File size: 20,561 Bytes
2685e1e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Nucleus-Image MoE DiT in MLX.

All dimensions verified against actual safetensors weight shapes.
"""

import math
from typing import Tuple

import mlx.core as mx
import mlx.nn as nn


# ── RoPE ──

def rope(pos, dim: int, theta: int = 10000):
    scale = mx.arange(0, dim, 2).astype(mx.float32) / dim
    omega = 1.0 / (theta ** scale)
    out = mx.expand_dims(pos.astype(mx.float32), -1) * mx.expand_dims(omega, 0)
    return out


def build_rope_freqs(axes_dim, theta, max_len=4096):
    """Precompute RoPE cos/sin for positive AND negative positions.

    Matches NucleusMoEEmbedRope: pos_freqs for [0..4095], neg_freqs for [-4096..-1].
    """
    pos_cos, pos_sin = [], []
    neg_cos, neg_sin = [], []
    for dim in axes_dim:
        freqs = 1.0 / (theta ** (mx.arange(0, dim, 2).astype(mx.float32) / dim))
        # Positive positions: 0, 1, ..., max_len-1
        t_pos = mx.arange(max_len).astype(mx.float32)
        angles_pos = mx.outer(t_pos, freqs)
        pos_cos.append(mx.cos(angles_pos))
        pos_sin.append(mx.sin(angles_pos))
        # Negative positions: -max_len, ..., -2, -1
        t_neg = (mx.arange(max_len).astype(mx.float32)[::-1] * -1 - 1)
        angles_neg = mx.outer(t_neg, freqs)
        neg_cos.append(mx.cos(angles_neg))
        neg_sin.append(mx.sin(angles_neg))
    return pos_cos, pos_sin, neg_cos, neg_sin


def compute_image_rope(height, width, axes_dim, pos_cos, pos_sin, neg_cos, neg_sin, scale_rope=True):
    """Compute RoPE frequencies for image patches.

    With scale_rope=True (default for Nucleus), uses centered positions:
      height: [-ceil(H/2), ..., -1, 0, 1, ..., floor(H/2)-1]
      width:  [-ceil(W/2), ..., -1, 0, 1, ..., floor(W/2)-1]

    Returns (img_cos, img_sin) each of shape [H*W, D//2]
    """
    frame = 1
    # Frame axis: position 0
    f_cos = mx.broadcast_to(pos_cos[0][:frame], (frame, 1, 1, axes_dim[0] // 2))
    f_sin = mx.broadcast_to(pos_sin[0][:frame], (frame, 1, 1, axes_dim[0] // 2))
    f_cos = mx.broadcast_to(f_cos, (frame, height, width, axes_dim[0] // 2))
    f_sin = mx.broadcast_to(f_sin, (frame, height, width, axes_dim[0] // 2))

    if scale_rope:
        # Centered: negative positions + positive positions
        # Height: neg[-N_neg:] + pos[:N_pos] where N_neg = H - H//2, N_pos = H//2
        n_neg_h = height - height // 2
        n_pos_h = height // 2
        h_cos = mx.concatenate([neg_cos[1][-n_neg_h:], pos_cos[1][:n_pos_h]], axis=0)
        h_sin = mx.concatenate([neg_sin[1][-n_neg_h:], pos_sin[1][:n_pos_h]], axis=0)
        # Width
        n_neg_w = width - width // 2
        n_pos_w = width // 2
        w_cos = mx.concatenate([neg_cos[2][-n_neg_w:], pos_cos[2][:n_pos_w]], axis=0)
        w_sin = mx.concatenate([neg_sin[2][-n_neg_w:], pos_sin[2][:n_pos_w]], axis=0)
    else:
        h_cos = pos_cos[1][:height]
        h_sin = pos_sin[1][:height]
        w_cos = pos_cos[2][:width]
        w_sin = pos_sin[2][:width]

    h_cos = mx.broadcast_to(h_cos.reshape(1, height, 1, -1), (frame, height, width, axes_dim[1] // 2))
    h_sin = mx.broadcast_to(h_sin.reshape(1, height, 1, -1), (frame, height, width, axes_dim[1] // 2))
    w_cos = mx.broadcast_to(w_cos.reshape(1, 1, width, -1), (frame, height, width, axes_dim[2] // 2))
    w_sin = mx.broadcast_to(w_sin.reshape(1, 1, width, -1), (frame, height, width, axes_dim[2] // 2))

    img_cos = mx.concatenate([f_cos, h_cos, w_cos], axis=-1).reshape(frame * height * width, -1)
    img_sin = mx.concatenate([f_sin, h_sin, w_sin], axis=-1).reshape(frame * height * width, -1)
    return img_cos, img_sin


def compute_text_rope(max_txt_len, max_vid_index, axes_dim, pos_cos, pos_sin):
    """Compute RoPE frequencies for text tokens.

    Text positions start after max_vid_index (= max(H/2, W/2) with scale_rope).
    """
    start = max_vid_index
    cos_parts = []
    sin_parts = []
    for i, dim in enumerate(axes_dim):
        cos_parts.append(pos_cos[i][start:start + max_txt_len])
        sin_parts.append(pos_sin[i][start:start + max_txt_len])
    txt_cos = mx.concatenate(cos_parts, axis=-1)
    txt_sin = mx.concatenate(sin_parts, axis=-1)
    return txt_cos, txt_sin


def apply_rotary_emb_complex(x, freqs_cos, freqs_sin):
    """Apply rotary embeddings using complex multiplication.

    x: [B, S, H, D]
    freqs_cos, freqs_sin: [S, D//2] (real and imaginary parts of complex exponentials)

    Complex multiplication: (a+bi)(c+di) = (ac-bd) + (ad+bc)i
    """
    # Reshape x into pairs: [B, S, H, D//2, 2]
    x_pairs = x.reshape(*x.shape[:-1], -1, 2)
    x_real = x_pairs[..., 0]  # [B, S, H, D//2]
    x_imag = x_pairs[..., 1]

    # Broadcast freqs: [1, S, 1, D//2]
    cos_ = freqs_cos[None, :, None, :].astype(x.dtype)
    sin_ = freqs_sin[None, :, None, :].astype(x.dtype)

    # Complex multiply
    out_real = x_real * cos_ - x_imag * sin_
    out_imag = x_real * sin_ + x_imag * cos_

    # Interleave back: [B, S, H, D//2, 2] β†’ [B, S, H, D]
    out = mx.stack([out_real, out_imag], axis=-1)
    return out.reshape(x.shape)


def timestep_embedding(t, dim: int):
    half = dim // 2
    freqs = mx.exp(-math.log(10000) * mx.arange(0, half).astype(mx.float32) / half)
    args = t.astype(mx.float32)[:, None] * freqs[None, :]
    return mx.concatenate([mx.cos(args), mx.sin(args)], axis=-1)


# ── TimestepEmbedder: linear_1 [8192, 2048], linear_2 [2048, 8192] ──

class TimestepEmbedder(nn.Module):
    def __init__(self, in_dim: int, expand_dim: int):
        super().__init__()
        self.linear_1 = nn.Linear(in_dim, expand_dim)
        self.linear_2 = nn.Linear(expand_dim, in_dim)

    def __call__(self, t):
        return self.linear_2(nn.silu(self.linear_1(t)))


# ── Gated projection (shared between dense FFN and shared expert) ──

class GatedProj(nn.Module):
    """net.0.proj: [hidden*2, in_dim]. SwiGLU: first_half * silu(second_half)."""
    def __init__(self, in_dim: int, hidden_dim: int):
        super().__init__()
        self.proj = nn.Linear(in_dim, hidden_dim * 2, bias=False)

    def __call__(self, x):
        hidden, gate = mx.split(self.proj(x), 2, axis=-1)
        return hidden * nn.silu(gate)


# ── Dense FFN (layers 0-2) ──
# net.0.proj: [10752, 2048] β†’ gated, hidden=5376
# net.2: [2048, 5376]

class DenseFFN(nn.Module):
    def __init__(self, in_dim: int, hidden_dim: int):
        super().__init__()
        self.net = [GatedProj(in_dim, hidden_dim), None, nn.Linear(hidden_dim, in_dim, bias=False)]

    def __call__(self, x):
        return self.net[2](self.net[0](x))


# ── MoE FFN (layers 3-31) ──

class SharedExpert(nn.Module):
    """shared_expert.net.0.proj: [2688, 2048], net.2: [2048, 1344]"""
    def __init__(self, in_dim: int, hidden_dim: int):
        super().__init__()
        self.net = [GatedProj(in_dim, hidden_dim), None, nn.Linear(hidden_dim, in_dim, bias=False)]

    def __call__(self, x):
        return self.net[2](self.net[0](x))


class Experts(nn.Module):
    """Packed expert weights.
    gate_up_proj: [64, in_dim, hidden*2] (weight shape from safetensors)
    down_proj: [64, hidden, in_dim]
    """
    def __init__(self, in_dim: int, hidden_dim: int, num_experts: int):
        super().__init__()
        self.gate_up_proj = mx.zeros((num_experts, in_dim, hidden_dim * 2))
        self.down_proj = mx.zeros((num_experts, hidden_dim, in_dim))


class MoEFFN(nn.Module):
    """gate: [64, 4096] β€” router input is concat(token, timestep)
    experts: packed, shared_expert: standard FFN

    Uses expert-choice routing: each expert picks its top-C tokens
    (capacity-based), matching the diffusers NucleusMoELayer.
    """

    def __init__(self, in_dim: int, expert_hidden_dim: int, num_experts: int = 64,
                 route_scale: float = 2.5, capacity_factor: float = 2.0):
        super().__init__()
        self.num_experts = num_experts
        self.route_scale = route_scale
        self.capacity_factor = capacity_factor

        # Router input is 2*in_dim (token + timestep concat)
        self.gate = nn.Linear(in_dim * 2, num_experts, bias=False)
        self.experts = Experts(in_dim, expert_hidden_dim, num_experts)
        self.shared_expert = SharedExpert(in_dim, expert_hidden_dim)

    def __call__(self, x, timestep_emb=None, unmodulated_x=None):
        """Expert-choice routing: each expert picks top-C tokens.

        C = ceil(capacity_factor * S / num_experts)
        Per-token gating normalization ensures tokens selected by
        multiple experts have their weights sum to route_scale.
        """
        B, S, D = x.shape

        # Decoupled routing: router sees unmodulated tokens + timestep
        if unmodulated_x is None:
            unmodulated_x = x
        if timestep_emb is not None:
            t_expanded = mx.broadcast_to(timestep_emb[:, None, :], (B, S, D))
            router_input = mx.concatenate([t_expanded, unmodulated_x], axis=-1)
        else:
            router_input = unmodulated_x

        logits = self.gate(router_input)  # [B, S, E]
        scores = mx.softmax(logits.astype(mx.float32), axis=-1).astype(x.dtype)

        # Expert-choice: transpose to [B, E, S], each expert scores all tokens
        affinity = mx.transpose(scores, (0, 2, 1))  # [B, E, S]
        capacity = max(1, math.ceil(self.capacity_factor * S / self.num_experts))

        # B=1 for inference β€” squeeze batch dim for simpler indexing
        aff = affinity[0]     # [E, S]
        x_flat = x[0]         # [S, D]

        # Each expert picks top-C tokens (argsort descending)
        sorted_idx = mx.argsort(-aff, axis=-1)[:, :capacity]  # [E, C]
        top_scores = mx.take_along_axis(aff, sorted_idx, axis=-1)  # [E, C]

        # Per-token normalization (tokens picked by multiple experts
        # have their scores normalized to sum to route_scale)
        flat_idx = sorted_idx.reshape(-1)       # [E*C]
        flat_scores = top_scores.reshape(-1)    # [E*C]

        # Indicator matrix for scatter operations: [E*C, S]
        indicator = (flat_idx[:, None] == mx.arange(S)[None, :]).astype(x.dtype)

        # Per-token score sums
        token_sums = indicator.T @ flat_scores   # [S]

        # Normalized gating weights
        flat_norm = flat_scores / (token_sums[flat_idx] + 1e-12) * self.route_scale

        # Gather selected tokens for all experts
        gathered = x_flat[flat_idx]  # [E*C, D]

        # Process through each expert's weights
        gu = self.experts.gate_up_proj  # [E, D, hidden*2]
        down = self.experts.down_proj   # [E, hidden, D]

        expert_outputs = []
        for e in range(self.num_experts):
            start = e * capacity
            end = start + capacity
            tokens_e = gathered[start:end]  # [C, D]

            h = tokens_e @ gu[e]  # [C, hidden*2]
            gh, uh = mx.split(h, 2, axis=-1)  # SwiGLUExperts: first=gate, second=up
            h = nn.silu(gh) * uh
            h = h @ down[e]  # [C, D]
            expert_outputs.append(h)

        all_expert_out = mx.concatenate(expert_outputs, axis=0)  # [E*C, D]
        all_expert_out = all_expert_out * flat_norm[:, None]       # weighted

        # Scatter back: indicator.T is [S, E*C], matmul gives [S, D]
        scattered = indicator.T @ all_expert_out

        # Shared expert (always runs on all tokens)
        shared_out = self.shared_expert(x)[0]  # [S, D]

        return (shared_out + scattered).reshape(1, S, D)


# ── Joint Attention ──

class JointAttention(nn.Module):
    def __init__(self, hidden_dim: int, num_heads: int, num_kv_heads: int, head_dim: int):
        super().__init__()
        self.num_heads = num_heads
        self.num_kv_heads = num_kv_heads
        self.head_dim = head_dim

        self.to_q = nn.Linear(hidden_dim, num_heads * head_dim, bias=False)
        self.to_k = nn.Linear(hidden_dim, num_kv_heads * head_dim, bias=False)
        self.to_v = nn.Linear(hidden_dim, num_kv_heads * head_dim, bias=False)
        self.to_out = [nn.Linear(num_heads * head_dim, hidden_dim, bias=False)]

        self.add_k_proj = nn.Linear(hidden_dim, num_kv_heads * head_dim, bias=False)
        self.add_v_proj = nn.Linear(hidden_dim, num_kv_heads * head_dim, bias=False)

        self.norm_q = nn.RMSNorm(head_dim)
        self.norm_k = nn.RMSNorm(head_dim)
        self.norm_added_q = nn.RMSNorm(head_dim)
        self.norm_added_k = nn.RMSNorm(head_dim)

    def __call__(self, img_x, txt_kv, img_rope=None, txt_rope=None):
        B, S_img, _ = img_x.shape
        S_txt = txt_kv.shape[1]

        q = self.to_q(img_x).reshape(B, S_img, self.num_heads, self.head_dim)
        k = self.to_k(img_x).reshape(B, S_img, self.num_kv_heads, self.head_dim)
        v = self.to_v(img_x).reshape(B, S_img, self.num_kv_heads, self.head_dim)

        q = self.norm_q(q)
        k = self.norm_k(k)

        txt_k = self.add_k_proj(txt_kv).reshape(B, S_txt, self.num_kv_heads, self.head_dim)
        txt_v = self.add_v_proj(txt_kv).reshape(B, S_txt, self.num_kv_heads, self.head_dim)
        txt_k = self.norm_added_k(txt_k)

        # Apply RoPE
        if img_rope is not None:
            img_cos, img_sin = img_rope
            q = apply_rotary_emb_complex(q, img_cos, img_sin)
            k = apply_rotary_emb_complex(k, img_cos, img_sin)
        if txt_rope is not None:
            txt_cos, txt_sin = txt_rope
            txt_k = apply_rotary_emb_complex(txt_k, txt_cos, txt_sin)

        k = mx.concatenate([k, txt_k], axis=1)
        v = mx.concatenate([v, txt_v], axis=1)

        if self.num_kv_heads < self.num_heads:
            r = self.num_heads // self.num_kv_heads
            k = mx.repeat(k, r, axis=2)
            v = mx.repeat(v, r, axis=2)

        q = q.transpose(0, 2, 1, 3)
        k = k.transpose(0, 2, 1, 3)
        v = v.transpose(0, 2, 1, 3)

        out = mx.fast.scaled_dot_product_attention(q, k, v, scale=1.0 / math.sqrt(self.head_dim))
        out = out.transpose(0, 2, 1, 3).reshape(B, S_img, -1)
        return self.to_out[0](out)


# ── Transformer Block ──
# img_mod.1: [8192, 2048] β†’ 4 modulations (shift_attn, scale_attn, shift_mlp, scale_mlp)

class TransformerBlock(nn.Module):
    def __init__(self, hidden_dim: int, num_heads: int, num_kv_heads: int,
                 head_dim: int, dense_hidden: int, is_moe: bool,
                 num_experts: int, expert_hidden: int, route_scale: float,
                 text_dim: int, capacity_factor: float = 2.0):
        super().__init__()
        self.is_moe = is_moe

        self.attn = JointAttention(hidden_dim, num_heads, num_kv_heads, head_dim)

        if is_moe:
            self.is_moe = True
            self.img_mlp = MoEFFN(hidden_dim, expert_hidden, num_experts, route_scale, capacity_factor)
        else:
            self.is_moe = False
            self.img_mlp = DenseFFN(hidden_dim, dense_hidden)

        # Pre-norms: LayerNorm without affine
        self.pre_attn_norm = nn.LayerNorm(hidden_dim, affine=False)
        self.pre_mlp_norm = nn.LayerNorm(hidden_dim, affine=False)

        # 4 modulations: scale1, gate1, scale2, gate2
        self.img_mod = [nn.SiLU(), nn.Linear(hidden_dim, hidden_dim * 4)]
        self.encoder_proj = nn.Linear(text_dim, hidden_dim)

    def __call__(self, img_x, txt_kv, c, img_rope=None, txt_rope=None):
        # 4 modulations: scale1, gate1, scale2, gate2
        mod = self.img_mod[1](self.img_mod[0](c))[:, None, :]  # [B, 1, 4*H]
        scale1, gate1, scale2, gate2 = mx.split(mod, 4, axis=-1)

        # Clamp gates
        gate1 = mx.clip(gate1, -2.0, 2.0)
        gate2 = mx.clip(gate2, -2.0, 2.0)

        txt_projected = self.encoder_proj(txt_kv)

        # Attention: LayerNorm β†’ scale β†’ attn β†’ tanh gate
        img_normed = self.pre_attn_norm(img_x)
        img_modulated = img_normed * (1 + scale1)
        attn_out = self.attn(img_modulated, txt_projected, img_rope=img_rope, txt_rope=txt_rope)
        img_x = img_x + mx.tanh(gate1) * attn_out

        # FFN: LayerNorm β†’ scale β†’ mlp β†’ tanh gate
        img_normed2 = self.pre_mlp_norm(img_x)
        img_modulated2 = img_normed2 * (1 + scale2)
        if self.is_moe:
            mlp_out = self.img_mlp(img_modulated2, timestep_emb=c, unmodulated_x=img_normed2)
        else:
            mlp_out = self.img_mlp(img_modulated2)
        img_x = img_x + mx.tanh(gate2) * mlp_out

        return img_x


# ── Full Model ──

class NucleusMoEDiT(nn.Module):
    def __init__(self, config: dict):
        super().__init__()
        num_heads = config.get("num_attention_heads", 16)
        head_dim = config.get("attention_head_dim", 128)
        hidden = num_heads * head_dim  # 2048
        num_kv_heads = config.get("num_key_value_heads", 4)
        num_layers = config.get("num_layers", 32)
        num_experts = config.get("num_experts", 64)
        expert_hidden = config.get("moe_intermediate_dim", 1344)
        route_scale = config.get("route_scale", 2.5)
        text_dim = config.get("joint_attention_dim", 4096)
        in_channels = config.get("in_channels", 64)
        out_channels = config.get("out_channels", 16)
        axes_dims = config.get("axes_dims_rope", [16, 56, 56])

        # Dense FFN hidden: from weight [10752, 2048] β†’ 10752/2 = 5376
        dense_hidden = 5376

        self.hidden_dim = hidden
        self.in_channels = in_channels
        self.out_channels = out_channels

        self.img_in = nn.Linear(in_channels, hidden)

        # Timestep: linear_1 [8192, 2048], linear_2 [2048, 8192]
        self.time_text_embed = {
            "timestep_embedder": TimestepEmbedder(hidden, 8192),
            "norm": nn.RMSNorm(hidden),
        }

        self.txt_norm = nn.RMSNorm(text_dim)
        self._axes_dim = axes_dims
        self._scale_rope = True  # Nucleus uses centered positions
        self._pos_cos, self._pos_sin, self._neg_cos, self._neg_sin = build_rope_freqs(axes_dims, 10000)

        capacity_factors = config.get("capacity_factors", [0.0] * 3 + [2.0] * 29)

        dense_layers = {0, 1, 2}
        self.transformer_blocks = [
            TransformerBlock(
                hidden, num_heads, num_kv_heads, head_dim, dense_hidden,
                is_moe=(i not in dense_layers),
                num_experts=num_experts, expert_hidden=expert_hidden,
                route_scale=route_scale, text_dim=text_dim,
                capacity_factor=capacity_factors[i],
            )
            for i in range(num_layers)
        ]

        # norm_out: AdaLayerNormContinuous(elementwise_affine=False, eps=1e-6)
        # Applies LayerNorm before adaptive modulation
        self.norm_out = {
            "norm": nn.LayerNorm(hidden, affine=False, eps=1e-6),
            "linear": nn.Linear(hidden, hidden * 2),
        }
        # proj_out: [64, 2048] β†’ 64 = patch_sizeΒ² * out_channels
        self.proj_out = nn.Linear(hidden, in_channels, bias=False)

    def __call__(self, hidden_states, timestep, txt_kv):
        B = hidden_states.shape[0]

        x = self.img_in(hidden_states)

        # Timesteps(scale=1000): multiply by 1000 then sinusoidal embed
        # Input timestep is already sigma*1000 from the scheduler
        t_emb = timestep_embedding(timestep * 1000, self.hidden_dim).astype(x.dtype)
        c = self.time_text_embed["timestep_embedder"](t_emb)
        c = self.time_text_embed["norm"](c)

        txt_kv = self.txt_norm(txt_kv)

        # Build RoPE: image patches are on a grid, text follows after
        N_img = hidden_states.shape[1]
        grid_h = int(N_img ** 0.5)
        grid_w = N_img // grid_h
        img_cos, img_sin = compute_image_rope(
            grid_h, grid_w, self._axes_dim,
            self._pos_cos, self._pos_sin, self._neg_cos, self._neg_sin,
            scale_rope=self._scale_rope,
        )

        T_txt = txt_kv.shape[1]
        # With scale_rope, text starts at max(H/2, W/2)
        max_vid_idx = max(grid_h // 2, grid_w // 2) if self._scale_rope else max(grid_h, grid_w)
        txt_cos, txt_sin = compute_text_rope(T_txt, max_vid_idx, self._axes_dim, self._pos_cos, self._pos_sin)

        img_rope = (img_cos, img_sin)
        txt_rope = (txt_cos, txt_sin)

        for block in self.transformer_blocks:
            x = block(x, txt_kv, c, img_rope=img_rope, txt_rope=txt_rope)

        # AdaLayerNormContinuous: norm first, then adaptive modulation
        mod = self.norm_out["linear"](nn.silu(c))
        scale, shift = mx.split(mod, 2, axis=-1)  # scale first, shift second
        x = self.norm_out["norm"](x) * (1 + scale[:, None, :]) + shift[:, None, :]
        x = self.proj_out(x)

        return x