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
//! Adapters on the device UNet's attention projections, and what a forward
//! pass keeps of itself for the backward one.
//!
//! This is the device side of [`crate::unet_lora`]: the same adapters, in the
//! same places, matched to that host reference. The forward pass is
//! [`DeviceUnet`]'s own, with a [`Recorder`] threaded through it that adds
//! each adapter after its projection and keeps what the backward pass reads.

use super::*;
use crate::unet_lora::{PROJECTIONS, UnetLora, transformers};
use cudarc::driver::DevicePtrMut;

/// One adapter: `scale * up . (down . x)`, both factors FP16, and both
/// factors' gradients in FP32.
pub(crate) struct DeviceLora {
    name: String,
    down: Tensor,
    up: Tensor,
    scale: f32,
    grads: [CudaSlice<f32>; 2],
}

/// A [`UnetLora`] on the device. One entry a transformer, in the order a
/// pass meets them; in each, one entry a block, holding `attn1` and `attn2`,
/// each holding `to_q`, `to_k`, `to_v` and `to_out.0`.
pub struct DeviceAdapters {
    transformers: Vec<Vec<[[Option<DeviceLora>; 4]; 2]>>,
}

impl std::fmt::Debug for DeviceAdapters {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("DeviceAdapters").finish_non_exhaustive()
    }
}

impl DeviceAdapters {
    pub fn upload(gpu: &ImageGpu, unet: &Unet, lora: &UnetLora) -> Result<Self, NetworkError> {
        let mut used = 0;
        let mut table = Vec::new();
        for (name, transformer) in transformers(unet) {
            let mut blocks = Vec::new();
            for index in 0..transformer.blocks.len() {
                let mut slots: [[Option<DeviceLora>; 4]; 2] = Default::default();
                for (attention, slot) in ["attn1", "attn2"].iter().zip(&mut slots) {
                    for (projection, target) in PROJECTIONS.iter().zip(slot.iter_mut()) {
                        let key =
                            format!("{name}.transformer_blocks.{index}.{attention}.{projection}");
                        if let Some(adapter) = lora.adapters.get(&key) {
                            used += 1;
                            let zeros = |len| {
                                let stream = &gpu.context.stream;
                                stream
                                    .alloc_zeros::<f32>(len)
                                    .map_err(cuda_err("device allocation"))
                            };
                            let (down, up) = (&adapter.down.value, &adapter.up.value);
                            *target = Some(DeviceLora {
                                grads: [zeros(down.data.len())?, zeros(up.data.len())?],
                                down: gpu.upload_matrix(down)?,
                                up: gpu.upload_matrix(up)?,
                                scale: adapter.scale,
                                name: key,
                            });
                        }
                    }
                }
                blocks.push(slots);
            }
            table.push(blocks);
        }
        if used != lora.adapters.len() {
            return Err(NetworkError::InvalidConfig(format!(
                "{} of {} adapters sit on no attention projection of this UNet",
                lora.adapters.len() - used,
                lora.adapters.len()
            )));
        }
        Ok(Self {
            transformers: table,
        })
    }

    /// Adds the gradients gathered on the device to `lora`'s, which is the
    /// file these adapters were uploaded from.
    pub fn add_gradients(&self, gpu: &ImageGpu, lora: &mut UnetLora) -> Result<(), NetworkError> {
        let all = self.transformers.iter().flatten().flatten().flatten();
        for adapter in all.flatten() {
            let target = lora.adapters.get_mut(&adapter.name).ok_or_else(|| {
                NetworkError::InvalidConfig(format!("no adapter {} in this file", adapter.name))
            })?;
            for (param, grad) in [&mut target.down, &mut target.up]
                .into_iter()
                .zip(&adapter.grads)
            {
                let grad = gpu
                    .context
                    .stream
                    .clone_dtoh(grad)
                    .map_err(cuda_err("device to host copy"))?;
                param
                    .grad
                    .data
                    .iter_mut()
                    .zip(grad)
                    .for_each(|(sum, value)| *sum += value);
            }
        }
        Ok(())
    }
}

/// Threaded through a forward pass: the adapters, which transformer is next,
/// and what has been kept so far.
pub(crate) struct Recorder<'a> {
    adapters: &'a DeviceAdapters,
    next: usize,
    pub(super) records: Records,
}

impl<'a> Recorder<'a> {
    /// The next transformer's adapters.
    pub(super) fn next(&mut self) -> Result<&'a [[[Option<DeviceLora>; 4]; 2]], NetworkError> {
        let adapters: &'a DeviceAdapters = self.adapters;
        let blocks = adapters.transformers.get(self.next).ok_or_else(|| {
            NetworkError::InvalidConfig("the adapters were uploaded for another UNet".into())
        })?;
        self.next += 1;
        Ok(blocks)
    }
}

/// Each list in the order the forward pass made it, which the backward pass
/// reads from the end.
#[derive(Default)]
pub(super) struct Records {
    /// A resnet's input and its hidden map before the second norm.
    pub(super) resnets: Vec<(Map, Map)>,
    /// A transformer's input and its blocks.
    pub(super) transformers: Vec<(Map, Vec<BlockRecord>)>,
    /// The size of every resampling convolution's input.
    pub(super) resampled: Vec<(usize, usize)>,
    /// On the way up, the channels that came from below rather than the skip.
    pub(super) joined: Vec<usize>,
}

/// The tokens before each of the three norms, the attentions, and the gated
/// feed-forward's projection before its bias.
pub(crate) struct BlockRecord {
    pub(super) before: [Tensor; 3],
    pub(super) own: AttentionRecord,
    pub(super) cross: AttentionRecord,
    pub(super) projected: Tensor,
}

/// Self-attention keeps its stacked `[q | k | v]`; cross-attention its queries
/// and its stacked `[k | v]` of the prompt.
pub(crate) struct AttentionRecord {
    pub(super) queries: Option<Tensor>,
    pub(super) projected: Tensor,
    pub(super) attended: Tensor,
}

/// What [`DeviceUnet::forward_adapted`] keeps for the backward pass.
pub struct DeviceTape {
    records: Records,
    /// The output norm's input.
    out: Map,
    text: Tensor,
}

impl std::fmt::Debug for DeviceTape {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("DeviceTape").finish_non_exhaustive()
    }
}

impl ImageGpu {
    /// `out += scale * (input . down^T) . up^T`, `out` having rows `stride`
    /// apart.
    pub(super) fn lora_add<O: DevicePtrMut<f16>>(
        &self,
        lora: &DeviceLora,
        input: &Tensor,
        out: &mut O,
        stride: usize,
    ) -> Result<(), NetworkError> {
        let rank = lora.down.rows;
        let mut hidden = self.uninit(input.rows, rank)?;
        act_rhs_transposed::<f16, _>(
            &self.context,
            &input.bytes(),
            input.cols,
            &lora.down.bytes(),
            lora.down.cols,
            true,
            true,
            &mut hidden.data,
            rank,
            input.rows,
            rank,
            input.cols,
            1.0,
            0.0,
        )?;
        act_rhs_transposed::<f16, _>(
            &self.context,
            &hidden.bytes(),
            rank,
            &lora.up.bytes(),
            rank,
            true,
            true,
            out,
            stride,
            input.rows,
            lora.up.rows,
            rank,
            lora.scale,
            1.0,
        )
    }
}

impl ImageGpu {
    /// Adds both factors' gradients for `grad`, the gradient of the adapter's
    /// output, and with `grad_input` the adapter's share of `dL/d(input)`.
    fn lora_backward(
        &self,
        lora: &mut DeviceLora,
        input: &Tensor,
        grad: &Tensor,
        grad_input: Option<&mut Tensor>,
    ) -> Result<(), NetworkError> {
        let (rows, rank, inputs) = (input.rows, lora.down.rows, input.cols);
        let mut hidden = self.uninit(rows, rank)?;
        act_rhs_transposed::<f16, _>(
            &self.context,
            &input.bytes(),
            inputs,
            &lora.down.bytes(),
            inputs,
            true,
            true,
            &mut hidden.data,
            rank,
            rows,
            rank,
            inputs,
            lora.scale,
            0.0,
        )?;
        let [grad_down, grad_up] = &mut lora.grads;
        // `dL/d(up) += grad^T . (scale * down . x)`.
        self.sum_lhs_transposed(
            (&grad.data.slice(..), grad.cols, 0),
            (&hidden.data.slice(..), rank, 0),
            (grad_up, rank, 0),
            (rows, grad.cols, rank, 1),
        )?;
        let mut grad_hidden = self.uninit(rows, rank)?;
        act_plain::<f16, _>(
            &self.context,
            &grad.bytes(),
            grad.cols,
            &lora.up.bytes(),
            rank,
            true,
            true,
            &mut grad_hidden.data,
            rank,
            rows,
            rank,
            grad.cols,
            lora.scale,
            0.0,
        )?;
        self.sum_lhs_transposed(
            (&grad_hidden.data.slice(..), rank, 0),
            (&input.data.slice(..), inputs, 0),
            (grad_down, inputs, 0),
            (rows, rank, inputs, 1),
        )?;
        if let Some(grad_input) = grad_input {
            act_plain::<f16, _>(
                &self.context,
                &grad_hidden.bytes(),
                rank,
                &lora.down.bytes(),
                inputs,
                true,
                true,
                &mut grad_input.data,
                inputs,
                rows,
                inputs,
                rank,
                1.0,
                1.0,
            )?;
        }
        Ok(())
    }

    /// `dL/d(input)` of a projection by `weight`, `[outputs, inputs]`:
    /// `grad . weight`, added to `out` with `beta` one.
    fn dense_backward_into(
        &self,
        weight: &CudaView<'_, u8>,
        inputs: usize,
        grad: &Tensor,
        out: &mut Tensor,
        beta: f32,
    ) -> Result<(), NetworkError> {
        act_plain::<f16, _>(
            &self.context,
            &grad.bytes(),
            grad.cols,
            weight,
            inputs,
            true,
            true,
            &mut out.data,
            inputs,
            grad.rows,
            inputs,
            grad.cols,
            1.0,
            beta,
        )
    }

    fn dense_backward(&self, weight: &Tensor, grad: &Tensor) -> Result<Tensor, NetworkError> {
        let mut out = self.uninit(grad.rows, weight.cols)?;
        self.dense_backward_into(&weight.bytes(), weight.cols, grad, &mut out, 0.0)?;
        Ok(out)
    }

    /// `joined`'s first `channels` channels, and the rest.
    fn split(&self, joined: &Map, channels: usize) -> Result<(Map, Map), NetworkError> {
        let (pixels, split) = (joined.pixels(), channels * joined.pixels());
        let part = |range: std::ops::Range<usize>| {
            let mut part = self.uninit(range.len() / pixels, pixels)?;
            self.context
                .stream
                .memcpy_dtod(&joined.tensor.data.slice(range), &mut part.data)
                .map_err(cuda_err("device to device copy"))?;
            Ok::<_, NetworkError>(Map {
                tensor: part,
                height: joined.height,
                width: joined.width,
            })
        };
        Ok((part(0..split)?, part(split..joined.tensor.len())?))
    }
}

fn pop<T>(list: &mut Vec<T>) -> Result<T, NetworkError> {
    list.pop()
        .ok_or_else(|| NetworkError::InvalidConfig("the tape is not from this UNet".into()))
}

impl DeviceUnet {
    /// [`DeviceUnet::forward`] with `adapters` on the attention projections,
    /// keeping what the backward pass reads. `latent` and `time` are what
    /// `unet_lora::prepare` makes of a latent and a noise level.
    pub fn forward_adapted(
        &self,
        adapters: &DeviceAdapters,
        latent: &FeatureMap,
        time: &[f32],
        text: &Matrix,
    ) -> Result<(FeatureMap, DeviceTape), NetworkError> {
        let gpu = &self.gpu;
        let time = self.time(time)?;
        let text = gpu.upload_matrix(text)?;
        let mut recorder = Recorder {
            adapters,
            next: 0,
            records: Records::default(),
        };
        let (sample, skips) = self.down(latent, &time, &text, None, Some(&mut recorder))?;
        let mut sample = self.up(sample, skips, &time, &text, Some(&mut recorder))?;
        if recorder.next != adapters.transformers.len() {
            return Err(NetworkError::InvalidConfig(
                "the adapters were uploaded for another UNet".into(),
            ));
        }
        let out = clone_map(gpu, &sample)?;
        let (norm_out, conv_out) = self.out.as_ref().ok_or_else(|| {
            NetworkError::InvalidConfig("a ControlNet has no output to train".into())
        })?;
        gpu.group_norm(norm_out, &mut sample, 0)?;
        let output = gpu.conv(conv_out, &sample)?;
        let tape = DeviceTape {
            records: recorder.records,
            out,
            text,
        };
        Ok((download_map(gpu, &output)?, tape))
    }

    /// One noise-prediction training step on `lora`, as diffusers'
    /// `train_text_to_image_lora` takes it: `noise` is added to `latent` at
    /// level `sigma`, the model predicts it, and the optimizer steps on the
    /// adapters with the gradient clipped to a norm of 1. Returns the mean
    /// squared error before the step.
    ///
    /// NOTE: the adapters go up and their gradients come down every
    /// step, a few MB at rank 16 on SD 1.5; keep them on the device with a
    /// device optimizer once that shows in a profile.
    #[allow(clippy::too_many_arguments)]
    pub fn noise_step(
        &self,
        unet: &Unet,
        lora: &mut UnetLora,
        latent: &FeatureMap,
        noise: &FeatureMap,
        sigma: f32,
        conditioning: &crate::mmdit::Conditioning,
        optimizer: &crate::optimizers::Optimizer,
        step: usize,
    ) -> Result<f32, NetworkError> {
        let mut noisy = latent.clone();
        noisy
            .data
            .iter_mut()
            .zip(&noise.data)
            .for_each(|(x, n)| *x += sigma * n);
        let (scaled, time) = crate::unet_lora::prepare(unet, &noisy, sigma, conditioning)?;
        let mut adapters = DeviceAdapters::upload(&self.gpu, unet, lora)?;
        let (prediction, tape) =
            self.forward_adapted(&adapters, &scaled, &time, &conditioning.text)?;
        let count = prediction.data.len() as f32;
        let mut grad = prediction;
        let mut loss = 0.0;
        for (p, n) in grad.data.iter_mut().zip(&noise.data) {
            let error = *p - n;
            loss += error * error;
            *p = 2.0 * error / count;
        }
        self.backward_adapted(&mut adapters, tape, &grad)?;
        adapters.add_gradients(&self.gpu, lora)?;
        crate::optimizers::step_clipped(&mut lora.params_mut(), optimizer, step, 1.0, 1.0)?;
        crate::optimizers::zero_grad(&mut lora.params_mut());
        Ok(loss / count)
    }

    /// Adds `dL/d(adapter)` to the gradients `adapters` keep on the device,
    /// given `dL/d(prediction)` of the pass that made `tape`. The way back
    /// reads the tape from its end: the output, the way up handing one
    /// gradient back to each skip, the middle, and the way down.
    pub fn backward_adapted(
        &self,
        adapters: &mut DeviceAdapters,
        tape: DeviceTape,
        grad: &FeatureMap,
    ) -> Result<(), NetworkError> {
        let gpu = &*self.gpu;
        let DeviceTape {
            mut records,
            out,
            text,
        } = tape;
        if records.transformers.len() != adapters.transformers.len() {
            return Err(NetworkError::InvalidConfig(
                "the tape is not from these adapters".into(),
            ));
        }
        let mut slots = adapters.transformers.iter_mut().rev();
        let mut transformer = |model: &DeviceTransformer, records: &mut Records, grad: Map| {
            let slot = slots
                .next()
                .expect("as many adapters as transformer records");
            self.transformer_backward(model, pop(&mut records.transformers)?, slot, &text, grad)
        };
        let (norm_out, conv_out) = self.out.as_ref().ok_or_else(|| {
            NetworkError::InvalidConfig("a ControlNet has no output to train".into())
        })?;
        let mut normed = clone_map(gpu, &out)?;
        gpu.group_norm(norm_out, &mut normed, -1)?;
        let (height, width) = (out.height, out.width);
        let mut grad = gpu.conv_backward(
            conv_out,
            &upload_map(gpu, grad)?,
            height,
            width,
            COLUMN_BUDGET,
        )?;
        gpu.silu_backward(&mut grad.tensor, &normed.tensor)?;
        gpu.group_norm_backward(norm_out, &out, &mut grad)?;
        drop((normed, out));

        let mut skips = Vec::new();
        for block in self.up.iter().rev() {
            if let Some(resampler) = &block.resampler {
                let (height, width) = pop(&mut records.resampled)?;
                let small = gpu.conv_backward(resampler, &grad, height, width, COLUMN_BUDGET)?;
                grad = gpu.upsample_backward(&small, 2)?;
            }
            for (index, resnet) in block.resnets.iter().enumerate().rev() {
                if let Some(model) = block.attentions.get(index) {
                    grad = transformer(model, &mut records, grad)?;
                }
                let joined = self.resnet_backward(resnet, pop(&mut records.resnets)?, &grad)?;
                let (below, skip) = gpu.split(&joined, pop(&mut records.joined)?)?;
                skips.push(skip);
                grad = below;
            }
        }

        let (first, middle, second) = &self.middle;
        grad = self.resnet_backward(second, pop(&mut records.resnets)?, &grad)?;
        grad = transformer(middle, &mut records, grad)?;
        grad = self.resnet_backward(first, pop(&mut records.resnets)?, &grad)?;

        for block in self.down.iter().rev() {
            if let Some(resampler) = &block.resampler {
                gpu.add(&mut grad.tensor, &pop(&mut skips)?.tensor)?;
                let (height, width) = pop(&mut records.resampled)?;
                grad = gpu.conv_backward(resampler, &grad, height, width, COLUMN_BUDGET)?;
            }
            for (index, resnet) in block.resnets.iter().enumerate().rev() {
                gpu.add(&mut grad.tensor, &pop(&mut skips)?.tensor)?;
                if let Some(model) = block.attentions.get(index) {
                    grad = transformer(model, &mut records, grad)?;
                }
                grad = self.resnet_backward(resnet, pop(&mut records.resnets)?, &grad)?;
            }
        }
        // What is left is the gradient of `conv_in`'s output and the first
        // skip's, which reach only frozen weights and the latent.
        Ok(())
    }

    /// `dL/d(input)` of [`DeviceUnet::resnet`], given its input and the
    /// hidden map its second norm read.
    fn resnet_backward(
        &self,
        resnet: &DeviceResnet,
        (input, hidden): (Map, Map),
        grad: &Map,
    ) -> Result<Map, NetworkError> {
        let gpu = &*self.gpu;
        let mut normed = clone_map(gpu, &hidden)?;
        gpu.group_norm(&resnet.norm2, &mut normed, -1)?;
        let (height, width) = (hidden.height, hidden.width);
        let mut inner = gpu.conv_backward(&resnet.conv2, grad, height, width, COLUMN_BUDGET)?;
        gpu.silu_backward(&mut inner.tensor, &normed.tensor)?;
        // The noise level's offset adds to `hidden` and passes the gradient on.
        gpu.group_norm_backward(&resnet.norm2, &hidden, &mut inner)?;
        drop((normed, hidden));

        let mut normed = clone_map(gpu, &input)?;
        gpu.group_norm(&resnet.norm1, &mut normed, -1)?;
        let (height, width) = (input.height, input.width);
        let mut output = gpu.conv_backward(&resnet.conv1, &inner, height, width, COLUMN_BUDGET)?;
        gpu.silu_backward(&mut output.tensor, &normed.tensor)?;
        gpu.group_norm_backward(&resnet.norm1, &input, &mut output)?;
        match &resnet.shortcut {
            Some(shortcut) => {
                let skip = gpu.conv_backward(shortcut, grad, height, width, COLUMN_BUDGET)?;
                gpu.add(&mut output.tensor, &skip.tensor)?;
            }
            None => gpu.add(&mut output.tensor, &grad.tensor)?,
        }
        Ok(output)
    }

    fn transformer_backward(
        &self,
        transformer: &DeviceTransformer,
        (input, blocks): (Map, Vec<BlockRecord>),
        adapters: &mut [[[Option<DeviceLora>; 4]; 2]],
        text: &Tensor,
        grad: Map,
    ) -> Result<Map, NetworkError> {
        let (gpu, eps, heads) = (&*self.gpu, self.eps, transformer.heads);
        let (height, width) = (grad.height, grad.width);
        let mut grad_tokens = match &transformer.output {
            DeviceProjection::Convolution(conv) => {
                gpu.to_tokens(&gpu.conv_backward(conv, &grad, height, width, COLUMN_BUDGET)?)?
            }
            DeviceProjection::Linear(dense) => {
                gpu.dense_backward(&dense.weight, &gpu.to_tokens(&grad)?)?
            }
        };
        for ((block, record), lora) in transformer
            .blocks
            .iter()
            .zip(blocks)
            .zip(adapters.iter_mut())
            .rev()
        {
            let BlockRecord {
                before: [before_self, before_cross, before_feed],
                own,
                cross,
                projected,
            } = record;
            let grad_hidden = gpu.dense_backward(&block.output.weight, &grad_tokens)?;
            let bias = block.gate.bias.as_ref();
            let grad_projected = gpu.gelu_gate_backward(&projected, bias, &grad_hidden)?;
            let mut grad_normed = gpu.dense_backward(&block.gate.weight, &grad_projected)?;
            gpu.layer_norm_backward(&block.norm3, &before_feed, eps, &mut grad_normed)?;
            gpu.add(&mut grad_tokens, &grad_normed)?;

            let normed = gpu.layer_norm(&block.norm2, &before_cross, eps)?;
            let (attention, lora_cross) = (&block.cross, &mut lora[1]);
            let mut grad_normed = self.attend_backward(
                attention,
                cross,
                &normed,
                Some(text),
                heads,
                lora_cross,
                &grad_tokens,
            )?;
            gpu.layer_norm_backward(&block.norm2, &before_cross, eps, &mut grad_normed)?;
            gpu.add(&mut grad_tokens, &grad_normed)?;

            let normed = gpu.layer_norm(&block.norm1, &before_self, eps)?;
            let (attention, lora_own) = (&block.attention, &mut lora[0]);
            let mut grad_normed =
                self.attend_backward(attention, own, &normed, None, heads, lora_own, &grad_tokens)?;
            gpu.layer_norm_backward(&block.norm1, &before_self, eps, &mut grad_normed)?;
            gpu.add(&mut grad_tokens, &grad_normed)?;
        }
        let mut output = match &transformer.input {
            DeviceProjection::Convolution(conv) => {
                let grad = gpu.to_map(&grad_tokens, height, width)?;
                gpu.conv_backward(conv, &grad, height, width, COLUMN_BUDGET)?
            }
            DeviceProjection::Linear(dense) => gpu.to_map(
                &gpu.dense_backward(&dense.weight, &grad_tokens)?,
                height,
                width,
            )?,
        };
        gpu.group_norm_backward(&transformer.norm, &input, &mut output)?;
        gpu.add(&mut output.tensor, &grad.tensor)?;
        Ok(output)
    }

    /// `dL/d(tokens)` of [`DeviceUnet::attend`], given the gradient of what it
    /// added to the residual, adding every adapter's gradients. The prompt is
    /// frozen, so cross-attention hands nothing back to it.
    #[allow(clippy::too_many_arguments)]
    fn attend_backward(
        &self,
        attention: &DeviceAttention,
        record: AttentionRecord,
        tokens: &Tensor,
        context: Option<&Tensor>,
        heads: usize,
        lora: &mut [Option<DeviceLora>; 4],
        grad: &Tensor,
    ) -> Result<Tensor, NetworkError> {
        let gpu = &*self.gpu;
        let units = attention.stacked.units();
        let AttentionRecord {
            queries,
            projected,
            attended,
        } = record;
        let mut grad_attended = gpu.dense_backward(&attention.output.weight, grad)?;
        if let Some(adapter) = &mut lora[3] {
            gpu.lora_backward(adapter, &attended, grad, Some(&mut grad_attended))?;
        }
        let (q, k, v) = match &queries {
            None => (
                projected.part(0, units),
                projected.part(1, units),
                projected.part(2, units),
            ),
            Some(queries) => (
                queries.view(),
                projected.part(0, units),
                projected.part(1, units),
            ),
        };
        let grads = gpu.attention_backward(q, k, v, heads, &grad_attended, SCORE_BUDGET)?;
        drop(grad_attended);
        let stack = &attention.stacked.dense.weight;
        let (stack_bytes, inputs) = (stack.bytes(), stack.cols);
        let rows = |slot: usize| {
            stack_bytes.slice(slot * units * inputs * 2..(slot + 1) * units * inputs * 2)
        };
        let mut grad_tokens = match &attention.query {
            Some(query) => gpu.dense_backward(&query.weight, &grads[0])?,
            None => {
                let mut out = gpu.uninit(tokens.rows, inputs)?;
                gpu.dense_backward_into(&rows(0), inputs, &grads[0], &mut out, 0.0)?;
                out
            }
        };
        if let Some(adapter) = &mut lora[0] {
            gpu.lora_backward(adapter, tokens, &grads[0], Some(&mut grad_tokens))?;
        }
        for slot in 1..3 {
            match context {
                None => {
                    gpu.dense_backward_into(
                        &rows(slot),
                        inputs,
                        &grads[slot],
                        &mut grad_tokens,
                        1.0,
                    )?;
                    if let Some(adapter) = &mut lora[slot] {
                        gpu.lora_backward(adapter, tokens, &grads[slot], Some(&mut grad_tokens))?;
                    }
                }
                Some(context) => {
                    if let Some(adapter) = &mut lora[slot] {
                        gpu.lora_backward(adapter, context, &grads[slot], None)?;
                    }
                }
            }
        }
        Ok(grad_tokens)
    }
}

#[cfg(test)]
mod tests {
    use super::super::tests::{assert_close, device};
    use super::*;
    use crate::unet::tests::{conditioning, latent, random, tiny};
    use crate::unet_lora::{prepare, tests::adapters};
    use rand::SeedableRng;

    #[test]
    fn an_adapted_device_pass_matches_the_host_or_skips_without_a_device() {
        let Some(gpu) = device() else { return };
        let config = tiny();
        let unet = random(&config, "rusting_brain_cuda_unet_lora.safetensors");
        // Ten times the host test's adapters, so their share of the
        // prediction stands well clear of FP16 rounding.
        let mut lora = adapters(&unet);
        lora.adapters
            .values_mut()
            .for_each(|adapter| adapter.scale *= 10.0);
        let (latent, conditioning) = (latent(&config, 8), conditioning(&config, 8, 1.0));
        let sigma = 1.5;
        let (expected, _) = lora.forward(&unet, &latent, sigma, &conditioning).unwrap();
        let (plain, _) = UnetLora::default()
            .forward(&unet, &latent, sigma, &conditioning)
            .unwrap();

        let model = DeviceUnet::upload(&gpu, &unet).unwrap();
        let adapters = DeviceAdapters::upload(&gpu, &unet, &lora).unwrap();
        let (scaled, time) = prepare(&unet, &latent, sigma, &conditioning).unwrap();
        let (actual, tape) = model
            .forward_adapted(&adapters, &scaled, &time, &conditioning.text)
            .unwrap();
        assert_close("adapted UNet", &actual.data, &expected.data, 0.05);

        // A pass that dropped an adapter could still pass the bar above, so
        // the adapters' own share is compared too: device minus unadapted device against host minus unadapted host.
        let bare = DeviceAdapters::upload(&gpu, &unet, &UnetLora::default()).unwrap();
        let (unadapted, _) = model
            .forward_adapted(&bare, &scaled, &time, &conditioning.text)
            .unwrap();
        let delta = |a: &[f32], b: &[f32]| a.iter().zip(b).map(|(a, b)| a - b).collect::<Vec<_>>();
        assert_close(
            "the adapters' share",
            &delta(&actual.data, &unadapted.data),
            &delta(&expected.data, &plain.data),
            0.05,
        );

        let records = &tape.records;
        assert_eq!(records.transformers.len(), adapters.transformers.len());
        let resnets = unet.encoder.down.iter().chain(&unet.up);
        let resnets = resnets.map(|block| block.resnets.len()).sum::<usize>() + 2;
        assert_eq!(records.resnets.len(), resnets);
        assert_eq!(
            records.joined.len(),
            unet.up.iter().map(|b| b.resnets.len()).sum::<usize>()
        );
    }

    #[test]
    fn noise_steps_lower_the_loss_or_skip_without_a_device() {
        let Some(gpu) = device() else { return };
        let config = tiny();
        let unet = random(&config, "rusting_brain_cuda_unet_noise_step.safetensors");
        let mut rng = rand::rngs::StdRng::seed_from_u64(0);
        let mut lora = UnetLora::new(&unet, 4, 4.0, &mut rng);
        let (latent, conditioning) = (latent(&config, 8), conditioning(&config, 8, 1.0));
        let noise = crate::diffusion::noise(1, latent.data.len(), Some(1));
        let noise = FeatureMap::from_vec(latent.channels, 8, 8, noise.data).unwrap();
        let model = DeviceUnet::upload(&gpu, &unet).unwrap();
        let optimizer = crate::optimizers::Optimizer::adam(1e-2);
        // Rank-4 adapters on the one attention level of the tiny UNet fit
        // fixed noise slowly: about 7% in 30 steps, 11% in 200.
        let losses: Vec<f32> = (1..=30)
            .map(|step| {
                model
                    .noise_step(
                        &unet,
                        &mut lora,
                        &latent,
                        &noise,
                        1.5,
                        &conditioning,
                        &optimizer,
                        step,
                    )
                    .unwrap()
            })
            .collect();
        assert!(
            losses[29] < 0.95 * losses[0],
            "the loss went from {} to {}",
            losses[0],
            losses[29]
        );
        assert!(
            lora.adapters
                .values()
                .any(|l| l.up.value.data.iter().any(|v| *v != 0.0))
        );
    }

    #[test]
    fn adapted_device_gradients_match_the_host_or_skip_without_a_device() {
        let Some(gpu) = device() else { return };
        let config = tiny();
        let unet = random(&config, "rusting_brain_cuda_unet_lora_grad.safetensors");
        let mut lora = adapters(&unet);
        lora.adapters
            .values_mut()
            .for_each(|adapter| adapter.scale *= 10.0);
        let (latent, conditioning) = (latent(&config, 8), conditioning(&config, 8, 1.0));
        let sigma = 1.5;
        let mut host = lora.clone();
        let (expected, tape) = host.forward(&unet, &latent, sigma, &conditioning).unwrap();
        let mut grad = expected.clone();
        grad.data
            .iter_mut()
            .enumerate()
            .for_each(|(i, value)| *value = ((i % 13) as f32 - 6.0) * 0.1);
        host.backward(&unet, &tape, &grad).unwrap();

        let model = DeviceUnet::upload(&gpu, &unet).unwrap();
        let mut adapters = DeviceAdapters::upload(&gpu, &unet, &lora).unwrap();
        let (scaled, time) = prepare(&unet, &latent, sigma, &conditioning).unwrap();
        let (_, tape) = model
            .forward_adapted(&adapters, &scaled, &time, &conditioning.text)
            .unwrap();
        model.backward_adapted(&mut adapters, tape, &grad).unwrap();
        let mut device = lora.clone();
        adapters.add_gradients(&gpu, &mut device).unwrap();

        // A missing or misplaced term puts its tensor near 100%; FP16 noise
        // stays at a few percent.
        let (mut total_norm, mut total_error) = (0.0f32, 0.0f32);
        let mut worst = (0.0f32, String::new());
        for (key, expected) in &host.adapters {
            let actual = &device.adapters[key];
            let pairs = [
                ("down", &expected.down.grad, &actual.down.grad),
                ("up", &expected.up.grad, &actual.up.grad),
            ];
            for (name, expected, actual) in pairs {
                let norm = expected.data.iter().map(|v| v * v).sum::<f32>().sqrt();
                let error = expected
                    .data
                    .iter()
                    .zip(&actual.data)
                    .map(|(a, b)| (a - b).powi(2))
                    .sum::<f32>()
                    .sqrt();
                assert!(norm > 0.0, "{key} {name} has no gradient");
                total_norm += norm * norm;
                total_error += error * error;
                if error / norm > worst.0 {
                    worst = (error / norm, format!("{key} {name}"));
                }
            }
        }
        let overall = (total_error / total_norm).sqrt();
        eprintln!(
            "relative gradient error {overall} overall, worst {} at {}",
            worst.0, worst.1
        );
        assert!(overall < 2e-2, "gradients off by {overall} overall");
        assert!(worst.0 < 1e-1, "{} off by {}", worst.1, worst.0);
    }

    /// G10 b-iii-c: one adapted forward and backward of Stable Diffusion
    /// 1.5's UNet at 512x512 (a 64x64 latent, 77 prompt tokens), timed for
    /// `docs/measurements.md`. The weights are random: the step's cost is its
    /// shapes, and the 1.5 checkpoint is not on this machine.
    ///
    /// ```bash
    /// cargo test --release --features cuda --lib an_sd15_step -- --ignored --nocapture
    /// ```
    #[test]
    #[ignore = "builds a 3.4 GB random UNet; run by hand to time a step"]
    fn an_sd15_step_at_512_is_timed() {
        let Some(gpu) = device() else { return };
        let config = crate::unet::UnetConfig::default();
        let unet = random(&config, "rusting_brain_sd15_step.safetensors");
        let lora = adapters(&unet);
        let mut conditioning = conditioning(&config, 64, 1.0);
        conditioning.text = Matrix::new(77, config.cross_dim);
        let latent = latent(&config, 64);
        let model = DeviceUnet::upload(&gpu, &unet).unwrap();
        let mut adapters = DeviceAdapters::upload(&gpu, &unet, &lora).unwrap();
        let (scaled, time) = prepare(&unet, &latent, 1.5, &conditioning).unwrap();
        let mut grad = FeatureMap::new(config.out_channels, 64, 64);
        grad.data
            .iter_mut()
            .enumerate()
            .for_each(|(i, value)| *value = ((i % 7) as f32 - 3.0) * 1e-3);

        let mut step = || {
            let started = std::time::Instant::now();
            let (_, tape) = model
                .forward_adapted(&adapters, &scaled, &time, &conditioning.text)
                .unwrap();
            gpu.synchronize().unwrap();
            let forward = started.elapsed();
            model.backward_adapted(&mut adapters, tape, &grad).unwrap();
            gpu.synchronize().unwrap();
            (forward, started.elapsed())
        };
        step();
        let runs: Vec<_> = (0..5).map(|_| step()).collect();
        let ms = |d: std::time::Duration| d.as_secs_f64() * 1e3;
        let best = runs.iter().min_by_key(|run| run.1).unwrap();
        eprintln!(
            "SD 1.5 adapted step at 512x512: {:.1} ms (forward {:.1} ms, backward {:.1} ms), best of 5; all {:?}",
            ms(best.1),
            ms(best.0),
            ms(best.1 - best.0),
            runs.iter().map(|run| ms(run.1).round()).collect::<Vec<_>>()
        );
    }
}