Repository navigation
Expand file tree
/
Copy pathconditioning_modules.cpp
More file actions
619 lines (540 loc) · 27.1 KB
/
Copy pathconditioning_modules.cpp
File metadata and controls
619 lines (540 loc) · 27.1 KB
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
#include "engine/framework/modules/conditioning_modules.h"
#include "tensor_layout_utils.h"
#include "engine/framework/modules/activation_modules.h"
#include "engine/framework/modules/linear_module.h"
#include "engine/framework/modules/primitive_modules.h"
#include "engine/framework/modules/structural_modules.h"
#include <algorithm>
#include <stdexcept>
namespace engine::modules {
namespace {
void require_positive(int64_t value, const char * name) {
if (value <= 0) {
throw std::runtime_error(std::string(name) + " must be positive");
}
}
core::TensorValue repeat_like(
core::ModuleBuildContext & ctx,
const core::TensorValue & value,
const core::TensorValue & like) {
return core::wrap_tensor(ggml_repeat(ctx.ggml, value.tensor, like.tensor), like.shape, GGML_TYPE_F32);
}
core::TensorValue expand_conditioning_like(
core::ModuleBuildContext & ctx,
const core::TensorValue & value,
const core::TensorValue & like) {
if (value.shape.rank == like.shape.rank) {
return value;
}
if (value.shape.rank + 1 != like.shape.rank) {
throw std::runtime_error("Conditioning tensor rank is not broadcast-compatible with target");
}
core::TensorShape reshaped = {};
reshaped.rank = like.shape.rank;
reshaped.dims[0] = value.shape.dims[0];
reshaped.dims[1] = 1;
for (size_t i = 1; i < value.shape.rank; ++i) {
reshaped.dims[i + 1] = value.shape.dims[i];
}
auto contiguous_value = tensor_layout::ensure_contiguous_layout_if_needed(ctx, value);
auto expanded = core::reshape_tensor(ctx, contiguous_value, reshaped);
return repeat_like(ctx, expanded, like);
}
core::TensorValue slice_last_dim(
core::ModuleBuildContext & ctx,
const core::TensorValue & input,
int64_t offset,
int64_t length) {
return SliceModule({static_cast<int>(input.shape.rank - 1), offset, length}).build(ctx, input);
}
core::TensorValue add_tensors(
core::ModuleBuildContext & ctx,
const core::TensorValue & lhs,
const core::TensorValue & rhs) {
return core::wrap_tensor(ggml_add(ctx.ggml, lhs.tensor, rhs.tensor), lhs.shape, GGML_TYPE_F32);
}
core::TensorValue mul(
core::ModuleBuildContext & ctx,
const core::TensorValue & lhs,
const core::TensorValue & rhs) {
return core::wrap_tensor(ggml_mul(ctx.ggml, lhs.tensor, rhs.tensor), lhs.shape, GGML_TYPE_F32);
}
core::TensorValue scale_tensor(
core::ModuleBuildContext & ctx,
const core::TensorValue & input,
float scale) {
return core::wrap_tensor(ggml_scale(ctx.ggml, input.tensor, scale), input.shape, GGML_TYPE_F32);
}
core::TensorValue apply_adaln(
core::ModuleBuildContext & ctx,
const core::TensorValue & input,
const core::TensorValue & shift,
const core::TensorValue & scale,
float eps,
const std::optional<NormWeights> & norm_weights) {
core::TensorValue normalized;
if (norm_weights.has_value()) {
normalized = LayerNormModule({input.shape.dims[input.shape.rank - 1], eps, true, true}).build(ctx, input, *norm_weights);
} else {
auto input_contiguous = tensor_layout::ensure_contiguous_layout_if_needed(ctx, input);
normalized = core::wrap_tensor(ggml_norm(ctx.ggml, input_contiguous.tensor, eps), input.shape, GGML_TYPE_F32);
}
normalized = tensor_layout::ensure_contiguous_layout_if_needed(ctx, normalized);
auto scaled = mul(ctx, normalized, scale);
auto modulated = add_tensors(ctx, normalized, scaled);
return add_tensors(ctx, modulated, shift);
}
core::TensorValue concat_last_dim(
core::ModuleBuildContext & ctx,
const core::TensorValue & lhs,
const core::TensorValue & rhs) {
return ConcatModule({static_cast<int>(lhs.shape.rank - 1)}).build(ctx, lhs, rhs);
}
core::TensorValue variance_rms_norm(
core::ModuleBuildContext & ctx,
const core::TensorValue & input,
const core::TensorValue & alpha,
float eps) {
core::validate_shape(alpha, core::TensorShape::from_dims({input.shape.dims[input.shape.rank - 1]}), "alpha");
core::TensorShape reduced_shape = input.shape;
reduced_shape.dims[reduced_shape.rank - 1] = 1;
auto mean = core::wrap_tensor(ggml_mean(ctx.ggml, input.tensor), reduced_shape, GGML_TYPE_F32);
auto mean_rep = repeat_like(ctx, mean, input);
auto centered = core::wrap_tensor(ggml_sub(ctx.ggml, input.tensor, mean_rep.tensor), input.shape, GGML_TYPE_F32);
auto squared = core::wrap_tensor(ggml_mul(ctx.ggml, centered.tensor, centered.tensor), input.shape, GGML_TYPE_F32);
auto variance = core::wrap_tensor(ggml_mean(ctx.ggml, squared.tensor), reduced_shape, GGML_TYPE_F32);
const int64_t hidden = input.shape.dims[input.shape.rank - 1];
if (hidden > 1) {
const float correction = static_cast<float>(hidden) / static_cast<float>(hidden - 1);
variance = core::wrap_tensor(ggml_scale(ctx.ggml, variance.tensor, correction), variance.shape, GGML_TYPE_F32);
}
auto alpha_first = SliceModule({0, 0, 1}).build(ctx, alpha);
auto ones = core::wrap_tensor(ggml_div(ctx.ggml, alpha_first.tensor, alpha_first.tensor), alpha_first.shape, GGML_TYPE_F32);
auto eps_tensor = scale_tensor(ctx, repeat_like(ctx, ones, variance), eps);
variance = add_tensors(ctx, variance, eps_tensor);
auto sqrt_var = core::wrap_tensor(ggml_sqrt(ctx.ggml, variance.tensor), variance.shape, GGML_TYPE_F32);
auto alpha_rep = repeat_like(ctx, alpha, input);
auto sqrt_rep = repeat_like(ctx, sqrt_var, input);
auto scale = core::wrap_tensor(ggml_div(ctx.ggml, alpha_rep.tensor, sqrt_rep.tensor), input.shape, GGML_TYPE_F32);
return core::wrap_tensor(ggml_mul(ctx.ggml, input.tensor, scale.tensor), input.shape, GGML_TYPE_F32);
}
const core::ModulePortSpec kSingleInput[] = {
{"input", core::PortKind::Activation, false},
};
const core::ModulePortSpec kSingleOutput[] = {
{"output", core::PortKind::Activation, false},
};
const core::ModulePortSpec kInputConditioningInputs[] = {
{"input", core::PortKind::Activation, false},
{"conditioning", core::PortKind::Activation, false},
};
const core::ModulePortSpec kInputSpeakerInputs[] = {
{"input", core::PortKind::Activation, false},
{"speaker", core::PortKind::Activation, false},
};
const core::ModuleSchema kSpeakerConditioningSchema = {
"SpeakerConditioning",
"nn.conditioning",
kInputSpeakerInputs,
2,
kSingleOutput,
1,
"Projects a speaker embedding to hidden size and adds it across all frames.",
};
const core::ModuleSchema kFiLMSchema = {
"FiLM",
"nn.conditioning",
kInputConditioningInputs,
2,
kSingleOutput,
1,
"Applies feature-wise linear modulation from a conditioning vector.",
};
const core::ModuleSchema kAdaptiveLayerNormSchema = {
"AdaptiveLayerNorm",
"nn.conditioning",
kInputConditioningInputs,
2,
kSingleOutput,
1,
"Applies table-backed adaptive affine or RMS-normalized affine modulation.",
};
const core::ModuleSchema kPriorEncoderBlockSchema = {
"PriorEncoderBlock",
"nn.conditioning",
kSingleInput,
1,
kSingleOutput,
1,
"Projects hidden states to a latent size, applies GELU, then layer norm.",
};
const core::ModulePortSpec kSqueezeExcite1dInputs[] = {
{"input", core::PortKind::Activation, false},
{"fc1_weight", core::PortKind::Parameter, false},
{"fc1_bias", core::PortKind::Parameter, true},
{"fc2_weight", core::PortKind::Parameter, false},
{"fc2_bias", core::PortKind::Parameter, true},
};
const core::ModuleSchema kSqueezeExcite1dSchema = {
"SqueezeExcite1d",
"nn.conditioning",
kSqueezeExcite1dInputs,
5,
kSingleOutput,
1,
"Applies squeeze-excite channel recalibration to channel-first 1D tensors.",
};
} // namespace
core::TensorValue LayerScaleModule::build(
core::ModuleBuildContext & ctx,
const core::TensorValue & input,
const LayerScaleWeights & weights) const {
auto scale = repeat_like(ctx, weights.scale, input);
return mul(ctx, input, scale);
}
TimestepEmbeddingModule::TimestepEmbeddingModule(TimestepEmbeddingConfig config) : config_(config) {
require_positive(config_.frequency_embedding_size, "TimestepEmbeddingConfig.frequency_embedding_size");
require_positive(config_.hidden_size, "TimestepEmbeddingConfig.hidden_size");
if (config_.frequency_embedding_size % 2 != 0) {
throw std::runtime_error("TimestepEmbeddingConfig.frequency_embedding_size must be even");
}
}
core::TensorValue TimestepEmbeddingModule::build(
core::ModuleBuildContext & ctx,
const core::TensorValue & timestep,
const TimestepEmbeddingWeights & weights) const {
core::validate_rank_between(timestep, 1, core::kMaxTensorRank, "timestep");
core::validate_last_dim(timestep, 1, "timestep");
core::validate_shape(weights.freqs, core::TensorShape::from_dims({config_.frequency_embedding_size / 2}), "freqs");
core::TensorShape expanded_shape = timestep.shape;
expanded_shape.dims[expanded_shape.rank - 1] = config_.frequency_embedding_size / 2;
core::TensorShape freqs_shape = {};
freqs_shape.rank = timestep.shape.rank;
for (size_t i = 0; i + 1 < freqs_shape.rank; ++i) {
freqs_shape.dims[i] = 1;
}
freqs_shape.dims[freqs_shape.rank - 1] = config_.frequency_embedding_size / 2;
auto timestep_expanded = RepeatModule({expanded_shape}).build(ctx, timestep);
auto freqs = core::reshape_tensor(ctx, weights.freqs, freqs_shape);
freqs = RepeatModule({expanded_shape}).build(ctx, freqs);
auto args = mul(ctx, timestep_expanded, freqs);
auto cos_part = core::wrap_tensor(ggml_cos(ctx.ggml, args.tensor), args.shape, GGML_TYPE_F32);
auto sin_part = core::wrap_tensor(ggml_sin(ctx.ggml, args.tensor), args.shape, GGML_TYPE_F32);
auto embedding = concat_last_dim(ctx, cos_part, sin_part);
auto hidden = LinearModule({config_.frequency_embedding_size, config_.hidden_size, true}).build(ctx, embedding, weights.fc1);
hidden = SiluModule().build(ctx, hidden);
hidden = LinearModule({config_.hidden_size, config_.hidden_size, true}).build(ctx, hidden, weights.fc2);
if (!weights.rms_weight.has_value()) {
throw std::runtime_error("TimestepEmbeddingWeights.rms_weight is required");
}
return variance_rms_norm(ctx, hidden, *weights.rms_weight, config_.rms_eps);
}
AdaLNResidualMLPModule::AdaLNResidualMLPModule(AdaLNResidualMLPConfig config) : config_(config) {
require_positive(config_.hidden_size, "AdaLNResidualMLPConfig.hidden_size");
require_positive(config_.intermediate_size, "AdaLNResidualMLPConfig.intermediate_size");
require_positive(config_.conditioning_size, "AdaLNResidualMLPConfig.conditioning_size");
}
core::TensorValue AdaLNResidualMLPModule::build(
core::ModuleBuildContext & ctx,
const core::TensorValue & input,
const core::TensorValue & conditioning,
const AdaLNResidualMLPWeights & weights) const {
auto modulation = LinearModule({config_.conditioning_size, config_.hidden_size * 3, config_.use_bias})
.build(ctx, SiluModule().build(ctx, conditioning), weights.modulation.projection);
auto shift = tensor_layout::ensure_contiguous_layout_if_needed(
ctx,
expand_conditioning_like(ctx, slice_last_dim(ctx, modulation, 0, config_.hidden_size), input));
auto scale = tensor_layout::ensure_contiguous_layout_if_needed(
ctx,
expand_conditioning_like(ctx, slice_last_dim(ctx, modulation, config_.hidden_size, config_.hidden_size), input));
auto gate = tensor_layout::ensure_contiguous_layout_if_needed(
ctx,
expand_conditioning_like(ctx, slice_last_dim(ctx, modulation, config_.hidden_size * 2, config_.hidden_size), input));
auto hidden = apply_adaln(ctx, input, shift, scale, config_.eps, weights.norm);
hidden = tensor_layout::ensure_contiguous_layout_if_needed(ctx, hidden);
hidden = LinearModule({config_.hidden_size, config_.intermediate_size, config_.use_bias}).build(ctx, hidden, weights.fc1);
hidden = SiluModule().build(ctx, hidden);
hidden = tensor_layout::ensure_contiguous_layout_if_needed(ctx, hidden);
hidden = LinearModule({config_.intermediate_size, config_.hidden_size, config_.use_bias}).build(ctx, hidden, weights.fc2);
return add_tensors(ctx, input, mul(ctx, gate, hidden));
}
FinalAdaLNProjectionModule::FinalAdaLNProjectionModule(AdaLNResidualMLPConfig config) : config_(config) {
require_positive(config_.hidden_size, "FinalAdaLNProjectionModule.hidden_size");
require_positive(config_.conditioning_size, "FinalAdaLNProjectionModule.conditioning_size");
}
core::TensorValue FinalAdaLNProjectionModule::build(
core::ModuleBuildContext & ctx,
const core::TensorValue & input,
const core::TensorValue & conditioning,
const FinalAdaLNProjectionWeights & weights) const {
auto modulation = LinearModule({config_.conditioning_size, config_.hidden_size * 2, config_.use_bias})
.build(ctx, SiluModule().build(ctx, conditioning), weights.modulation.projection);
auto shift = tensor_layout::ensure_contiguous_layout_if_needed(
ctx,
expand_conditioning_like(ctx, slice_last_dim(ctx, modulation, 0, config_.hidden_size), input));
auto scale = tensor_layout::ensure_contiguous_layout_if_needed(
ctx,
expand_conditioning_like(ctx, slice_last_dim(ctx, modulation, config_.hidden_size, config_.hidden_size), input));
auto input_contiguous = tensor_layout::ensure_contiguous_layout_if_needed(ctx, input);
auto normalized = core::wrap_tensor(ggml_norm(ctx.ggml, input_contiguous.tensor, config_.eps), input.shape, GGML_TYPE_F32);
normalized = tensor_layout::ensure_contiguous_layout_if_needed(ctx, normalized);
auto scaled = mul(ctx, normalized, scale);
auto modulated = add_tensors(ctx, normalized, scaled);
auto hidden = add_tensors(ctx, modulated, shift);
hidden = tensor_layout::ensure_contiguous_layout_if_needed(ctx, hidden);
return LinearModule({config_.hidden_size, config_.intermediate_size, config_.use_bias}).build(ctx, hidden, weights.projection);
}
ConditionedFlowMLPModule::ConditionedFlowMLPModule(ConditionedFlowMLPConfig config) : config_(config) {
require_positive(config_.input_size, "ConditionedFlowMLPConfig.input_size");
require_positive(config_.hidden_size, "ConditionedFlowMLPConfig.hidden_size");
require_positive(config_.condition_size, "ConditionedFlowMLPConfig.condition_size");
require_positive(config_.output_size, "ConditionedFlowMLPConfig.output_size");
require_positive(config_.layers, "ConditionedFlowMLPConfig.layers");
}
core::TensorValue ConditionedFlowMLPModule::build(
core::ModuleBuildContext & ctx,
const core::TensorValue & condition,
const core::TensorValue & input,
const ConditionedFlowMLPWeights & weights) const {
if (static_cast<int64_t>(weights.residual_layers.size()) != config_.layers) {
throw std::runtime_error("ConditionedFlowMLPWeights layer count does not match ConditionedFlowMLPConfig.layers");
}
auto x = LinearModule({config_.input_size, config_.hidden_size, config_.use_bias}).build(ctx, input, weights.input_projection);
auto c = LinearModule({config_.condition_size, config_.hidden_size, config_.use_bias}).build(ctx, condition, weights.condition_projection);
for (const auto & layer_weights : weights.residual_layers) {
x = AdaLNResidualMLPModule({
config_.hidden_size,
config_.hidden_size,
config_.hidden_size,
config_.eps,
config_.use_bias,
}).build(ctx, x, c, layer_weights);
}
return FinalAdaLNProjectionModule({
config_.hidden_size,
config_.output_size,
config_.hidden_size,
config_.eps,
config_.use_bias,
}).build(ctx, x, c, weights.output_projection);
}
TimedConditionedFlowMLPModule::TimedConditionedFlowMLPModule(ConditionedFlowMLPConfig config) : config_(config) {
require_positive(config_.input_size, "TimedConditionedFlowMLPConfig.input_size");
require_positive(config_.hidden_size, "TimedConditionedFlowMLPConfig.hidden_size");
require_positive(config_.condition_size, "TimedConditionedFlowMLPConfig.condition_size");
require_positive(config_.output_size, "TimedConditionedFlowMLPConfig.output_size");
require_positive(config_.layers, "TimedConditionedFlowMLPConfig.layers");
}
core::TensorValue TimedConditionedFlowMLPModule::build(
core::ModuleBuildContext & ctx,
const core::TensorValue & condition,
const core::TensorValue & start_time,
const core::TensorValue & end_time,
const core::TensorValue & input,
const TimedConditionedFlowMLPWeights & weights) const {
if (static_cast<int64_t>(weights.residual_layers.size()) != config_.layers) {
throw std::runtime_error("TimedConditionedFlowMLPWeights layer count does not match config.layers");
}
auto x = LinearModule({config_.input_size, config_.hidden_size, config_.use_bias}).build(ctx, input, weights.input_projection);
auto c = LinearModule({config_.condition_size, config_.hidden_size, config_.use_bias}).build(ctx, condition, weights.condition_projection);
auto t0 = TimestepEmbeddingModule({256, config_.hidden_size, 1.0e-5F}).build(ctx, start_time, weights.start_time_embedding);
auto t1 = TimestepEmbeddingModule({256, config_.hidden_size, 1.0e-5F}).build(ctx, end_time, weights.end_time_embedding);
auto y = add_tensors(ctx, c, scale_tensor(ctx, add_tensors(ctx, t0, t1), 0.5F));
for (const auto & layer_weights : weights.residual_layers) {
x = AdaLNResidualMLPModule({
config_.hidden_size,
config_.hidden_size,
config_.hidden_size,
config_.eps,
config_.use_bias,
}).build(ctx, x, y, layer_weights);
}
return FinalAdaLNProjectionModule({
config_.hidden_size,
config_.output_size,
config_.hidden_size,
config_.eps,
config_.use_bias,
}).build(ctx, x, y, weights.output_projection);
}
SpeakerConditioningModule::SpeakerConditioningModule(SpeakerConditioningConfig config) : config_(config) {
require_positive(config_.hidden_size, "SpeakerConditioningConfig.hidden_size");
require_positive(config_.speaker_dim, "SpeakerConditioningConfig.speaker_dim");
}
const SpeakerConditioningConfig & SpeakerConditioningModule::config() const noexcept {
return config_;
}
const core::ModuleSchema & SpeakerConditioningModule::schema() const noexcept {
return static_schema();
}
core::TensorValue SpeakerConditioningModule::build(
core::ModuleBuildContext & ctx,
const core::TensorValue & input,
const core::TensorValue & speaker,
const SpeakerConditioningWeights & weights) const {
core::validate_shape(speaker, core::TensorShape::from_dims({input.shape.dims[0], config_.speaker_dim}), "speaker");
core::validate_last_dim(input, config_.hidden_size, "input");
const auto projected = LinearModule({config_.speaker_dim, config_.hidden_size, config_.use_bias}).build(
ctx,
speaker,
LinearWeights{weights.proj_weight, weights.proj_bias});
return add_tensors(ctx, input, expand_conditioning_like(ctx, projected, input));
}
const core::ModuleSchema & SpeakerConditioningModule::static_schema() noexcept {
return kSpeakerConditioningSchema;
}
FiLMModule::FiLMModule(FiLMConfig config) : config_(config) {
require_positive(config_.hidden_size, "FiLMConfig.hidden_size");
require_positive(config_.conditioning_dim, "FiLMConfig.conditioning_dim");
}
const FiLMConfig & FiLMModule::config() const noexcept {
return config_;
}
const core::ModuleSchema & FiLMModule::schema() const noexcept {
return static_schema();
}
core::TensorValue FiLMModule::build(
core::ModuleBuildContext & ctx,
const core::TensorValue & input,
const core::TensorValue & conditioning,
const FiLMWeights & weights) const {
core::validate_shape(
conditioning,
core::TensorShape::from_dims({input.shape.dims[0], config_.conditioning_dim}),
"conditioning");
core::validate_last_dim(input, config_.hidden_size, "input");
const LinearModule gamma({config_.conditioning_dim, config_.hidden_size, config_.use_bias});
const LinearModule beta({config_.conditioning_dim, config_.hidden_size, config_.use_bias});
auto gamma_vec = gamma.build(ctx, conditioning, LinearWeights{weights.gamma_weight, weights.gamma_bias});
auto beta_vec = beta.build(ctx, conditioning, LinearWeights{weights.beta_weight, weights.beta_bias});
auto scaled = mul(ctx, input, expand_conditioning_like(ctx, gamma_vec, input));
return add_tensors(ctx, scaled, expand_conditioning_like(ctx, beta_vec, input));
}
const core::ModuleSchema & FiLMModule::static_schema() noexcept {
return kFiLMSchema;
}
AdaptiveLayerNormModule::AdaptiveLayerNormModule(AdaptiveLayerNormConfig config) : config_(config) {
require_positive(config_.hidden_size, "AdaptiveLayerNormConfig.hidden_size");
if (config_.scale_index < 0) {
throw std::runtime_error("AdaptiveLayerNormConfig.scale_index must be non-negative");
}
if (config_.mode != AdaptiveLayerNormMode::Scale && config_.shift_index < 0) {
throw std::runtime_error("AdaptiveLayerNormConfig.shift_index must be non-negative unless mode is Scale");
}
}
const AdaptiveLayerNormConfig & AdaptiveLayerNormModule::config() const noexcept {
return config_;
}
const core::ModuleSchema & AdaptiveLayerNormModule::schema() const noexcept {
return static_schema();
}
core::TensorValue AdaptiveLayerNormModule::build(
core::ModuleBuildContext & ctx,
const core::TensorValue & input,
const core::TensorValue & modulation,
const AdaptiveLayerNormWeights & weights) const {
core::validate_last_dim(input, config_.hidden_size, "input");
if (modulation.shape.rank != input.shape.rank) {
throw std::runtime_error("AdaptiveLayerNorm modulation rank must match input rank");
}
for (size_t dim = 0; dim + 1 < input.shape.rank; ++dim) {
if (modulation.shape.dims[dim] != input.shape.dims[dim] && modulation.shape.dims[dim] != 1) {
throw std::runtime_error("AdaptiveLayerNorm modulation shape must match or broadcast to input except the last dimension");
}
}
if (weights.table.shape.rank != 2 || weights.table.shape.dims[1] != config_.hidden_size) {
throw std::runtime_error("AdaptiveLayerNorm table shape mismatch");
}
const int64_t max_index = std::max(config_.shift_index, config_.scale_index);
if (weights.table.shape.dims[0] <= max_index ||
modulation.shape.dims[modulation.shape.rank - 1] < (max_index + 1) * config_.hidden_size) {
throw std::runtime_error("AdaptiveLayerNorm modulation/table index out of range");
}
auto scale = SliceModule({static_cast<int>(modulation.shape.rank - 1), config_.scale_index * config_.hidden_size, config_.hidden_size})
.build(ctx, modulation);
auto scale_table = SliceModule({0, config_.scale_index, 1}).build(ctx, weights.table);
core::TensorShape table_shape = {};
table_shape.rank = modulation.shape.rank;
for (size_t dim = 0; dim + 1 < table_shape.rank; ++dim) {
table_shape.dims[dim] = 1;
}
table_shape.dims[table_shape.rank - 1] = config_.hidden_size;
scale_table = core::reshape_tensor(ctx, scale_table, table_shape);
scale = add_tensors(ctx, scale, scale_table);
if (config_.mode == AdaptiveLayerNormMode::Scale) {
return mul(ctx, input, scale);
}
auto shift = SliceModule({static_cast<int>(modulation.shape.rank - 1), config_.shift_index * config_.hidden_size, config_.hidden_size})
.build(ctx, modulation);
auto shift_table = SliceModule({0, config_.shift_index, 1}).build(ctx, weights.table);
shift_table = core::reshape_tensor(ctx, shift_table, table_shape);
shift = add_tensors(ctx, shift, shift_table);
const auto normalized = config_.mode == AdaptiveLayerNormMode::RmsAffine
? RMSNormModule({config_.hidden_size, config_.eps, false, false}).build(ctx, input, {})
: input;
auto one_plus_scale = core::wrap_tensor(
ggml_scale_bias(ctx.ggml, scale.tensor, 1.0F, 1.0F),
scale.shape,
GGML_TYPE_F32);
auto scaled = mul(ctx, normalized, one_plus_scale);
return add_tensors(ctx, scaled, shift);
}
const core::ModuleSchema & AdaptiveLayerNormModule::static_schema() noexcept {
return kAdaptiveLayerNormSchema;
}
PriorEncoderBlockModule::PriorEncoderBlockModule(PriorEncoderBlockConfig config) : config_(config) {
require_positive(config_.hidden_size, "PriorEncoderBlockConfig.hidden_size");
require_positive(config_.latent_size, "PriorEncoderBlockConfig.latent_size");
}
const PriorEncoderBlockConfig & PriorEncoderBlockModule::config() const noexcept {
return config_;
}
const core::ModuleSchema & PriorEncoderBlockModule::schema() const noexcept {
return static_schema();
}
core::TensorValue PriorEncoderBlockModule::build(
core::ModuleBuildContext & ctx,
const core::TensorValue & input,
const PriorEncoderBlockWeights & weights) const {
core::validate_last_dim(input, config_.hidden_size, "input");
const LinearModule proj({config_.hidden_size, config_.latent_size, config_.use_bias});
const GeluModule gelu({GeluApproximation::ExactErf});
const LayerNormModule norm({config_.latent_size, config_.eps, true, true});
auto output = proj.build(ctx, input, LinearWeights{weights.proj_weight, weights.proj_bias});
output = gelu.build(ctx, output);
return norm.build(ctx, output, weights.norm);
}
const core::ModuleSchema & PriorEncoderBlockModule::static_schema() noexcept {
return kPriorEncoderBlockSchema;
}
SqueezeExcite1dModule::SqueezeExcite1dModule(SqueezeExcite1dConfig config) : config_(config) {
require_positive(config_.channels, "SqueezeExcite1dConfig.channels");
require_positive(config_.hidden_channels, "SqueezeExcite1dConfig.hidden_channels");
}
const SqueezeExcite1dConfig & SqueezeExcite1dModule::config() const noexcept {
return config_;
}
const core::ModuleSchema & SqueezeExcite1dModule::schema() const noexcept {
return static_schema();
}
core::TensorValue SqueezeExcite1dModule::build(
core::ModuleBuildContext & ctx,
const core::TensorValue & input,
const SqueezeExcite1dWeights & weights) const {
core::validate_rank_between(input, 3, 3, "input");
core::validate_shape(
input,
core::TensorShape::from_dims({input.shape.dims[0], config_.channels, input.shape.dims[2]}),
"input");
auto pooled = ReduceMeanModule({static_cast<int>(input.shape.rank - 1)}).build(ctx, input);
auto gate = Conv1dModule({config_.channels, config_.hidden_channels, 1, 1, 0, 1, config_.use_bias}).build(ctx, pooled, weights.fc1);
gate = ReluModule().build(ctx, gate);
gate = Conv1dModule({config_.hidden_channels, config_.channels, 1, 1, 0, 1, config_.use_bias}).build(ctx, gate, weights.fc2);
gate = SigmoidModule().build(ctx, gate);
gate = RepeatModule({input.shape}).build(ctx, gate);
return MulModule().build(ctx, input, gate);
}
const core::ModuleSchema & SqueezeExcite1dModule::static_schema() noexcept {
return kSqueezeExcite1dSchema;
}
} // namespace engine::modules