File size: 22,261 Bytes
247228a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Reusable DiT building blocks needed for inference.
Extracted from networks/DiT/dit.py so the inference set avoids importing
the full training-only model classes in that file.
"""
import math

import numpy as np
import torch
import torch.nn as nn
from einops import rearrange
from timm.layers.mlp import SwiGLU
from timm.models.vision_transformer import Attention, Mlp


def get_norm_layer(norm_layer):
    if isinstance(norm_layer, str):
        if norm_layer == 'layer_norm':
            return nn.LayerNorm
        elif norm_layer == 'rms_norm':
            return nn.RMSNorm
        else:
            raise ValueError(f"Unsupported norm layer: {norm_layer}")
    return norm_layer


def modulate(x, shift, scale):
    n = x.ndim - shift.ndim
    for _ in range(n):
        shift = shift.unsqueeze(-2)
        scale = scale.unsqueeze(-2)
    return x * (1 + scale) + shift


def _broadcast_gate(gate, x):
    n = x.ndim - gate.ndim
    for _ in range(n):
        gate = gate.unsqueeze(-2)
    return gate


class FrequencyEncoder:
   def __init__(self, embed_dim, freq_min=1, freq_max=5):
       """
       Deterministic frequency encoder with fixed normalization.


       Args:
           embed_dim (int): Dimensionality of the token embeddings.
           freq_min (float): Minimum frequency value.
           freq_max (float): Maximum frequency value.
       """
       self.embed_dim = embed_dim
       self.freq_min = freq_min
       self.freq_max = freq_max


   def encode(self, frequencies):
       """
       Encodes frequencies into embeddings using sine-cosine features.


       Args:
           frequencies (torch.Tensor): Tensor of shape (batch_size,) containing frequencies.


       Returns:
           torch.Tensor: Encoded frequency embeddings of shape (batch_size, embed_dim).
       """
       batch_size = frequencies.size(0)


       # Fixed normalization: Scale frequencies to [0, 1]
       normalized_freq = (frequencies - self.freq_min) / (self.freq_max - self.freq_min)


       # Generate positional features using sine and cosine. `frequencies` supplies the device
       # via new_tensor/new_zeros instead of a bare `.device` read, which crashes Dynamo on
       # some torch builds (see TimestepEmbedder in modules/dit.py for the same issue).
       positions = frequencies.new_tensor(range(self.embed_dim), dtype=torch.float32)
       scaling_factors = 1 / (10000 ** (2 * (positions // 2) / self.embed_dim))
       frequency_features = normalized_freq.unsqueeze(1) * scaling_factors  # Shape: (batch_size, embed_dim)


       # Apply sine to even indices and cosine to odd indices
       encoded_freq = frequencies.new_zeros(batch_size, self.embed_dim, dtype=torch.get_default_dtype())
       encoded_freq[:, 0::2] = torch.sin(frequency_features[:, 0::2])  # Sine for even indices
       encoded_freq[:, 1::2] = torch.cos(frequency_features[:, 1::2])  # Cosine for odd indices


       return encoded_freq


def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0):
   """
   grid_size: int of the grid height and width
   return:
   pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
   """
   grid_h = np.arange(grid_size[0], dtype=np.float32)
   grid_w = np.arange(grid_size[1], dtype=np.float32)
   grid = np.meshgrid(grid_w, grid_h)  # here w goes first
   grid = np.stack(grid, axis=0)


   grid = grid.reshape([2, 1, grid_size[0], grid_size[1]])
   pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
   if cls_token and extra_tokens > 0:
       pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0)
   return pos_embed


def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
   assert embed_dim % 2 == 0


   # use half of dimensions to encode grid_h
   emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
   emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)


   emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
   return emb


def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
   """
   embed_dim: output dimension for each position
   pos: a list of positions to be encoded: size (M,)
   out: (M, D)
   """
   assert embed_dim % 2 == 0
   omega = np.arange(embed_dim // 2, dtype=np.float64)
   omega /= embed_dim / 2.
   omega = 1. / 10000**omega  # (D/2,)


   pos = pos.reshape(-1)  # (M,)
   out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product


   emb_sin = np.sin(out) # (M, D/2)
   emb_cos = np.cos(out) # (M, D/2)


   emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
   return emb


class TimestepEmbedder(nn.Module):
   """
   Embeds scalar timesteps into vector representations.
   """
   def __init__(self, hidden_size, frequency_embedding_size=256, max_period=10000):
       super().__init__()
       self.mlp = nn.Sequential(
           nn.Linear(frequency_embedding_size, hidden_size, bias=True),
           nn.SiLU(),
           nn.Linear(hidden_size, hidden_size, bias=True),
       )
       self.frequency_embedding_size = frequency_embedding_size
       # https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
       half = frequency_embedding_size // 2
       freqs = torch.exp(
           -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
       )
       # Registered as a buffer (not recomputed per call) so it moves with the module's
       # device/dtype casts. Under torch.compile, reading `t.device` here instead used to
       # crash dynamo (ConstantVariable can't wrap torch.device on some builds); reusing the
       # buffer's own device sidesteps that op entirely.
       self.register_buffer("_freqs", freqs, persistent=False)

   def timestep_embedding(self, t, dim):
       """
       Create sinusoidal timestep embeddings.
       :param t: a 1-D Tensor of N indices, one per batch element.
                         These may be fractional.
       :param dim: the dimension of the output.
       :return: an (N, D) Tensor of positional embeddings.
       """
       args = t[:, None].float() * self._freqs[None]
       embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
       if dim % 2:
           embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
       return embedding


   def forward(self, t):
       t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
       t_emb = self.mlp(t_freq)
       return t_emb


class STBlock(nn.Module):
   # Used for temporal compression in context
   def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, dropout_rate=0.0,
                norm_layer=nn.LayerNorm,
                mlp_block='mlp',
                act_layer=lambda: nn.GELU(approximate="tanh"),
                **block_kwargs):
        super().__init__()
        mlp_hidden_dim = int(hidden_size * mlp_ratio)
        self.norm1 = norm_layer(hidden_size, elementwise_affine=False, eps=1e-6)
        self.norm2 = norm_layer(hidden_size, elementwise_affine=False, eps=1e-6)
        self.norm3 = norm_layer(hidden_size, elementwise_affine=False, eps=1e-6)
        self.norm4 = norm_layer(hidden_size, elementwise_affine=False, eps=1e-6)
        self.space_attn = Attention(hidden_size, num_heads=num_heads, qkv_bias=True, qk_norm=True,
                                    attn_drop=dropout_rate, proj_drop=dropout_rate, norm_layer=norm_layer, **block_kwargs)
        self.time_attn = Attention(hidden_size, num_heads=num_heads, qkv_bias=True, qk_norm=True,
                                    attn_drop=dropout_rate, proj_drop=dropout_rate, norm_layer=norm_layer, **block_kwargs)
        if mlp_block == 'mlp':
            self.space_mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=act_layer, norm_layer=norm_layer, drop=0)
            self.time_mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=act_layer, norm_layer=norm_layer, drop=0)
        elif mlp_block == 'swiglu':
            self.space_mlp = SwiGLU(in_features=hidden_size, hidden_features=(mlp_hidden_dim*2)//3)
            self.time_mlp = SwiGLU(in_features=hidden_size, hidden_features=(mlp_hidden_dim*2)//3)
        else:
            raise NotImplementedError(f"mlp_block {mlp_block} not implemented")

   def forward(self, x):
       B, F, N, D = x.shape
       x = rearrange(x, 'b f n d -> (b f) n d')
       x = x + self.space_attn(self.norm1(x))
       x = x + self.space_mlp(self.norm2(x))
       x = rearrange(x, '(b f) n d -> (b n) f d', b=B, f=F, n=N)
       x = x + self.time_attn(self.norm3(x))
       x = x + self.time_mlp(self.norm4(x))
       x = rearrange(x, '(b n) f d -> b f n d', b=B, n=N, f=F)
       return x


class DiTBlock(nn.Module):
   """
   A DiT block with adaptive layer norm zero (adaLN-Zero) conditioning.
   """
   def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, dropout_rate=0.0, norm_layer=nn.LayerNorm, mlp_block='mlp', **block_kwargs):
        super().__init__()
        if isinstance(norm_layer, str):
            norm_layer = get_norm_layer(norm_layer)
        self.norm1 = norm_layer(hidden_size, elementwise_affine=False, eps=1e-6)
        self.attn = Attention(hidden_size, num_heads=num_heads, qkv_bias=True, qk_norm=True, norm_layer=norm_layer, attn_drop=dropout_rate, proj_drop=dropout_rate, **block_kwargs)
        self.norm2 = norm_layer(hidden_size, elementwise_affine=False, eps=1e-6)
        mlp_hidden_dim = int(hidden_size * mlp_ratio)
        if mlp_block == 'mlp':
            approx_gelu = lambda: nn.GELU(approximate="tanh")
            self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, norm_layer=norm_layer, drop=0)
        elif mlp_block == 'swiglu':
            self.mlp = SwiGLU(in_features=hidden_size, hidden_features=(mlp_hidden_dim*2)//3, bias=True)
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, 6 * hidden_size, bias=True)
        )

   def initialize_adaln_weights(self, gate_init_std=0.0):
       nn.init.constant_(self.adaLN_modulation[-1].weight, 0)
       nn.init.constant_(self.adaLN_modulation[-1].bias, 0)
       if gate_init_std != 0.0:
           hidden_size = self.adaLN_modulation[-1].out_features // 6
           for gate_idx in (2, 5):
               start = gate_idx * hidden_size
               end = start + hidden_size
               nn.init.normal_(self.adaLN_modulation[-1].weight[start:end], std=gate_init_std)

   def forward(self, x, c):
       shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(6, dim=1)
       x = x + gate_msa.unsqueeze(1) * self.attn(modulate(self.norm1(x), shift_msa, scale_msa))
       x = x + gate_mlp.unsqueeze(1) * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp))
       return x


class CDiTBlock(nn.Module):
   """
   A DiT block with cross-attention conditioning.
   """
   def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, norm_layer=nn.LayerNorm, mlp_block='mlp', **block_kwargs):
        super().__init__()
        self.norm1 = norm_layer(hidden_size, elementwise_affine=False, eps=1e-6)
        self.attn = Attention(hidden_size, num_heads=num_heads, qkv_bias=True, norm_layer=norm_layer, **block_kwargs)
        self.norm2 = norm_layer(hidden_size, elementwise_affine=False, eps=1e-6)
        self.norm_cond = norm_layer(hidden_size, elementwise_affine=False, eps=1e-6)
        self.cttn = nn.MultiheadAttention(hidden_size, num_heads=num_heads, add_bias_kv=True, bias=True, batch_first=True, **block_kwargs)
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, 11 * hidden_size, bias=True)
        )
        self.norm3 = norm_layer(hidden_size, elementwise_affine=False, eps=1e-6)
        mlp_hidden_dim = int(hidden_size * mlp_ratio)
        if mlp_block == 'mlp':
            approx_gelu = lambda: nn.GELU(approximate="tanh")
            self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, norm_layer=norm_layer, drop=0)
        elif mlp_block == 'swiglu':
            self.mlp = SwiGLU(in_features=hidden_size, hidden_features=(mlp_hidden_dim*2)//3, bias=True)

   def initialize_adaln_weights(self, gate_init_std=0.0):
       nn.init.constant_(self.adaLN_modulation[-1].weight, 0)
       nn.init.constant_(self.adaLN_modulation[-1].bias, 0)
       if gate_init_std != 0.0:
           hidden_size = self.adaLN_modulation[-1].out_features // 11
           for gate_idx in (2, 7, 10):
               start = gate_idx * hidden_size
               end = start + hidden_size
               nn.init.normal_(self.adaLN_modulation[-1].weight[start:end], std=gate_init_std)

   def forward(self, x, c, x_cond):
       shift_msa, scale_msa, gate_msa, shift_ca_xcond, scale_ca_xcond, shift_ca_x, scale_ca_x, gate_ca_x, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(11, dim=1)
       x = x + gate_msa.unsqueeze(1) * self.attn(modulate(self.norm1(x), shift_msa, scale_msa))
       x_cond_norm = modulate(self.norm_cond(x_cond), shift_ca_xcond, scale_ca_xcond)
       x = x + gate_ca_x.unsqueeze(1) * self.cttn(query=modulate(self.norm2(x), shift_ca_x, scale_ca_x), key=x_cond_norm, value=x_cond_norm, need_weights=False)[0]
       x = x + gate_mlp.unsqueeze(1) * self.mlp(modulate(self.norm3(x), shift_mlp, scale_mlp))
       return x


class STDiTBlock(nn.Module):
    """
    A DiT block with adaptive layer norm zero (adaLN-Zero) conditioning.
    """
    def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, dropout_rate=0.0, causal_time_attn=False, modulate_time_attn=False, norm_layer=nn.LayerNorm, mlp_block='mlp', **block_kwargs):
        super().__init__()
        self.norm1 = norm_layer(hidden_size, elementwise_affine=False, eps=1e-6)
        self.space_attn = Attention(hidden_size, num_heads=num_heads, qkv_bias=True, qk_norm=True, norm_layer=norm_layer, attn_drop=dropout_rate, proj_drop=dropout_rate, **block_kwargs)
        self.time_attn = Attention(hidden_size, num_heads=num_heads, qkv_bias=True, qk_norm=True, norm_layer=norm_layer, attn_drop=dropout_rate, proj_drop=dropout_rate, **block_kwargs)
        self.norm2 = norm_layer(hidden_size, elementwise_affine=False, eps=1e-6)
        self.norm3 = norm_layer(hidden_size, elementwise_affine=False, eps=1e-6)
        self.norm4 = norm_layer(hidden_size, elementwise_affine=False, eps=1e-6)
        mlp_hidden_dim = int(hidden_size * mlp_ratio)

        if mlp_block == 'mlp':
            approx_gelu = lambda: nn.GELU(approximate="tanh")
            self.space_mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, norm_layer=None, drop=0)
            self.time_mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, norm_layer=None, drop=0)
        elif mlp_block == 'swiglu':
            self.space_mlp = SwiGLU(in_features=hidden_size, hidden_features=(mlp_hidden_dim*2)//3)
            self.time_mlp = SwiGLU(in_features=hidden_size, hidden_features=(mlp_hidden_dim*2)//3)
        else:
            raise NotImplementedError(f"mlp_block {mlp_block} not implemented")

        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, 9 * hidden_size, bias=True)
        )
        self.causal_time_attn = causal_time_attn
        self.modulate_time_attn = modulate_time_attn
        self.layer_idx = block_kwargs.get("layer_idx")
        self.log_adaln_mean_abs = False
        self._last_adaln_mean_abs = None

        if modulate_time_attn:
            self.adaLN_time_attn_modulation = nn.Sequential(
                nn.SiLU(),
                nn.Linear(hidden_size, 3 * hidden_size, bias=True)
            )
            self.norm_time_attn = norm_layer(hidden_size, elementwise_affine=False, eps=1e-6)
            # initialize
            nn.init.constant_(self.adaLN_time_attn_modulation[-1].weight, 0)
            nn.init.constant_(self.adaLN_time_attn_modulation[-1].bias, 0)
        else:
            self.norm_time_attn = nn.Identity()

    def initialize_adaln_weights(self, gate_init_std=0.0):
        nn.init.constant_(self.adaLN_modulation[-1].weight, 0)
        nn.init.constant_(self.adaLN_modulation[-1].bias, 0)
        if gate_init_std != 0.0:
            hidden_size = self.adaLN_modulation[-1].out_features // 9
            for gate_idx in (2, 5, 8):
                start = gate_idx * hidden_size
                end = start + hidden_size
                nn.init.normal_(self.adaLN_modulation[-1].weight[start:end], std=gate_init_std)

        if hasattr(self, "adaLN_time_attn_modulation"):
            nn.init.constant_(self.adaLN_time_attn_modulation[-1].weight, 0)
            nn.init.constant_(self.adaLN_time_attn_modulation[-1].bias, 0)
            if gate_init_std != 0.0:
                hidden_size = self.adaLN_time_attn_modulation[-1].out_features // 3
                start = 2 * hidden_size
                end = start + hidden_size
                nn.init.normal_(self.adaLN_time_attn_modulation[-1].weight[start:end], std=gate_init_std)

    def _collect_adaln_mean_abs(self, **tensors):
        self._last_adaln_mean_abs = {
            name: tensor.detach().abs().flatten(1).mean(dim=1)
            for name, tensor in tensors.items()
        }

    def forward(self, x, c):
        B, F, N, D = x.shape

        # chunk into 9 [B, C] vectors
        (shift_msa, scale_msa, gate_msa,
        shift_mlp_s, scale_mlp_s, gate_mlp_s,
        shift_mlp_t, scale_mlp_t, gate_mlp_t) = self.adaLN_modulation(c).chunk(9, dim=-1)

        x_modulated = modulate(self.norm1(x), shift_msa, scale_msa)
        x_modulated = rearrange(x_modulated, 'b f n d -> (b f) n d', b=B, f=F)
        x_ = self.space_attn(x_modulated)
        x_ = rearrange(x_, '(b f) n d -> b f n d', b=B, f=F)
        x = x + _broadcast_gate(gate_msa, x) * x_

        x_modulated = modulate(self.norm2(x), shift_mlp_s, scale_mlp_s)
        x = x + _broadcast_gate(gate_mlp_s, x) * self.space_mlp(x_modulated)

        # β€” temporal attention path β€”
        if self.modulate_time_attn:
            shift_mta, scale_mta, gate_mta = self.adaLN_time_attn_modulation(c).chunk(3, dim=-1)
        else:
            shift_mta, scale_mta, gate_mta = torch.zeros_like(shift_mlp_t), torch.zeros_like(scale_mlp_t), torch.ones_like(gate_mlp_t)

        if self.log_adaln_mean_abs:
            self._collect_adaln_mean_abs(
                msa_shift=shift_msa,
                msa_scale=scale_msa,
                msa_gate=gate_msa,
                mlp_s_shift=shift_mlp_s,
                mlp_s_scale=scale_mlp_s,
                mlp_s_gate=gate_mlp_s,
                mta_shift=shift_mta,
                mta_scale=scale_mta,
                mta_gate=gate_mta,
                mlp_t_shift=shift_mlp_t,
                mlp_t_scale=scale_mlp_t,
                mlp_t_gate=gate_mlp_t,
            )

        x_modulated = modulate(self.norm_time_attn(x), shift_mta, scale_mta)
        x_modulated = rearrange(x_modulated, 'b f n d -> (b n) f d', b=B, f=F, n=N)
        # x supplies the device via new_ones instead of a bare `.device` read, which crashes
        # Dynamo on some torch builds (see TimestepEmbedder above for the same issue).
        time_attn_mask = torch.tril(x.new_ones(F, F, dtype=torch.get_default_dtype())) if self.causal_time_attn else None
        x_ = self.time_attn(x_modulated, attn_mask=time_attn_mask)
        x_ = rearrange(x_, '(b n) f d -> b f n d', b=B, n=N, f=F)
        x = x + _broadcast_gate(gate_mta, x) * x_

        x_modulated = modulate(self.norm3(x), shift_mlp_t, scale_mlp_t)
        x = x + _broadcast_gate(gate_mlp_t, x) * self.time_mlp(x_modulated)

        return x


class FinalLayer(nn.Module):
   """
   The final layer of DiT.
   """
   def __init__(self, hidden_size, patch_size, out_channels, norm_layer=nn.LayerNorm, act_layer=nn.SiLU):
       super().__init__()
       self.norm_final = norm_layer(hidden_size, elementwise_affine=False, eps=1e-6)
       self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
       self.adaLN_modulation = nn.Sequential(
           act_layer(),
           nn.Linear(hidden_size, 2 * hidden_size, bias=True)
       )


   def forward(self, x, c):
       shift, scale = self.adaLN_modulation(c).chunk(2, dim=-1)
       x = modulate(self.norm_final(x), shift, scale)
       x = self.linear(x)
       return x


class STDiTBlockWithSpatialL2(STDiTBlock):
    """
    STDiTBlock augmented with per-token gated spatial conditioning from L2 tokens.

    z_spatial (B, F, N, D) β†’ (shift_s, scale_s, gate_s) per token via a small MLP.
    Applied as an additive gated adaLN before the standard block operations:
        x = x + gate_s * modulate(spatial_norm(x), shift_s, scale_s)

    Zero-init of spatial_to_mod guarantees a no-op at the start of training,
    so pre-trained STDiTDF weights can be fine-tuned without disruption.
    """

    def __init__(self, hidden_size, num_heads, norm_layer=nn.LayerNorm, **kwargs):
        super().__init__(hidden_size, num_heads, norm_layer=norm_layer, **kwargs)
        self.spatial_norm = norm_layer(hidden_size, elementwise_affine=False, eps=1e-6)
        self.spatial_to_mod = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, 3 * hidden_size, bias=True),
        )
        # Small weight init so shift/scale start as genuine L2-dependent perturbations.
        # Gate bias initialised to 2.0 β†’ sigmoid(2.0) β‰ˆ 0.88, so the gate starts open
        # and the model must learn to close it rather than having to learn to open it.
        nn.init.normal_(self.spatial_to_mod[-1].weight, std=0.02)
        nn.init.zeros_(self.spatial_to_mod[-1].bias)
        self.spatial_to_mod[-1].bias.data[2 * hidden_size:].fill_(2.0)

    def forward(self, x, c, z_spatial):
        # z_spatial: (B, F, N, D) β€” pre-computed L2 spatial tokens (patchified + position)
        shift_s, scale_s, gate_s = self.spatial_to_mod(z_spatial).chunk(3, dim=-1)
        x = x + torch.sigmoid(gate_s) * modulate(self.spatial_norm(x), shift_s, scale_s)
        return super().forward(x, c)