File size: 26,434 Bytes
ac69be1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Modified Wan2.2 Diffusion Transformer for LoomVideo.

Extends the HuggingFace Diffusers Wan transformer to support:
- Separate self-attention and cross-attention execution paths
- Variable-length sequence handling with cu_seq_lens
- Source video conditioning via learnable patch embedding
- Reference image/video conditioning with custom temporal RoPE offsets

Reference: https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer_wan.py
"""

import math
from typing import Optional, Tuple, List

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F

from diffusers.configuration_utils import register_to_config
from diffusers.utils import logging
from diffusers.models.attention import FeedForward
from diffusers.models.embeddings import get_1d_rotary_pos_embed
from diffusers.models.normalization import FP32LayerNorm
from diffusers.models.transformers.transformer_wan import (
    WanAttention,
    WanRotaryPosEmbed,
    WanTimeTextImageEmbedding,
    WanAttnProcessor,
    WanTransformerBlock,
    WanTransformer3DModel,
    _get_qkv_projections,
)

logger = logging.get_logger(__name__)


class WanAttnProcessor(WanAttnProcessor):
    """Attention processor with QKV extraction for cross-attention in LoomVideo."""

    _attention_backend = None

    def __init__(self):
        super().__init__()
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError("WanAttnProcessor requires PyTorch 2.0+.")

    def get_qkv(
        self,
        attn: "WanAttention",
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
    ):
        """
        Extract Q, K, V projections with optional rotary embeddings.

        Returns:
            Tuple of (query, key, value, encoder_hidden_states, encoder_hidden_states_img).
        """
        encoder_hidden_states_img = None
        if attn.add_k_proj is not None:
            # 512 is the context length of the text encoder
            image_context_length = encoder_hidden_states.shape[1] - 512
            encoder_hidden_states_img = encoder_hidden_states[:, :image_context_length]
            encoder_hidden_states = encoder_hidden_states[:, image_context_length:]

        query, key, value = _get_qkv_projections(
            attn, hidden_states, encoder_hidden_states
        )

        query = attn.norm_q(query)
        key = attn.norm_k(key)

        query = query.unflatten(2, (attn.heads, -1))
        key = key.unflatten(2, (attn.heads, -1))
        value = value.unflatten(2, (attn.heads, -1))

        if rotary_emb is not None:

            def apply_rotary_emb(
                hidden_states: torch.Tensor,
                freqs_cos: torch.Tensor,
                freqs_sin: torch.Tensor,
            ):
                x1, x2 = hidden_states.unflatten(-1, (-1, 2)).unbind(-1)
                cos = freqs_cos[..., 0::2]
                sin = freqs_sin[..., 1::2]
                out = torch.empty_like(hidden_states)
                out[..., 0::2] = x1 * cos - x2 * sin
                out[..., 1::2] = x1 * sin + x2 * cos
                return out.type_as(hidden_states)

            query = apply_rotary_emb(query, *rotary_emb)
            key = apply_rotary_emb(key, *rotary_emb)

        # [B, L, Num_heads, Head_dims] -> [B, Num_heads, L, Head_dims]
        query = query.transpose(1, 2)
        key = key.transpose(1, 2)
        value = value.transpose(1, 2)

        return query, key, value, encoder_hidden_states, encoder_hidden_states_img


class WanTransformerBlock(WanTransformerBlock):
    """
    Extended transformer block with separate self-attention and cross-attention paths.

    Adds methods for split execution:
    - forward_selfattn: self-attention only (for training with flash cross-attn)
    - forward_crossattn_later_layer: cross-attn + FFN (paired with forward_selfattn)
    """

    def __init__(
        self,
        dim: int,
        ffn_dim: int,
        num_heads: int,
        qk_norm: str = "rms_norm_across_heads",
        cross_attn_norm: bool = False,
        eps: float = 1e-6,
        added_kv_proj_dim: Optional[int] = None,
    ):
        super().__init__(
            dim, ffn_dim, num_heads, qk_norm, cross_attn_norm, eps, added_kv_proj_dim
        )

        self.dim = dim
        self.num_heads = num_heads
        self.eps = eps

        # 1. Self-attention
        self.norm1 = FP32LayerNorm(dim, eps, elementwise_affine=False)
        self.attn1 = WanAttention(
            dim=dim,
            heads=num_heads,
            dim_head=dim // num_heads,
            eps=eps,
            cross_attention_dim_head=None,
            processor=WanAttnProcessor(),
        )

        # 2. Cross-attention
        self.attn2 = WanAttention(
            dim=dim,
            heads=num_heads,
            dim_head=dim // num_heads,
            eps=eps,
            added_kv_proj_dim=added_kv_proj_dim,
            cross_attention_dim_head=dim // num_heads,
            processor=WanAttnProcessor(),
        )
        self.norm2 = (
            FP32LayerNorm(dim, eps, elementwise_affine=True)
            if cross_attn_norm
            else nn.Identity()
        )

        # 3. Feed-forward
        self.ffn = FeedForward(dim, inner_dim=ffn_dim, activation_fn="gelu-approximate")
        self.norm3 = FP32LayerNorm(dim, eps, elementwise_affine=False)

        self.scale_shift_table = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5)

    def _parse_temb(self, temb: torch.Tensor):
        """Parse timestep embedding into shift/scale/gate components."""
        if temb.ndim == 4:
            shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
                self.scale_shift_table.unsqueeze(0) + temb.float()
            ).chunk(6, dim=2)
            shift_msa = shift_msa.squeeze(2)
            scale_msa = scale_msa.squeeze(2)
            gate_msa = gate_msa.squeeze(2)
            c_shift_msa = c_shift_msa.squeeze(2)
            c_scale_msa = c_scale_msa.squeeze(2)
            c_gate_msa = c_gate_msa.squeeze(2)
        else:
            shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
                self.scale_shift_table + temb.float()
            ).chunk(6, dim=1)
        return shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa

    def forward_selfattn(
        self,
        hidden_states: torch.Tensor,
        temb: torch.Tensor,
        rotary_emb: torch.Tensor,
        cu_seq_lens: list = None,
    ) -> torch.Tensor:
        """
        Execute only the self-attention portion of this block.

        Args:
            hidden_states: Input tensor.
            temb: Timestep embedding (adaptive normalization parameters).
            rotary_emb: Rotary position embeddings.
            cu_seq_lens: Cumulative sequence lengths for per-sample attention.

        Returns:
            Hidden states after self-attention (before cross-attention and FFN).
        """
        shift_msa, scale_msa, gate_msa, _, _, _ = self._parse_temb(temb)

        norm_hidden_states = (self.norm1(hidden_states.float()) * (1 + scale_msa) + shift_msa).type_as(hidden_states)

        if cu_seq_lens is not None and len(cu_seq_lens) > 2:
            # Per-sample self-attention for variable-length sequences
            attn_outputs = []
            for i in range(len(cu_seq_lens) - 1):
                start = int(cu_seq_lens[i])
                end = int(cu_seq_lens[i + 1])
                sample_norm_hidden = norm_hidden_states[:, start:end, :]

                if isinstance(rotary_emb, (list, tuple)):
                    sample_rotary = [r[:, start:end, ...] if r is not None else None for r in rotary_emb]
                elif rotary_emb is not None:
                    sample_rotary = rotary_emb[:, start:end, ...]
                else:
                    sample_rotary = None
                sample_attn_output = self.attn1(sample_norm_hidden, None, None, sample_rotary)
                attn_outputs.append(sample_attn_output)

            attn_output = torch.cat(attn_outputs, dim=1)
        else:
            attn_output = self.attn1(norm_hidden_states, None, None, rotary_emb)

        hidden_states = (hidden_states.float() + attn_output * gate_msa).type_as(hidden_states)

        return hidden_states

    def forward_crossattn_later_layer(
        self,
        hidden_states: torch.Tensor,
        attn_output: torch.Tensor,
        temb: torch.Tensor,
    ) -> torch.Tensor:
        """
        Execute cross-attention residual addition and feed-forward network.

        Args:
            hidden_states: Current hidden states (after self-attention).
            attn_output: Cross-attention output to be added as residual.
            temb: Timestep embedding for adaptive normalization.

        Returns:
            Hidden states after cross-attention residual and FFN.
        """
        _, _, _, c_shift_msa, c_scale_msa, c_gate_msa = self._parse_temb(temb)

        hidden_states = hidden_states + attn_output

        # Feed-forward network
        norm_hidden_states = (self.norm3(hidden_states.float()) * (1 + c_scale_msa) + c_shift_msa).type_as(hidden_states)
        ff_output = self.ffn(norm_hidden_states)
        hidden_states = (hidden_states.float() + ff_output.float() * c_gate_msa).type_as(hidden_states)

        return hidden_states


class WanTransformer3DModel(WanTransformer3DModel):
    """
    Extended Wan 3D Transformer with source/reference conditioning support.

    Adds:
    - Source video conditioning via learnable patch embedding
    - Reference conditioning with custom temporal RoPE offsets
    - Separate early-layer and output-projection methods for flexible fusion
    """

    _supports_gradient_checkpointing = True
    _skip_layerwise_casting_patterns = ["patch_embedding", "condition_embedder", "norm"]
    _no_split_modules = ["WanTransformerBlock"]
    _keep_in_fp32_modules = [
        "time_embedder",
        "scale_shift_table",
        "norm1",
        "norm2",
        "norm3",
    ]
    _keys_to_ignore_on_load_unexpected = ["norm_added_q"]
    _repeated_blocks = ["WanTransformerBlock"]

    @register_to_config
    def __init__(
        self,
        patch_size: Tuple[int] = (1, 2, 2),
        num_attention_heads: int = 40,
        attention_head_dim: int = 128,
        in_channels: int = 16,
        out_channels: int = 16,
        text_dim: int = 4096,
        freq_dim: int = 256,
        ffn_dim: int = 13824,
        num_layers: int = 40,
        cross_attn_norm: bool = True,
        qk_norm: Optional[str] = "rms_norm_across_heads",
        eps: float = 1e-6,
        image_dim: Optional[int] = None,
        added_kv_proj_dim: Optional[int] = None,
        rope_max_seq_len: int = 1024,
        pos_embed_seq_len: Optional[int] = None,
    ):
        super().__init__(
            patch_size,
            num_attention_heads,
            attention_head_dim,
            in_channels,
            out_channels,
            text_dim,
            freq_dim,
            ffn_dim,
            num_layers,
            cross_attn_norm,
            qk_norm,
            eps,
            image_dim,
            added_kv_proj_dim,
            rope_max_seq_len,
            pos_embed_seq_len,
        )

        inner_dim = num_attention_heads * attention_head_dim
        out_channels = out_channels or in_channels
        self.in_channels = in_channels
        self.ffn_dim = ffn_dim
        self.qk_norm = qk_norm
        self.cross_attn_norm = cross_attn_norm
        self.attention_head_dim = attention_head_dim
        self.num_attention_heads = num_attention_heads
        self.eps = eps
        self.added_kv_proj_dim = added_kv_proj_dim
        self.num_layers = num_layers
        self.inner_dim = inner_dim

        # 1. Patch & position embedding
        self.rope = WanRotaryPosEmbed(attention_head_dim, patch_size, rope_max_seq_len)
        self.patch_embedding = nn.Conv3d(
            in_channels, inner_dim, kernel_size=patch_size, stride=patch_size
        )

        # 2. Condition embeddings
        self.condition_embedder = WanTimeTextImageEmbedding(
            dim=inner_dim,
            time_freq_dim=freq_dim,
            time_proj_dim=inner_dim * 6,
            text_embed_dim=text_dim,
            image_embed_dim=image_dim,
            pos_embed_seq_len=pos_embed_seq_len,
        )

        # 3. Transformer blocks
        self.blocks = nn.ModuleList(
            [
                WanTransformerBlock(
                    inner_dim,
                    ffn_dim,
                    num_attention_heads,
                    qk_norm,
                    cross_attn_norm,
                    eps,
                    added_kv_proj_dim,
                )
                for _ in range(num_layers)
            ]
        )

        # 4. Output norm & projection
        self.norm_out = FP32LayerNorm(inner_dim, eps, elementwise_affine=False)
        self.proj_out = nn.Linear(inner_dim, out_channels * math.prod(patch_size))
        self.scale_shift_table = nn.Parameter(
            torch.randn(1, 2, inner_dim) / inner_dim**0.5
        )

        self.gradient_checkpointing = True

    def compute_ref_rotary_emb(
        self,
        ref_shape: Tuple[int, ...],
        time_offset: int,
        device: torch.device,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Compute 3D RoPE for a reference latent with a custom temporal offset.

        Reference frames share the same temporal position (given by time_offset),
        while spatial dimensions use normal indices starting from 0.

        Args:
            ref_shape: (B, C, T, H, W) shape of the reference latent tensor.
            time_offset: Temporal position index for all frames (e.g., -10, -20).
            device: Target device.

        Returns:
            Tuple of (freqs_cos, freqs_sin) tensors.
        """
        _, _, num_frames, height, width = ref_shape
        p_t, p_h, p_w = self.config.patch_size
        ppf = num_frames // p_t
        pph = height // p_h
        ppw = width // p_w

        rope_module = self.rope
        freqs_dtype = torch.float64

        # Temporal: all frames share the same time_offset
        time_pos = np.array([time_offset] * ppf, dtype=np.float64)
        freq_cos_t, freq_sin_t = get_1d_rotary_pos_embed(
            rope_module.t_dim, time_pos, theta=10000.0,
            use_real=True, repeat_interleave_real=True, freqs_dtype=freqs_dtype,
        )

        # Spatial: normal indices [0, 1, ...]
        split_sizes = [rope_module.t_dim, rope_module.h_dim, rope_module.w_dim]
        precomputed_cos = rope_module.freqs_cos.split(split_sizes, dim=1)
        precomputed_sin = rope_module.freqs_sin.split(split_sizes, dim=1)

        freq_cos_h = precomputed_cos[1][:pph]
        freq_sin_h = precomputed_sin[1][:pph]
        freq_cos_w = precomputed_cos[2][:ppw]
        freq_sin_w = precomputed_sin[2][:ppw]

        # Broadcast to (ppf, pph, ppw, dim_*)
        freq_cos_t = freq_cos_t.to(device).view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
        freq_cos_h = freq_cos_h.to(device).view(1, pph, 1, -1).expand(ppf, pph, ppw, -1)
        freq_cos_w = freq_cos_w.to(device).view(1, 1, ppw, -1).expand(ppf, pph, ppw, -1)

        freq_sin_t = freq_sin_t.to(device).view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
        freq_sin_h = freq_sin_h.to(device).view(1, pph, 1, -1).expand(ppf, pph, ppw, -1)
        freq_sin_w = freq_sin_w.to(device).view(1, 1, ppw, -1).expand(ppf, pph, ppw, -1)

        freqs_cos = torch.cat([freq_cos_t, freq_cos_h, freq_cos_w], dim=-1).reshape(1, ppf * pph * ppw, 1, -1)
        freqs_sin = torch.cat([freq_sin_t, freq_sin_h, freq_sin_w], dim=-1).reshape(1, ppf * pph * ppw, 1, -1)

        return freqs_cos, freqs_sin

    def forward_early_layers(
        self,
        hidden_states: List[torch.Tensor],
        timestep: torch.LongTensor,
        encoder_hidden_states: torch.Tensor,
        source_hidden_states: Optional[List[torch.Tensor]] = None,
        source_scale: Optional[torch.Tensor] = None,
        ref_hidden_states: Optional[List[torch.Tensor]] = None,
    ):
        """
        Process patch embedding, timestep conditioning, source/ref concatenation.

        Prepares all inputs for the main transformer blocks by:
        1. Applying patch embedding to each latent sample
        2. Adding source conditioning (scaled by timestep)
        3. Concatenating reference latents with custom temporal RoPE
        4. Computing timestep projections

        Args:
            hidden_states: List of latent tensors (None for non-generation samples).
            timestep: Diffusion timestep for each generation sample.
            encoder_hidden_states: T5 text encoder hidden states.
            source_hidden_states: Optional source video latents.
            source_scale: Timestep-dependent scale for source conditioning.
            ref_hidden_states: Optional list of reference latent lists.

        Returns:
            Tuple of processed tensors and metadata for the transformer blocks.
        """
        hidden_states_list = []
        shape_list = []
        rotary_emb_cos = []
        rotary_emb_sin = []
        cu_seq_lens = [0]
        sample_index = []
        seq_lens = []
        gen_seq_lens = []

        for index, hidden_state in enumerate(hidden_states):
            if hidden_state is not None:
                valid_sample_idx = len(seq_lens)
                shape = hidden_state.shape

                hidden_state = self.patch_embedding(hidden_state)
                hidden_state = hidden_state.flatten(2).transpose(1, 2)  # [1, L, C]

                # Add source video conditioning
                if (source_hidden_states is not None
                        and source_hidden_states[index] is not None
                        and source_scale is not None
                        and hasattr(self, 'source_patch_embedding')):
                    source_input = source_hidden_states[index]
                    source_encoded = self.source_patch_embedding(source_input)
                    source_encoded = source_encoded.flatten(2).transpose(1, 2)
                    source_scale_idx = index if source_scale.shape[0] == len(hidden_states) else valid_sample_idx
                    hidden_state = hidden_state + source_encoded * source_scale[source_scale_idx]

                # Record original sequence length (for loss computation, excluding ref tokens)
                original_seq_len = hidden_state.shape[1]
                gen_seq_lens.append(original_seq_len)

                # Compute rotary embeddings for the main latent
                fake_tensor = torch.empty(shape, device=hidden_state.device, dtype=hidden_state.dtype)
                rotary_emb = self.rope(fake_tensor)

                # Concatenate reference latents with negative temporal offsets
                if (ref_hidden_states is not None and ref_hidden_states[index] is not None):
                    ref_list = ref_hidden_states[index]
                    for ref_idx, ref_input in enumerate(ref_list):
                        ref_encoded = self.patch_embedding(ref_input)
                        ref_encoded = ref_encoded.flatten(2).transpose(1, 2)
                        hidden_state = torch.cat([hidden_state, ref_encoded], dim=1)

                        # Negative temporal offset: -10, -20, -30, ... for each ref
                        ref_time_offset = -10 * (ref_idx + 1)
                        ref_rotary_emb = self.compute_ref_rotary_emb(
                            ref_shape=ref_input.shape,
                            time_offset=ref_time_offset,
                            device=hidden_state.device,
                        )
                        rotary_emb = (
                            torch.cat([rotary_emb[0], ref_rotary_emb[0]], dim=1),
                            torch.cat([rotary_emb[1], ref_rotary_emb[1]], dim=1),
                        )

                hidden_states_list.append(hidden_state)
                rotary_emb_cos.append(rotary_emb[0])
                rotary_emb_sin.append(rotary_emb[1])
                cu_seq_lens.append(cu_seq_lens[-1] + hidden_state.shape[1])
                sample_index.append(index)
                shape_list.append(shape)
                seq_lens.append(hidden_state.shape[1])

        hidden_states = torch.cat(hidden_states_list, dim=1)  # [1, sum(L), C]
        rotary_emb_cos = torch.cat(rotary_emb_cos, dim=1)
        rotary_emb_sin = torch.cat(rotary_emb_sin, dim=1)
        rotary_emb = (rotary_emb_cos, rotary_emb_sin)

        # Timestep conditioning
        if timestep.ndim == 2:
            ts_seq_len = timestep.shape[1]
            timestep = timestep.flatten()
        else:
            ts_seq_len = None

        temb, timestep_proj, encoder_hidden_states, _ = (
            self.condition_embedder(
                timestep,
                encoder_hidden_states,
                timestep_seq_len=ts_seq_len,
            )
        )
        if ts_seq_len is not None:
            timestep_proj = timestep_proj.unflatten(2, (6, -1))
        else:
            timestep_proj = timestep_proj.unflatten(1, (6, -1))

        # Expand timestep projections to match sequence lengths (zero for ref tokens)
        num_valid_samples = len(seq_lens)
        if timestep_proj.shape[0] == num_valid_samples and num_valid_samples > 0:
            expanded_timestep_proj = []
            expanded_temb = []
            for i, length in enumerate(seq_lens):
                gen_len = gen_seq_lens[i]
                ref_len = length - gen_len
                if ts_seq_len is None:
                    expanded_timestep_proj.append(timestep_proj[i:i + 1].expand(gen_len, -1, -1))
                    if temb is not None:
                        expanded_temb.append(temb[i:i + 1].expand(gen_len, -1))
                    # Zero timestep embedding for ref tokens
                    if ref_len > 0:
                        expanded_timestep_proj.append(
                            torch.zeros(ref_len, timestep_proj.shape[1], timestep_proj.shape[2],
                                        device=timestep_proj.device, dtype=timestep_proj.dtype)
                        )
                        if temb is not None:
                            expanded_temb.append(
                                torch.zeros(ref_len, temb.shape[1], device=temb.device, dtype=temb.dtype)
                            )
                else:
                    expanded_timestep_proj.append(timestep_proj[i])
                    if temb is not None:
                        expanded_temb.append(temb[i])

            if ts_seq_len is None:
                timestep_proj = torch.cat(expanded_timestep_proj, dim=0).unsqueeze(0)
                if temb is not None:
                    temb = torch.cat(expanded_temb, dim=0).unsqueeze(0)
            else:
                timestep_proj = torch.cat(expanded_timestep_proj, dim=0).unsqueeze(0)
                if temb is not None:
                    temb = torch.cat(expanded_temb, dim=0).unsqueeze(0)

        return (
            hidden_states,
            encoder_hidden_states,
            timestep_proj,
            rotary_emb,
            temb,
            shape_list,
            cu_seq_lens,
            sample_index,
            gen_seq_lens,
        )

    def get_output(self, hidden_states: torch.Tensor, temb: torch.Tensor, cu_seq_lens, shape, gen_seq_lens=None):
        """
        Project hidden states back to pixel space via unpatchify.

        Args:
            hidden_states: Transformer output tensor.
            temb: Timestep embedding for final adaptive normalization.
            cu_seq_lens: Cumulative sequence lengths.
            shape: Original latent shapes per sample.
            gen_seq_lens: Original sequence lengths (excluding ref tokens).
                If provided, only these tokens are projected (ref tokens discarded).

        Returns:
            List of output tensors, one per sample.
        """
        output_list = []

        for index in range(len(cu_seq_lens) - 1):
            start_idx = cu_seq_lens[index]
            # Only project original tokens (exclude ref tokens)
            if gen_seq_lens is not None:
                end_idx = cu_seq_lens[index] + gen_seq_lens[index]
            else:
                end_idx = cu_seq_lens[index + 1]
            hidden_state = hidden_states[:, start_idx:end_idx, :]

            if temb.ndim == 3 and temb.shape[1] == hidden_states.shape[1]:
                sample_temb = temb[:, start_idx:end_idx, :]
            elif temb.ndim == 2 and temb.shape[0] > 1:
                sample_temb = temb[index:index + 1]
            else:
                sample_temb = temb

            batch_size, num_channels, num_frames, height, width = shape[index]
            p_t, p_h, p_w = self.config.patch_size
            post_patch_num_frames = num_frames // p_t
            post_patch_height = height // p_h
            post_patch_width = width // p_w

            # Adaptive normalization for output
            if sample_temb.ndim == 3:
                shift, scale = (
                    self.scale_shift_table.unsqueeze(0) + sample_temb.unsqueeze(2)
                ).chunk(2, dim=2)
                shift = shift.squeeze(2)
                scale = scale.squeeze(2)
            else:
                shift, scale = (self.scale_shift_table + sample_temb.unsqueeze(1)).chunk(2, dim=1)

            shift = shift.to(hidden_state.device)
            scale = scale.to(hidden_state.device)

            hidden_state = (
                self.norm_out(hidden_state.float()) * (1 + scale) + shift
            ).type_as(hidden_state)
            hidden_state = self.proj_out(hidden_state)

            # Unpatchify: reshape back to video dimensions
            hidden_state = hidden_state.reshape(
                batch_size,
                post_patch_num_frames,
                post_patch_height,
                post_patch_width,
                p_t,
                p_h,
                p_w,
                -1,
            )
            hidden_state = hidden_state.permute(0, 7, 1, 4, 2, 5, 3, 6)
            output = hidden_state.flatten(6, 7).flatten(4, 5).flatten(2, 3)
            output_list.append(output)

        return output_list