rusting_brain 5.0.0

A Rust deep-learning library for training and running transformer language models, dense and mixture-of-experts, on CPU or CUDA.
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
//! Low-rank adapters on a UNet's attention projections, and the backward pass
//! that trains them.
//!
//! This is the host reference the device path is checked against, as
//! [`crate::dit_lora`] is for the DiT. Adapters sit where diffusers' own
//! `train_text_to_image_lora` puts them: `to_q`, `to_k`, `to_v` and
//! `to_out.0` of both attentions in every transformer block. The base weights
//! and the conditioning are frozen, so the backward keeps only adapter
//! gradients, but it still runs back through every resnet, sampler and skip,
//! because an adapter on the way down changes everything after it.
//!
//! NOTE: no ControlNet, no inpainting channels; they are frozen inputs a
//! training step can add once someone trains with them.

use crate::conv::{Conv2d, Dense, FeatureMap, GroupNorm, silu, upsample_nearest};
use crate::dit_lora::{attention_backward, layer_norm_backward};
use crate::ffn::{gelu, gelu_derivative, silu_derivative};
use crate::matrix::Matrix;
use crate::mmdit::{Conditioning, attention};
use crate::network::NetworkError;
use crate::param::{Lora, Param};
use crate::unet::{Attention, Projection, Resnet, Transformer, Unet, concatenate};
use std::collections::BTreeMap;

/// Adapters by the diffusers name of the projection they add to, such as
/// `down_blocks.0.attentions.1.transformer_blocks.0.attn2.to_k`.
#[derive(Clone, Debug, Default)]
pub struct UnetLora {
    pub adapters: BTreeMap<String, Lora>,
}

pub(crate) const PROJECTIONS: [&str; 4] = ["to_q", "to_k", "to_v", "to_out.0"];

/// Every transformer in the model, under its diffusers name.
pub(crate) fn transformers(unet: &Unet) -> Vec<(String, &Transformer)> {
    let mut list = Vec::new();
    for (level, block) in unet.encoder.down.iter().enumerate() {
        for (index, transformer) in block.attentions.iter().enumerate() {
            list.push((
                format!("down_blocks.{level}.attentions.{index}"),
                transformer,
            ));
        }
    }
    list.push(("mid_block.attentions.0".into(), &unet.encoder.middle.1));
    for (level, block) in unet.up.iter().enumerate() {
        for (index, transformer) in block.attentions.iter().enumerate() {
            list.push((format!("up_blocks.{level}.attentions.{index}"), transformer));
        }
    }
    list
}

/// Every projection an adapter can sit on, by name.
fn projections(unet: &Unet) -> BTreeMap<String, &Dense> {
    let mut map = BTreeMap::new();
    for (prefix, transformer) in transformers(unet) {
        for (index, block) in transformer.blocks.iter().enumerate() {
            for (name, attention) in [("attn1", &block.attention), ("attn2", &block.cross)] {
                let denses = [
                    &attention.query,
                    &attention.key,
                    &attention.value,
                    &attention.output,
                ];
                for (projection, dense) in PROJECTIONS.iter().zip(denses) {
                    let key = format!("{prefix}.transformer_blocks.{index}.{name}.{projection}");
                    map.insert(key, dense);
                }
            }
        }
    }
    map
}

impl UnetLora {
    /// Reads a PEFT file trained against a diffusers UNet (`{module}.lora_A`
    /// and `lora_B`, optionally under `unet.`), with `scale` the
    /// `alpha / rank` it was trained at.
    pub fn read_peft(
        path: impl AsRef<std::path::Path>,
        unet: &Unet,
        scale: f32,
    ) -> Result<Self, NetworkError> {
        let targets = projections(unet);
        let mut file = crate::safetensors::SafeTensors::open(path)?;
        let names: Vec<String> = file.names().map(str::to_string).collect();
        let mut result = Self::default();
        for name in &names {
            let Some(module) = name.strip_suffix(".lora_A.weight") else {
                continue;
            };
            let module = module.strip_prefix("unet.").unwrap_or(module);
            let dense = targets.get(module).ok_or_else(|| {
                NetworkError::InvalidConfig(format!("{name} adapts no attention projection"))
            })?;
            let down = file.matrix(name)?;
            let up = file.matrix(&name.replace(".lora_A.", ".lora_B."))?;
            if down.cols != dense.in_dim() || up.rows != dense.out_dim() || up.cols != down.rows {
                return Err(NetworkError::InvalidConfig(format!(
                    "the adapter on {module} does not fit its {}x{} projection",
                    dense.out_dim(),
                    dense.in_dim()
                )));
            }
            let lora = Lora {
                down: Param::new(down),
                up: Param::new(up),
                scale,
            };
            result.adapters.insert(module.to_string(), lora);
        }
        Ok(result)
    }

    /// Fresh adapters of `rank` on every attention projection, as PEFT starts
    /// them: `down` random, `up` zero, so the model is unchanged until the
    /// first step, and `alpha / rank` the scale.
    pub fn new(unet: &Unet, rank: usize, alpha: f32, rng: &mut rand::rngs::StdRng) -> Self {
        let adapters = projections(unet)
            .into_iter()
            .map(|(name, dense)| {
                let lora = Lora::new(dense.in_dim(), dense.out_dim(), rank, alpha, rng);
                (name, lora)
            })
            .collect();
        Self { adapters }
    }

    /// Writes the adapters as a diffusers PEFT file (`unet.{module}.lora_A`
    /// and `lora_B`), each adapter's scale folded into `lora_B`, so
    /// [`read_peft`](Self::read_peft) at scale 1 and diffusers'
    /// `load_lora_weights` (alpha defaults to the rank) give back the same
    /// model.
    pub fn write_peft(&self, path: impl AsRef<std::path::Path>) -> Result<(), NetworkError> {
        let mut tensors = Vec::new();
        for (module, lora) in &self.adapters {
            let (down, up) = (&lora.down.value, &lora.up.value);
            tensors.push((
                format!("unet.{module}.lora_A.weight"),
                vec![down.rows, down.cols],
                down.data.clone(),
            ));
            tensors.push((
                format!("unet.{module}.lora_B.weight"),
                vec![up.rows, up.cols],
                up.data.iter().map(|value| value * lora.scale).collect(),
            ));
        }
        let views: Vec<_> = tensors
            .iter()
            .map(|(name, shape, data)| (name.as_str(), &shape[..], &data[..]))
            .collect();
        crate::safetensors::write(path, &views, &BTreeMap::new())
    }

    /// Every adapter factor, for an optimizer.
    pub fn params_mut(&mut self) -> Vec<&mut Param> {
        self.adapters
            .values_mut()
            .flat_map(Lora::params_mut)
            .collect()
    }

    /// [`Unet::forward`] with the adapters added, and what the backward pass
    /// needs of it.
    pub fn forward(
        &self,
        unet: &Unet,
        latent: &FeatureMap,
        sigma: f32,
        conditioning: &Conditioning,
    ) -> Result<(FeatureMap, Tape), NetworkError> {
        let config = unet.config();
        let (scaled, time) = prepare(unet, latent, sigma, conditioning)?;
        let text = &conditioning.text;
        let eps = config.eps;
        let names: Vec<String> = transformers(unet)
            .into_iter()
            .map(|(name, _)| name)
            .collect();
        let mut names = names.iter();
        let mut run = |transformer: &Transformer, sample: &FeatureMap| {
            let name = names.next().expect("one name a transformer");
            self.transformer(transformer, name, sample, text, eps)
        };

        let mut sample = unet.encoder.conv_in.forward(&scaled)?;
        let mut skips = vec![sample.clone()];
        let mut down = Vec::new();
        for block in &unet.encoder.down {
            let mut rungs = Vec::new();
            for (index, resnet) in block.resnets.iter().enumerate() {
                let (output, resnet_tape) = resnet_forward(resnet, &sample, &time)?;
                sample = output;
                let mut transformer_tape = None;
                if let Some(transformer) = block.attentions.get(index) {
                    let (output, tape) = run(transformer, &sample)?;
                    sample = output;
                    transformer_tape = Some(tape);
                }
                rungs.push((0, resnet_tape, transformer_tape));
                skips.push(sample.clone());
            }
            let resampled = match &block.resampler {
                Some(resampler) => {
                    let input = sample.clone();
                    sample = resampler.forward(&sample)?;
                    skips.push(sample.clone());
                    Some(input)
                }
                None => None,
            };
            down.push((rungs, resampled));
        }

        let (first, middle, second) = &unet.encoder.middle;
        let (output, first) = resnet_forward(first, &sample, &time)?;
        let (output, middle) = run(middle, &output)?;
        let (output, second) = resnet_forward(second, &output, &time)?;
        sample = output;
        let middle = (first, middle, second);

        let mut up = Vec::new();
        for block in &unet.up {
            let mut rungs = Vec::new();
            for (index, resnet) in block.resnets.iter().enumerate() {
                let skip = skips.pop().ok_or_else(|| {
                    NetworkError::InvalidConfig(
                        "the way up has more rungs than the way down".into(),
                    )
                })?;
                let channels = sample.channels;
                let (output, resnet_tape) =
                    resnet_forward(resnet, &concatenate(&sample, &skip), &time)?;
                sample = output;
                let mut transformer_tape = None;
                if let Some(transformer) = block.attentions.get(index) {
                    let (output, tape) = run(transformer, &sample)?;
                    sample = output;
                    transformer_tape = Some(tape);
                }
                rungs.push((channels, resnet_tape, transformer_tape));
            }
            let resampled = match &block.resampler {
                Some(resampler) => {
                    let input = upsample_nearest(&sample, 2);
                    sample = resampler.forward(&input)?;
                    Some(input)
                }
                None => None,
            };
            up.push((rungs, resampled));
        }
        if !skips.is_empty() {
            return Err(NetworkError::InvalidConfig(
                "the way down has rungs the way up never reads".into(),
            ));
        }

        let norm_input = sample.clone();
        unet.norm_out.forward(&mut sample)?;
        let normed = sample.clone();
        silu(&mut sample);
        let output = unet.conv_out.forward(&sample)?;
        let tape = Tape {
            down,
            middle,
            up,
            norm_input,
            normed,
        };
        Ok((output, tape))
    }

    /// Adds `dL/d(adapter)` to every adapter's gradients, given
    /// `dL/d(prediction)`.
    pub fn backward(
        &mut self,
        unet: &Unet,
        tape: &Tape,
        grad: &FeatureMap,
    ) -> Result<(), NetworkError> {
        let eps = unet.config().eps;
        let mut list = transformers(unet);
        let mut next = || list.pop().expect("one transformer a tape");

        let grad = conv_backward(&unet.conv_out, grad, &tape.normed);
        let grad = silu_backward(&grad, &tape.normed);
        let mut grad = group_norm_backward(&unet.norm_out, &tape.norm_input, &grad);

        // The way up hands back one gradient a skip, in the order the way down
        // pushed them.
        let mut skips = Vec::new();
        for (block, (rungs, resampled)) in unet.up.iter().zip(&tape.up).rev() {
            if let (Some(resampler), Some(input)) = (&block.resampler, resampled) {
                grad = upsample_backward(&conv_backward(resampler, &grad, input));
            }
            for (resnet, (channels, resnet_tape, transformer_tape)) in
                block.resnets.iter().zip(rungs).rev()
            {
                if let Some(transformer_tape) = transformer_tape {
                    let (name, transformer) = next();
                    grad =
                        self.transformer_backward(transformer, &name, transformer_tape, &grad, eps);
                }
                let joined = resnet_backward(resnet, resnet_tape, &grad);
                let split = channels * joined.pixels();
                skips.push(FeatureMap {
                    channels: joined.channels - channels,
                    height: joined.height,
                    width: joined.width,
                    data: joined.data[split..].to_vec(),
                });
                grad = FeatureMap {
                    channels: *channels,
                    height: joined.height,
                    width: joined.width,
                    data: joined.data[..split].to_vec(),
                };
            }
        }

        let (first, middle, second) = &unet.encoder.middle;
        grad = resnet_backward(second, &tape.middle.2, &grad);
        let (name, _) = next();
        grad = self.transformer_backward(middle, &name, &tape.middle.1, &grad, eps);
        grad = resnet_backward(first, &tape.middle.0, &grad);

        let mut add_skip = |grad: &mut FeatureMap| {
            let skip = skips.pop().expect("a skip gradient for every rung down");
            add_maps(grad, &skip);
        };
        for (block, (rungs, resampled)) in unet.encoder.down.iter().zip(&tape.down).rev() {
            if let (Some(resampler), Some(input)) = (&block.resampler, resampled) {
                add_skip(&mut grad);
                grad = conv_backward(resampler, &grad, input);
            }
            for (resnet, (_, resnet_tape, transformer_tape)) in
                block.resnets.iter().zip(rungs).rev()
            {
                add_skip(&mut grad);
                if let Some(transformer_tape) = transformer_tape {
                    let (name, transformer) = next();
                    grad =
                        self.transformer_backward(transformer, &name, transformer_tape, &grad, eps);
                }
                grad = resnet_backward(resnet, resnet_tape, &grad);
            }
        }
        // What is left is the gradient of `conv_in`'s output and the first
        // skip's, which reach only frozen weights and the input.
        Ok(())
    }

    fn transformer(
        &self,
        transformer: &Transformer,
        name: &str,
        input: &FeatureMap,
        text: &Matrix,
        eps: f32,
    ) -> Result<(FeatureMap, TransformerTape), NetworkError> {
        let (height, width) = (input.height, input.width);
        let mut normed = input.clone();
        transformer.norm.forward(&mut normed)?;
        let mut tokens = match &transformer.input {
            Projection::Convolution(conv) => conv.forward(&normed)?.to_tokens(),
            Projection::Linear(dense) => dense.forward(&normed.to_tokens())?,
        };
        let mut blocks = Vec::new();
        for (index, block) in transformer.blocks.iter().enumerate() {
            let prefix = format!("{name}.transformer_blocks.{index}");
            let before_self = tokens.clone();
            let normed1 = block.norm1.forward(&tokens, eps);
            let (attended, own) = self.attention(
                &block.attention,
                &format!("{prefix}.attn1"),
                &normed1,
                &normed1,
                transformer.heads,
            )?;
            add(&mut tokens, &attended);

            let before_cross = tokens.clone();
            let normed2 = block.norm2.forward(&tokens, eps);
            let (attended, cross) = self.attention(
                &block.cross,
                &format!("{prefix}.attn2"),
                &normed2,
                text,
                transformer.heads,
            )?;
            add(&mut tokens, &attended);

            let before_feed = tokens.clone();
            let normed3 = block.norm3.forward(&tokens, eps);
            let projected = block.gate.forward(&normed3)?;
            let hidden = geglu(&projected);
            add(&mut tokens, &block.output.forward(&hidden)?);
            blocks.push(BlockTape {
                before: [before_self, before_cross, before_feed],
                own,
                cross,
                projected,
            });
        }
        let mut output = match &transformer.output {
            Projection::Convolution(conv) => {
                conv.forward(&FeatureMap::from_tokens(&tokens, height, width)?)?
            }
            Projection::Linear(dense) => {
                FeatureMap::from_tokens(&dense.forward(&tokens)?, height, width)?
            }
        };
        add_maps(&mut output, input);
        let tape = TransformerTape {
            input: input.clone(),
            normed,
            blocks,
            tokens,
        };
        Ok((output, tape))
    }

    fn transformer_backward(
        &mut self,
        transformer: &Transformer,
        name: &str,
        tape: &TransformerTape,
        grad: &FeatureMap,
        eps: f32,
    ) -> FeatureMap {
        let (height, width) = (grad.height, grad.width);
        let mut grad_tokens = match &transformer.output {
            Projection::Convolution(conv) => {
                let input = FeatureMap::from_tokens(&tape.tokens, height, width).unwrap();
                conv_backward(conv, grad, &input).to_tokens()
            }
            Projection::Linear(dense) => dense_backward(dense, &grad.to_tokens()),
        };
        for (index, (block, block_tape)) in transformer
            .blocks
            .iter()
            .zip(&tape.blocks)
            .enumerate()
            .rev()
        {
            let prefix = format!("{name}.transformer_blocks.{index}");
            let [before_self, before_cross, before_feed] = &block_tape.before;

            let grad_hidden = dense_backward(&block.output, &grad_tokens);
            let grad_projected = geglu_backward(&block_tape.projected, &grad_hidden);
            let grad_normed = dense_backward(&block.gate, &grad_projected);
            add(
                &mut grad_tokens,
                &norm_backward(&block.norm3, before_feed, eps, &grad_normed),
            );

            let (grad_normed, _) = self.attention_backward(
                &block.cross,
                &format!("{prefix}.attn2"),
                &block_tape.cross,
                transformer.heads,
                &grad_tokens,
            );
            add(
                &mut grad_tokens,
                &norm_backward(&block.norm2, before_cross, eps, &grad_normed),
            );

            let (mut grad_normed, grad_source) = self.attention_backward(
                &block.attention,
                &format!("{prefix}.attn1"),
                &block_tape.own,
                transformer.heads,
                &grad_tokens,
            );
            add(&mut grad_normed, &grad_source);
            add(
                &mut grad_tokens,
                &norm_backward(&block.norm1, before_self, eps, &grad_normed),
            );
        }
        let grad_normed = match &transformer.input {
            Projection::Convolution(conv) => {
                let grad = FeatureMap::from_tokens(&grad_tokens, height, width).unwrap();
                conv_backward(conv, &grad, &tape.normed)
            }
            Projection::Linear(dense) => {
                FeatureMap::from_tokens(&dense_backward(dense, &grad_tokens), height, width)
                    .unwrap()
            }
        };
        let mut output = group_norm_backward(&transformer.norm, &tape.input, &grad_normed);
        add_maps(&mut output, grad);
        output
    }

    fn attention(
        &self,
        attention_layer: &Attention,
        name: &str,
        input: &Matrix,
        source: &Matrix,
        heads: usize,
    ) -> Result<(Matrix, AttentionTape), NetworkError> {
        let queries = self.project(&attention_layer.query, &format!("{name}.to_q"), input)?;
        let keys = self.project(&attention_layer.key, &format!("{name}.to_k"), source)?;
        let values = self.project(&attention_layer.value, &format!("{name}.to_v"), source)?;
        let attended = attention(&queries, &keys, &values, heads, heads, false);
        let output = self.project(
            &attention_layer.output,
            &format!("{name}.to_out.0"),
            &attended,
        )?;
        let tape = AttentionTape {
            input: input.clone(),
            source: source.clone(),
            queries,
            keys,
            values,
            attended,
        };
        Ok((output, tape))
    }

    /// `dL/d(input)` and `dL/d(source)`, adding the adapters' gradients.
    fn attention_backward(
        &mut self,
        attention_layer: &Attention,
        name: &str,
        tape: &AttentionTape,
        heads: usize,
        grad: &Matrix,
    ) -> (Matrix, Matrix) {
        let grad_attended = self.project_backward(
            &attention_layer.output,
            &format!("{name}.to_out.0"),
            &tape.attended,
            grad,
        );
        let (grad_queries, grad_keys, grad_values) = attention_backward(
            &tape.queries,
            &tape.keys,
            &tape.values,
            heads,
            &grad_attended,
        );
        let grad_input = self.project_backward(
            &attention_layer.query,
            &format!("{name}.to_q"),
            &tape.input,
            &grad_queries,
        );
        let mut grad_source = self.project_backward(
            &attention_layer.key,
            &format!("{name}.to_k"),
            &tape.source,
            &grad_keys,
        );
        add(
            &mut grad_source,
            &self.project_backward(
                &attention_layer.value,
                &format!("{name}.to_v"),
                &tape.source,
                &grad_values,
            ),
        );
        (grad_input, grad_source)
    }

    fn project(&self, dense: &Dense, name: &str, input: &Matrix) -> Result<Matrix, NetworkError> {
        let mut output = dense.forward(input)?;
        if let Some(lora) = self.adapters.get(name) {
            lora.add_forward(input, &mut output);
        }
        Ok(output)
    }

    fn project_backward(
        &mut self,
        dense: &Dense,
        name: &str,
        input: &Matrix,
        grad: &Matrix,
    ) -> Matrix {
        let mut grad_input = dense_backward(dense, grad);
        if let Some(lora) = self.adapters.get_mut(name) {
            lora.backward(input, grad, &mut grad_input);
        }
        grad_input
    }
}

/// The latent scaled for its noise level and the embedded time, which is what
/// [`Unet::forward`] works out on the host before its device pass too.
pub(crate) fn prepare(
    unet: &Unet,
    latent: &FeatureMap,
    sigma: f32,
    conditioning: &Conditioning,
) -> Result<(FeatureMap, Vec<f32>), NetworkError> {
    let config = unet.config();
    let mut scaled = latent.clone();
    let scale = (sigma * sigma + 1.0).sqrt().recip();
    let noisy = config.out_channels.min(latent.channels) * latent.pixels();
    scaled.data[..noisy]
        .iter_mut()
        .for_each(|value| *value *= scale);
    let time = unet
        .encoder
        .time(config, unet.timestep(sigma), latent, &conditioning.pooled)?;
    Ok((scaled, time))
}

/// What [`UnetLora::backward`] reads of a forward pass.
pub struct Tape {
    down: Vec<(Vec<Rung>, Option<FeatureMap>)>,
    middle: (ResnetTape, TransformerTape, ResnetTape),
    up: Vec<(Vec<Rung>, Option<FeatureMap>)>,
    norm_input: FeatureMap,
    normed: FeatureMap,
}

/// A resnet and the transformer after it, if there is one; on the way up, the
/// channels that came from below rather than from the skip.
type Rung = (usize, ResnetTape, Option<TransformerTape>);

struct ResnetTape {
    input: FeatureMap,
    normed1: FeatureMap,
    hidden: FeatureMap,
    normed2: FeatureMap,
}

struct TransformerTape {
    input: FeatureMap,
    normed: FeatureMap,
    blocks: Vec<BlockTape>,
    tokens: Matrix,
}

struct BlockTape {
    before: [Matrix; 3],
    own: AttentionTape,
    cross: AttentionTape,
    projected: Matrix,
}

struct AttentionTape {
    input: Matrix,
    source: Matrix,
    queries: Matrix,
    keys: Matrix,
    values: Matrix,
    attended: Matrix,
}

fn resnet_forward(
    resnet: &Resnet,
    input: &FeatureMap,
    time: &[f32],
) -> Result<(FeatureMap, ResnetTape), NetworkError> {
    let mut normed1 = input.clone();
    resnet.norm1.forward(&mut normed1)?;
    let mut activated = normed1.clone();
    silu(&mut activated);
    let mut hidden = resnet.conv1.forward(&activated)?;
    let time: Vec<f32> = time.iter().map(|value| crate::ffn::silu(*value)).collect();
    let per_channel = resnet.time.apply(&time)?;
    let pixels = hidden.pixels();
    for (plane, offset) in hidden.data.chunks_mut(pixels).zip(&per_channel) {
        plane.iter_mut().for_each(|value| *value += offset);
    }
    let mut normed2 = hidden.clone();
    resnet.norm2.forward(&mut normed2)?;
    let mut activated = normed2.clone();
    silu(&mut activated);
    let mut output = match &resnet.shortcut {
        Some(shortcut) => shortcut.forward(input)?,
        None => input.clone(),
    };
    add_maps(&mut output, &resnet.conv2.forward(&activated)?);
    let tape = ResnetTape {
        input: input.clone(),
        normed1,
        hidden,
        normed2,
    };
    Ok((output, tape))
}

fn resnet_backward(resnet: &Resnet, tape: &ResnetTape, grad: &FeatureMap) -> FeatureMap {
    let grad_hidden = conv_backward(&resnet.conv2, grad, &tape.normed2);
    let grad_hidden = silu_backward(&grad_hidden, &tape.normed2);
    // The noise level's offset adds to `hidden` and passes the gradient on.
    let grad_hidden = group_norm_backward(&resnet.norm2, &tape.hidden, &grad_hidden);
    let grad_hidden = conv_backward(&resnet.conv1, &grad_hidden, &tape.normed1);
    let grad_hidden = silu_backward(&grad_hidden, &tape.normed1);
    let mut output = group_norm_backward(&resnet.norm1, &tape.input, &grad_hidden);
    match &resnet.shortcut {
        Some(shortcut) => add_maps(&mut output, &conv_backward(shortcut, grad, &tape.input)),
        None => add_maps(&mut output, grad),
    }
    output
}

/// `dL/d(input)` of `conv.forward(input)`; only `input`'s shape is read.
pub(crate) fn conv_backward(conv: &Conv2d, grad: &FeatureMap, input: &FeatureMap) -> FeatureMap {
    let (kernel, stride, padding) = (conv.kernel, conv.stride, conv.padding);
    let patch = conv.in_channels * kernel * kernel;
    let mut columns = Matrix::new(grad.pixels(), patch);
    grad.to_tokens().dot(&conv.weight, &mut columns);
    // The reverse of `im2col`: every column entry goes back to the input
    // value it was read from.
    let mut output = FeatureMap::new(input.channels, input.height, input.width);
    for pixel in 0..grad.pixels() {
        let (out_y, out_x) = (pixel / grad.width, pixel % grad.width);
        let top = (out_y * stride) as isize - padding as isize;
        let left = (out_x * stride) as isize - padding as isize;
        let row = columns.row(pixel);
        for channel in 0..conv.in_channels {
            for dy in 0..kernel {
                let y = top + dy as isize;
                if y < 0 || y >= input.height as isize {
                    continue;
                }
                for dx in 0..kernel {
                    let x = left + dx as isize;
                    if x >= 0 && x < input.width as isize {
                        let target =
                            (channel * input.height + y as usize) * input.width + x as usize;
                        output.data[target] += row[(channel * kernel + dy) * kernel + dx];
                    }
                }
            }
        }
    }
    output
}

/// `dL/dx` of [`GroupNorm::forward`], given `x`.
pub(crate) fn group_norm_backward(
    norm: &GroupNorm,
    input: &FeatureMap,
    grad: &FeatureMap,
) -> FeatureMap {
    let pixels = input.pixels();
    let per_group = input.channels / norm.groups;
    let span = per_group * pixels;
    let mut output = FeatureMap::new(input.channels, input.height, input.width);
    for (group, ((values, grads), target)) in input
        .data
        .chunks(span)
        .zip(grad.data.chunks(span))
        .zip(output.data.chunks_mut(span))
        .enumerate()
    {
        let mean = values.iter().sum::<f32>() / span as f32;
        let variance = values
            .iter()
            .map(|value| (value - mean).powi(2))
            .sum::<f32>()
            / span as f32;
        let inverse = (variance + norm.eps).sqrt().recip();
        let scaled: Vec<f32> = grads
            .iter()
            .enumerate()
            .map(|(index, grad)| grad * norm.weight[group * per_group + index / pixels])
            .collect();
        let grad_mean = scaled.iter().sum::<f32>() / span as f32;
        let projection = values
            .iter()
            .zip(&scaled)
            .map(|(value, grad)| (value - mean) * inverse * grad)
            .sum::<f32>()
            / span as f32;
        for ((target, value), grad) in target.iter_mut().zip(values).zip(&scaled) {
            *target = inverse * (grad - grad_mean - (value - mean) * inverse * projection);
        }
    }
    output
}

/// `dL/dx` of the affine layer norm [`crate::clip::Norm`].
pub(crate) fn norm_backward(
    norm: &crate::clip::Norm,
    input: &Matrix,
    eps: f32,
    grad: &Matrix,
) -> Matrix {
    let mut scaled = grad.clone();
    scaled.data.chunks_mut(grad.cols).for_each(|row| {
        row.iter_mut()
            .zip(&norm.weight)
            .for_each(|(value, weight)| *value *= weight)
    });
    layer_norm_backward(input, eps, &scaled)
}

/// Half the projection is the value and half its gate.
fn geglu(projected: &Matrix) -> Matrix {
    let inner = projected.cols / 2;
    let mut hidden = Matrix::new(projected.rows, inner);
    for (target, row) in hidden
        .data
        .chunks_mut(inner)
        .zip(projected.data.chunks(projected.cols))
    {
        for (index, value) in target.iter_mut().enumerate() {
            *value = row[index] * gelu(row[inner + index]);
        }
    }
    hidden
}

pub(crate) fn geglu_backward(projected: &Matrix, grad: &Matrix) -> Matrix {
    let inner = grad.cols;
    let mut output = Matrix::new(projected.rows, projected.cols);
    for ((target, row), grad) in output
        .data
        .chunks_mut(projected.cols)
        .zip(projected.data.chunks(projected.cols))
        .zip(grad.data.chunks(inner))
    {
        for index in 0..inner {
            let (value, gate) = (row[index], row[inner + index]);
            target[index] = grad[index] * gelu(gate);
            target[inner + index] = grad[index] * value * gelu_derivative(gate);
        }
    }
    output
}

/// `dL/d(input)` of a frozen [`Dense`].
fn dense_backward(dense: &Dense, grad: &Matrix) -> Matrix {
    let weight = dense.weight_f32();
    let mut output = Matrix::new(grad.rows, weight.cols);
    grad.dot(&weight, &mut output);
    output
}

pub(crate) fn silu_backward(grad: &FeatureMap, pre: &FeatureMap) -> FeatureMap {
    let mut output = grad.clone();
    output
        .data
        .iter_mut()
        .zip(&pre.data)
        .for_each(|(grad, pre)| *grad *= silu_derivative(*pre));
    output
}

/// Each output pixel's gradient goes back to the one it was copied from.
pub(crate) fn upsample_backward(grad: &FeatureMap) -> FeatureMap {
    let (height, width) = (grad.height / 2, grad.width / 2);
    let mut output = FeatureMap::new(grad.channels, height, width);
    for channel in 0..grad.channels {
        for y in 0..grad.height {
            for x in 0..grad.width {
                output.data[(channel * height + y / 2) * width + x / 2] +=
                    grad.data[(channel * grad.height + y) * grad.width + x];
            }
        }
    }
    output
}

fn add(target: &mut Matrix, source: &Matrix) {
    target
        .data
        .iter_mut()
        .zip(&source.data)
        .for_each(|(target, source)| *target += source);
}

fn add_maps(target: &mut FeatureMap, source: &FeatureMap) {
    target
        .data
        .iter_mut()
        .zip(&source.data)
        .for_each(|(target, source)| *target += source);
}

#[cfg(test)]
pub(crate) mod tests {
    use super::*;
    use crate::unet::tests::{conditioning, latent, random, tiny};
    use rand::{Rng, SeedableRng};

    pub(crate) fn adapters(unet: &Unet) -> UnetLora {
        let mut rng = rand::rngs::StdRng::seed_from_u64(0);
        let mut random = |rows, cols| {
            let data = (0..rows * cols).map(|_| rng.gen_range(-0.3..0.3)).collect();
            Param::new(Matrix::from_vec(rows, cols, data))
        };
        let adapters = projections(unet)
            .into_iter()
            .map(|(name, dense)| {
                let lora = Lora {
                    down: random(2, dense.in_dim()),
                    up: random(dense.out_dim(), 2),
                    scale: 0.5,
                };
                (name, lora)
            })
            .collect();
        UnetLora { adapters }
    }

    /// Needs `RB_UNET_LORA_TRAINED` from `scripts/unet_lora_trained_check.py
    /// check`, run on what `train_unet_lora` wrote.
    #[test]
    #[ignore]
    fn a_trained_unet_lora_loads_as_diffusers_loads_it() {
        let dir = std::path::PathBuf::from(std::env::var("RB_UNET_LORA_TRAINED").unwrap());
        let reference: serde_json::Value =
            serde_json::from_str(&std::fs::read_to_string(dir.join("reference.json")).unwrap())
                .unwrap();
        let floats = |value: &serde_json::Value| -> Vec<f32> {
            let values = value.as_array().unwrap().iter();
            values.map(|x| x.as_f64().unwrap() as f32).collect()
        };
        let model = std::path::PathBuf::from(reference["model"].as_str().unwrap());
        let config = crate::unet::UnetConfig::from_file(model.join("unet/config.json")).unwrap();
        let unet = Unet::load_at(
            model.join("unet/diffusion_pytorch_model.safetensors"),
            config.clone(),
            crate::transformer::Precision::F32,
        )
        .unwrap();
        let latent = FeatureMap::from_vec(4, 8, 8, floats(&reference["latent"])).unwrap();
        let conditioning = Conditioning {
            text: Matrix::from_vec(5, config.cross_dim, floats(&reference["text"])),
            pooled: vec![],
            height: 8,
            width: 8,
            guidance: 0.0,
        };
        let sigma = reference["sigma"].as_f64().unwrap() as f32;
        let worst = |got: &[f32], want: &[f32]| {
            let largest = want.iter().fold(0.0f32, |max, value| max.max(value.abs()));
            let error = got
                .iter()
                .zip(want)
                .fold(0.0f32, |max, (got, want)| max.max((got - want).abs()));
            error / largest
        };
        let (plain, adapted) = (floats(&reference["plain"]), floats(&reference["adapted"]));
        // Training has to have moved the model, or matching says nothing.
        let moved = worst(&adapted, &plain);
        assert!(
            moved > 1e-3,
            "the LoRA moves diffusers' prediction by {moved}"
        );
        assert_eq!(
            reference["adapters"].as_u64().unwrap() as usize,
            projections(&unet).len()
        );

        let file = reference["lora"].as_str().unwrap();
        let lora = UnetLora::read_peft(file, &unet, 1.0).unwrap();
        let (ours, _) = lora.forward(&unet, &latent, sigma, &conditioning).unwrap();
        let error = worst(&ours.data, &adapted);
        eprintln!("read_peft is off diffusers by {error}; the LoRA moves it by {moved}");
        assert!(
            error < 1e-4,
            "read_peft is off diffusers by {error}, moved {moved}"
        );
    }

    #[test]
    fn written_adapters_read_back_to_the_same_prediction() {
        let config = tiny();
        let unet = random(&config, "rusting_brain_unet_lora_write.safetensors");
        let fresh = UnetLora::new(&unet, 2, 4.0, &mut rand::rngs::StdRng::seed_from_u64(0));
        assert_eq!(fresh.adapters.len(), projections(&unet).len());
        assert!(fresh.adapters.values().all(|lora| lora.scale == 2.0));
        let (latent, conditioning) = (latent(&config, 8), conditioning(&config, 8, 1.0));
        let (plain, _) = UnetLora::default()
            .forward(&unet, &latent, 1.5, &conditioning)
            .unwrap();
        let (untrained, _) = fresh.forward(&unet, &latent, 1.5, &conditioning).unwrap();
        assert_eq!(untrained.data, plain.data, "a zero `up` changes nothing");

        let lora = adapters(&unet);
        let path = std::env::temp_dir().join("rusting_brain_unet_lora_written.safetensors");
        lora.write_peft(&path).unwrap();
        let read = UnetLora::read_peft(&path, &unet, 1.0).unwrap();
        std::fs::remove_file(path).ok();
        assert_eq!(read.adapters.len(), lora.adapters.len());
        let (expected, _) = lora.forward(&unet, &latent, 1.5, &conditioning).unwrap();
        let (actual, _) = read.forward(&unet, &latent, 1.5, &conditioning).unwrap();
        let worst = actual
            .data
            .iter()
            .zip(&expected.data)
            .fold(0.0f32, |max, (a, b)| max.max((a - b).abs()));
        assert!(worst < 1e-5, "{worst}");
        assert_ne!(expected.data, plain.data);
    }

    #[test]
    fn adapter_gradients_match_finite_differences() {
        let config = tiny();
        let unet = random(&config, "rusting_brain_unet_lora.safetensors");
        let (latent, conditioning) = (latent(&config, 8), conditioning(&config, 8, 1.0));
        let sigma = 1.5;

        // Without adapters the pass is the model's own.
        let (plain, _) = UnetLora::default()
            .forward(&unet, &latent, sigma, &conditioning)
            .unwrap();
        let own = unet.forward(&latent, sigma, &conditioning).unwrap();
        let error = plain
            .data
            .iter()
            .zip(&own.data)
            .fold(0.0f32, |max, (a, b)| max.max((a - b).abs()));
        assert!(error < 1e-5, "the taped pass is off by {error}");

        let mut lora = adapters(&unet);
        let (output, tape) = lora.forward(&unet, &latent, sigma, &conditioning).unwrap();
        let mut weights = output.clone();
        for (index, value) in weights.data.iter_mut().enumerate() {
            *value = ((index * 7 % 11) as f32 - 5.0) * 0.1;
        }
        lora.backward(&unet, &tape, &weights).unwrap();
        let loss = |lora: &UnetLora| -> f64 {
            let (output, _) = lora.forward(&unet, &latent, sigma, &conditioning).unwrap();
            let pairs = output.data.iter().zip(&weights.data);
            pairs.map(|(a, b)| (a * b) as f64).sum()
        };

        // Each adapter on its own: the loss along its own gradient changes by
        // the gradient's length.
        let names: Vec<String> = lora.adapters.keys().cloned().collect();
        assert_eq!(names.len(), 88);
        let step = 0.5;
        for name in names {
            for factor in 0..2 {
                fn pick<'a>(lora: &'a mut UnetLora, name: &str, factor: usize) -> &'a mut Param {
                    let adapter = lora.adapters.get_mut(name).unwrap();
                    if factor == 0 {
                        &mut adapter.down
                    } else {
                        &mut adapter.up
                    }
                }
                let grad = pick(&mut lora, &name, factor).grad.data.clone();
                let length = grad.iter().map(|g| g * g).sum::<f32>().sqrt();
                let mut moved = [lora.clone(), lora.clone()];
                for (copy, sign) in moved.iter_mut().zip([1.0, -1.0]) {
                    let param = pick(copy, &name, factor);
                    for (value, grad) in param.value.data.iter_mut().zip(&grad) {
                        *value += sign * step * grad / length;
                    }
                }
                let slope = (loss(&moved[0]) - loss(&moved[1])) / (2.0 * step as f64);
                // f32 rounding in the loss leaves about 5e-7 of noise in the slope.
                assert!(
                    (slope as f32 - length).abs() < 1e-2 * length + 1e-6,
                    "{name} factor {factor}: slope {slope}, gradient length {length}"
                );
                assert!(
                    error < 2e-2,
                    "{name} factor {factor}: slope {slope}, gradient length {length}"
                );
            }
        }
    }

    /// Needs `RB_UNET_LORA_REF` from `scripts/unet_lora_gradients.py`.
    #[test]
    #[ignore]
    fn unet_lora_gradients_match_pytorch() {
        let dir = std::path::PathBuf::from(std::env::var("RB_UNET_LORA_REF").unwrap());
        let reference: serde_json::Value =
            serde_json::from_str(&std::fs::read_to_string(dir.join("reference.json")).unwrap())
                .unwrap();
        let floats = |value: &serde_json::Value| -> Vec<f32> {
            let values = value.as_array().unwrap().iter();
            values.map(|x| x.as_f64().unwrap() as f32).collect()
        };
        let config = crate::unet::UnetConfig::from_file(dir.join("unet/config.json")).unwrap();
        let unet = Unet::load_at(
            dir.join("unet/diffusion_pytorch_model.safetensors"),
            config.clone(),
            crate::transformer::Precision::F32,
        )
        .unwrap();
        let scale = reference["scale"].as_f64().unwrap() as f32;
        let mut lora = UnetLora::read_peft(dir.join("adapters.safetensors"), &unet, scale).unwrap();
        assert_eq!(lora.adapters.len(), 48);
        let latent = FeatureMap::from_vec(4, 8, 8, floats(&reference["latent"])).unwrap();
        let conditioning = Conditioning {
            text: Matrix::from_vec(5, config.cross_dim, floats(&reference["text"])),
            pooled: vec![],
            height: 8,
            width: 8,
            guidance: 0.0,
        };
        let sigma = reference["sigma"].as_f64().unwrap() as f32;
        let (output, tape) = lora.forward(&unet, &latent, sigma, &conditioning).unwrap();
        let worst = |got: &[f32], want: &[f32]| {
            assert_eq!(got.len(), want.len());
            let largest = want.iter().fold(0.0f32, |max, value| max.max(value.abs()));
            let error = got
                .iter()
                .zip(want)
                .fold(0.0f32, |max, (got, want)| max.max((got - want).abs()));
            error / largest
        };
        let error = worst(&output.data, &floats(&reference["output"]));
        assert!(error < 1e-4, "the prediction is off by {error} relative");

        let grad = FeatureMap::from_vec(4, 8, 8, floats(&reference["weights"])).unwrap();
        lora.backward(&unet, &tape, &grad).unwrap();
        let mut largest = 0.0f32;
        for (name, want) in reference["gradients"].as_object().unwrap() {
            let (module, kind) = name.split_once(".lora_").unwrap();
            let adapter = &lora.adapters[module];
            let got = match kind.starts_with('A') {
                true => &adapter.down.grad.data,
                false => &adapter.up.grad.data,
            };
            let error = worst(got, &floats(want));
            assert!(error < 1e-4, "{name} is off by {error} relative");
            largest = largest.max(error);
        }
        println!("prediction off by {error}, worst gradient off by {largest}");
    }
}