iter().map(|v| v * v).sum();
        assert!(
            final_loss < initial_loss,
            "expected loss to decrease: initial={initial_loss} final={final_loss}"
        );
        assert!(parameter.data().iter().all(|value| value.is_finite()));
    }

    #[test]
    fn offload_moments_round_trips_through_disk_between_steps() {
        let parameter = Tensor::scalar(2.0, true);
        let mut optimizer = AdamW::new(
            &[parameter.clone()],
            AdamWConfig {
                weight_decay: 0.0,
                offload_moments: true,
                ..Default::default()
            },
        )
        .unwrap();

        parameter.powf(2.0).backward();
        optimizer.step(&[parameter.clone()], 0.1).unwrap();
        let after_first_step = parameter.item();
        assert!(after_first_step < 2.0);
        assert!(optimizer.moments_paged_out);
        let offload_path = optimizer
            .offload_path
            .clone()
            .expect("offload path should be recorded once paged out");
        assert!(
            offload_path.exists(),
            "moment offload file should exist between steps"
        );

        parameter.zero_grad();
        parameter.powf(2.0).backward();
        optimizer.step(&[parameter.clone()], 0.1).unwrap();
        assert!(parameter.item() < after_first_step);
        assert!(optimizer.moments_paged_out);
    }

    #[test]
    fn validate_rejects_fp16_moments_outside_adamw_and_lion() {
        let mut config = AdamWConfig {
            kind: OptimizerKind::Adafactor,
            moment_dtype: MomentDtype::Fp16,
            ..Default::default()
        };
        assert!(config.validate().is_err());

        config.kind = OptimizerKind::Sgd;
        assert!(config.validate().is_err());

        config.kind = OptimizerKind::Lion;
        assert!(config.validate().is_ok());

        config.kind = OptimizerKind::AdamW;
        assert!(config.validate().is_ok());
    }

    #[test]
    fn moments_are_sized_according_to_optimizer_kind() {
        let parameter = Tensor::new(vec![0.0; 6], vec![2, 3], true);

        let sgd_no_momentum = AdamW::new(
            &[parameter.clone()],
            AdamWConfig {
                kind: OptimizerKind::Sgd,
                momentum: 0.0,
                ..Default::default()
            },
        )
        .unwrap();
        assert_eq!(sgd_no_momentum.state().moments[0].first.len(), 0);
        assert_eq!(sgd_no_momentum.state().moments[0].second.len(), 0);

        let sgd_with_momentum = AdamW::new(
            &[parameter.clone()],
            AdamWConfig {
                kind: OptimizerKind::Sgd,
                momentum: 0.9,
                ..Default::default()
            },
        )
        .unwrap();
        assert_eq!(sgd_with_momentum.state().moments[0].first.len(), 6);
        assert_eq!(sgd_with_momentum.state().moments[0].second.len(), 0);

        let lion = AdamW::new(
            &[parameter.clone()],
            AdamWConfig {
                kind: OptimizerKind::Lion,
                ..Default::default()
            },
        )
        .unwrap();
        assert_eq!(lion.state().moments[0].first.len(), 6);
        assert_eq!(lion.state().moments[0].second.len(), 0);

        let adafactor = AdamW::new(
            &[parameter.clone()],
            AdamWConfig {
                kind: OptimizerKind::Adafactor,
                ..Default::default()
            },
        )
        .unwrap();
        assert_eq!(adafactor.state().moments[0].first.len(), 2);
        assert_eq!(adafactor.state().moments[0].second.len(), 3);

        let adamw = AdamW::new(&[parameter], AdamWConfig::default()).unwrap();
        assert_eq!(adamw.state().moments[0].first.len(), 6);
        assert_eq!(adamw.state().moments[0].second.len(), 6);
    }
}