File size: 24,782 Bytes
1cf04f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Alexandria β€” Text Reconstruction via Geometric Encoding
=========================================================
Wikipedia β†’ UTF-8 bytes β†’ (3, H, W) β†’ PatchSVAE β†’ reconstruct β†’ bytes β†’ text

The Library of Alexandria, rebuilt in geometry.

Text bytes are a structured subset of noise. Johanna already knows
how to invert the projection for arbitrary byte patterns. Alexandria
fine-tunes that knowledge specifically for text.

Byte accuracy is the metric that matters. A single wrong byte is
a wrong character. Text demands perfection.
"""

import os
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
import time
import numpy as np
from tqdm import tqdm

# ── HuggingFace auth from Colab secrets ──
try:
    from google.colab import userdata
    os.environ["HF_TOKEN"] = userdata.get('HF_TOKEN')
    from huggingface_hub import login
    login(token=os.environ["HF_TOKEN"])
except Exception:
    pass

# ── SVD Backend ──────────────────────────────────────────────────

try:
    from geolip_core.linalg.eigh import FLEigh, _FL_MAX_N
    HAS_FL = True
except ImportError:
    HAS_FL = False


def gram_eigh_svd_fp64(A):
    orig_dtype = A.dtype
    with torch.amp.autocast('cuda', enabled=False):
        A_d = A.double()
        G = torch.bmm(A_d.transpose(1, 2), A_d)
        G.diagonal(dim1=-2, dim2=-1).add_(1e-12)
        eigenvalues, V = torch.linalg.eigh(G)
        eigenvalues = eigenvalues.flip(-1)
        V = V.flip(-1)
        S = torch.sqrt(eigenvalues.clamp(min=1e-24))
        U = torch.bmm(A_d, V) / S.unsqueeze(1).clamp(min=1e-16)
        Vh = V.transpose(-2, -1).contiguous()
    return U.to(orig_dtype), S.to(orig_dtype), Vh.to(orig_dtype)


def svd_fp64(A):
    B, M, N = A.shape
    if HAS_FL and N <= _FL_MAX_N and A.is_cuda:
        orig_dtype = A.dtype
        with torch.amp.autocast('cuda', enabled=False):
            A_d = A.double()
            G = torch.bmm(A_d.transpose(1, 2), A_d)
            eigenvalues, V = FLEigh()(G.float())
            eigenvalues = eigenvalues.double().flip(-1)
            V = V.double().flip(-1)
            S = torch.sqrt(eigenvalues.clamp(min=1e-24))
            U = torch.bmm(A_d, V) / S.unsqueeze(1).clamp(min=1e-16)
            Vh = V.transpose(-2, -1).contiguous()
        return U.to(orig_dtype), S.to(orig_dtype), Vh.to(orig_dtype)
    else:
        return gram_eigh_svd_fp64(A)


# ── CV Monitoring ────────────────────────────────────────────────

def cayley_menger_vol2(points):
    B, N, D = points.shape
    pts = points.double()
    gram = torch.bmm(pts, pts.transpose(1, 2))
    norms = torch.diagonal(gram, dim1=1, dim2=2)
    d2 = F.relu(norms.unsqueeze(2) + norms.unsqueeze(1) - 2 * gram)
    cm = torch.zeros(B, N + 1, N + 1, device=points.device, dtype=torch.float64)
    cm[:, 0, 1:] = 1.0
    cm[:, 1:, 0] = 1.0
    cm[:, 1:, 1:] = d2
    k = N - 1
    sign = (-1.0) ** (k + 1)
    fact = math.factorial(k)
    return sign * torch.linalg.det(cm) / ((2 ** k) * (fact ** 2))


def cv_of(emb, n_samples=200):
    if emb.dim() != 2 or emb.shape[0] < 5:
        return 0.0
    N, D = emb.shape
    pool = min(N, 512)
    indices = torch.stack([torch.randperm(pool, device=emb.device)[:5] for _ in range(n_samples)])
    vol2 = cayley_menger_vol2(emb[:pool][indices])
    valid = vol2 > 1e-20
    if valid.sum() < 10:
        return 0.0
    vols = vol2[valid].sqrt()
    return (vols.std() / (vols.mean() + 1e-8)).item()


# ── Wikipedia Text Dataset ───────────────────────────────────────

class WikiTextAsImage(torch.utils.data.Dataset):
    """Wikipedia text packed as (3, H, W) byte tensors.

    Streams Wikipedia, concatenates into a byte buffer,
    serves random chunks as "images". The model never knows
    it's reading β€” it just sees numbers in a grid.

    Byte normalization: [0, 255] β†’ [-1, 1]
    """
    def __init__(self, size=200000, img_size=128, split='train'):
        self.size = size
        self.img_size = img_size
        self.n_bytes = 3 * img_size * img_size

        print(f"  Loading Wikipedia ({split})...")
        from datasets import load_dataset
        ds = load_dataset('wikipedia', '20220301.en', split=split,
                          streaming=True)

        # Accumulate enough text β€” need at least size * n_bytes
        target_bytes = min(size * self.n_bytes, 500_000_000)  # cap at 500MB
        chunks = []
        total = 0
        for article in ds:
            text = article['text']
            if text.strip():
                chunks.append(text)
                total += len(text)
            if total >= target_bytes:
                break

        self.raw_bytes = '\n'.join(chunks).encode('utf-8')
        print(f"  Corpus: {len(self.raw_bytes):,} bytes ({len(self.raw_bytes)/1024/1024:.1f}MB)")
        print(f"  Samples: {size:,} Γ— {self.n_bytes:,} bytes = {self.n_bytes} bytes/sample")

    def __len__(self):
        return self.size

    def __getitem__(self, idx):
        max_start = max(0, len(self.raw_bytes) - self.n_bytes)
        start = torch.randint(0, max_start + 1, (1,)).item()

        chunk = self.raw_bytes[start:start + self.n_bytes]
        if len(chunk) < self.n_bytes:
            chunk = chunk + b'\x00' * (self.n_bytes - len(chunk))

        arr = np.frombuffer(chunk, dtype=np.uint8).copy()
        tensor = torch.from_numpy(arr).float()
        tensor = (tensor / 127.5) - 1.0  # [0,255] β†’ [-1, 1]
        tensor = tensor.reshape(3, self.img_size, self.img_size)

        return tensor, 0


# ── Patch Utilities ──────────────────────────────────────────────

def extract_patches(images, patch_size=16):
    B, C, H, W = images.shape
    gh, gw = H // patch_size, W // patch_size
    patches = images.reshape(B, C, gh, patch_size, gw, patch_size)
    patches = patches.permute(0, 2, 4, 1, 3, 5)
    return patches.reshape(B, gh * gw, C * patch_size * patch_size), gh, gw


def stitch_patches(patches, gh, gw, patch_size=16):
    B = patches.shape[0]
    patches = patches.reshape(B, gh, gw, 3, patch_size, patch_size)
    patches = patches.permute(0, 3, 1, 4, 2, 5)
    return patches.reshape(B, 3, gh * patch_size, gw * patch_size)


class BoundarySmooth(nn.Module):
    def __init__(self, channels=3, mid=16):
        super().__init__()
        self.net = nn.Sequential(
            nn.Conv2d(channels, mid, 3, padding=1), nn.GELU(),
            nn.Conv2d(mid, channels, 3, padding=1))
        nn.init.zeros_(self.net[-1].weight)
        nn.init.zeros_(self.net[-1].bias)
    def forward(self, x):
        return x + self.net(x)


class SpectralCrossAttention(nn.Module):
    def __init__(self, D, n_heads=4, max_alpha=0.2, alpha_init=-2.0):
        super().__init__()
        self.n_heads = n_heads
        self.head_dim = D // n_heads
        self.max_alpha = max_alpha
        assert D % n_heads == 0
        self.qkv = nn.Linear(D, 3 * D)
        self.out_proj = nn.Linear(D, D)
        self.norm = nn.LayerNorm(D)
        self.scale = self.head_dim ** -0.5
        self.alpha_logits = nn.Parameter(torch.full((D,), alpha_init))

    @property
    def alpha(self):
        return self.max_alpha * torch.sigmoid(self.alpha_logits)

    def forward(self, S):
        B, N, D = S.shape
        S_normed = self.norm(S)
        qkv = self.qkv(S_normed).reshape(B, N, 3, self.n_heads, self.head_dim)
        qkv = qkv.permute(2, 0, 3, 1, 4)
        q, k, v = qkv[0], qkv[1], qkv[2]
        attn = (q @ k.transpose(-2, -1)) * self.scale
        attn = attn.softmax(dim=-1)
        out = (attn @ v).transpose(1, 2).reshape(B, N, D)
        gate = torch.tanh(self.out_proj(out))
        return S * (1.0 + self.alpha.unsqueeze(0).unsqueeze(0) * gate)


class PatchSVAE(nn.Module):
    def __init__(self, matrix_v=256, D=16, patch_size=16, hidden=768,
                 depth=4, n_cross_layers=2):
        super().__init__()
        self.matrix_v = matrix_v
        self.D = D
        self.patch_size = patch_size
        self.patch_dim = 3 * patch_size * patch_size
        self.mat_dim = matrix_v * D

        self.enc_in = nn.Linear(self.patch_dim, hidden)
        self.enc_blocks = nn.ModuleList([
            nn.Sequential(nn.LayerNorm(hidden), nn.Linear(hidden, hidden),
                          nn.GELU(), nn.Linear(hidden, hidden))
            for _ in range(depth)])
        self.enc_out = nn.Linear(hidden, self.mat_dim)

        self.dec_in = nn.Linear(self.mat_dim, hidden)
        self.dec_blocks = nn.ModuleList([
            nn.Sequential(nn.LayerNorm(hidden), nn.Linear(hidden, hidden),
                          nn.GELU(), nn.Linear(hidden, hidden))
            for _ in range(depth)])
        self.dec_out = nn.Linear(hidden, self.patch_dim)

        nn.init.orthogonal_(self.enc_out.weight)

        self.cross_attn = nn.ModuleList([
            SpectralCrossAttention(D, n_heads=min(4, D))
            for _ in range(n_cross_layers)])
        self.boundary_smooth = BoundarySmooth(channels=3, mid=16)

    def encode_patches(self, patches):
        B, N, _ = patches.shape
        flat = patches.reshape(B * N, -1)
        h = F.gelu(self.enc_in(flat))
        for block in self.enc_blocks:
            h = h + block(h)
        M = self.enc_out(h).reshape(B * N, self.matrix_v, self.D)
        M = F.normalize(M, dim=-1)
        U, S, Vt = svd_fp64(M)
        U = U.reshape(B, N, self.matrix_v, self.D)
        S = S.reshape(B, N, self.D)
        Vt = Vt.reshape(B, N, self.D, self.D)
        M = M.reshape(B, N, self.matrix_v, self.D)
        S_coord = S
        for layer in self.cross_attn:
            S_coord = layer(S_coord)
        return {'U': U, 'S_orig': S, 'S': S_coord, 'Vt': Vt, 'M': M}

    def decode_patches(self, U, S, Vt):
        B, N, V, D = U.shape
        U_flat = U.reshape(B * N, V, D)
        S_flat = S.reshape(B * N, D)
        Vt_flat = Vt.reshape(B * N, D, D)
        M_hat = torch.bmm(U_flat * S_flat.unsqueeze(1), Vt_flat)
        h = F.gelu(self.dec_in(M_hat.reshape(B * N, -1)))
        for block in self.dec_blocks:
            h = h + block(h)
        return self.dec_out(h).reshape(B, N, -1)

    def forward(self, images):
        patches, gh, gw = extract_patches(images, self.patch_size)
        svd = self.encode_patches(patches)
        decoded = self.decode_patches(svd['U'], svd['S'], svd['Vt'])
        recon = stitch_patches(decoded, gh, gw, self.patch_size)
        recon = self.boundary_smooth(recon)
        return {'recon': recon, 'svd': svd, 'gh': gh, 'gw': gw}

    @staticmethod
    def effective_rank(S):
        p = S / (S.sum(-1, keepdim=True) + 1e-8)
        p = p.clamp(min=1e-8)
        return (-(p * p.log()).sum(-1)).exp()


# ── Byte Accuracy ────────────────────────────────────────────────

def byte_accuracy(recon, target):
    """Compute exact byte recovery rate."""
    orig = ((target.flatten(1) + 1.0) * 127.5).round().clamp(0, 255).long()
    pred = ((recon.flatten(1) + 1.0) * 127.5).round().clamp(0, 255).long()
    return (orig == pred).float().mean().item()


def sample_text_reconstruction(model, dataset, device, n=3):
    """Show actual text reconstruction examples."""
    model.eval()
    img_size = dataset.img_size

    for i in range(n):
        tensor, _ = dataset[i * 1000]  # spread samples across corpus
        tensor = tensor.unsqueeze(0).to(device)

        with torch.no_grad():
            out = model(tensor)
            recon = out['recon']

        # Decode original
        orig_bytes = ((tensor.squeeze(0).cpu().flatten() + 1.0) * 127.5).round().clamp(0, 255).byte().numpy()
        orig_text = orig_bytes.tobytes().decode('utf-8', errors='replace')[:200]

        # Decode reconstruction
        recon_bytes = ((recon.squeeze(0).cpu().flatten() + 1.0) * 127.5).round().clamp(0, 255).byte().numpy()
        recon_text = recon_bytes.tobytes().decode('utf-8', errors='replace')[:200]

        acc = byte_accuracy(recon, tensor)
        print(f"\n  Sample {i+1}:")
        print(f"    Original: {repr(orig_text[:100])}")
        print(f"    Recon:    {repr(recon_text[:100])}")
        print(f"    Byte acc: {acc*100:.1f}%")


# ── Training ─────────────────────────────────────────────────────

def train():
    # ── Config ──
    V, D, patch_size = 256, 16, 16
    hidden, depth = 768, 4
    n_cross_layers = 2
    batch_size = 128
    lr = 1e-4
    epochs = 100
    target_cv = 0.125
    cv_weight, boost, sigma = 0.3, 0.5, 0.15
    img_size = 128

    save_dir = '/content/checkpoints'
    save_every = 10
    report_every = 2000
    hf_repo = 'AbstractPhil/geolip-SVAE'
    hf_version = 'v17_alexandria'
    tb_dir = '/content/runs'

    # ── Pretrained: load from Johanna omega or Fresnel ──
    # Johanna omega knows arbitrary bytes. Fresnel knows images.
    # Johanna is the better starting point for text.
    pretrained_repo = 'AbstractPhil/geolip-SVAE'
    pretrained_file = 'v16_johanna_omega/checkpoints/best.pt'
    # Fallback: Gaussian Johanna if omega not ready yet
    pretrained_fallback = 'v14_noise/checkpoints/epoch_0200.pt'

    os.makedirs(save_dir, exist_ok=True)
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

    # ── TensorBoard ──
    from torch.utils.tensorboard import SummaryWriter
    run_name = f"alexandria_V{V}_D{D}_h{hidden}_d{depth}"
    tb_path = os.path.join(tb_dir, run_name)
    writer = SummaryWriter(tb_path)
    print(f"  TensorBoard: {tb_path}")

    # ── HuggingFace ──
    hf_enabled = False
    try:
        from huggingface_hub import HfApi, hf_hub_download
        api = HfApi()
        api.whoami()
        hf_enabled = True
        hf_prefix = f"{hf_version}/checkpoints"
        print(f"  HuggingFace: {hf_repo}/{hf_prefix}")
    except Exception as e:
        print(f"  HuggingFace: disabled ({e})")

    def upload_to_hf(local_path, remote_name):
        if not hf_enabled:
            return
        try:
            api.upload_file(path_or_fileobj=local_path,
                            path_in_repo=f"{hf_prefix}/{remote_name}",
                            repo_id=hf_repo, repo_type="model")
            print(f"  ☁️  Uploaded: {hf_repo}/{hf_prefix}/{remote_name}")
        except Exception as e:
            print(f"  ⚠️  HF upload failed: {e}")

    # ── Load pretrained ──
    print(f"\n  Loading pretrained weights...")
    ckpt = None
    for fname in [pretrained_file, pretrained_fallback]:
        try:
            ckpt_path = hf_hub_download(repo_id=pretrained_repo,
                                         filename=fname, repo_type="model")
            ckpt = torch.load(ckpt_path, map_location='cpu', weights_only=False)
            print(f"  Loaded: {fname}")
            print(f"  Epoch: {ckpt['epoch']}, MSE: {ckpt['test_mse']:.6f}")
            break
        except Exception as e:
            print(f"  {fname}: {e}")

    # ── Model ──
    model = PatchSVAE(matrix_v=V, D=D, patch_size=patch_size,
                      hidden=hidden, depth=depth,
                      n_cross_layers=n_cross_layers).to(device)

    if ckpt is not None:
        model.load_state_dict(ckpt['model_state_dict'], strict=True)
        print(f"  Loaded pretrained weights into model")
    else:
        print(f"  ⚠️  No pretrained weights β€” training from scratch")

    total_params = sum(p.numel() for p in model.parameters())
    print(f"  Params: {total_params:,}")

    opt = torch.optim.Adam(model.parameters(), lr=lr)
    sched = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=epochs)

    # ── Data: Wikipedia ──
    print(f"\n  Loading Wikipedia corpus...")
    train_ds = WikiTextAsImage(size=200000, img_size=img_size, split='train')
    val_ds = WikiTextAsImage(size=5000, img_size=img_size, split='train')

    train_loader = torch.utils.data.DataLoader(
        train_ds, batch_size=batch_size, shuffle=True,
        num_workers=4, pin_memory=True, drop_last=True)
    test_loader = torch.utils.data.DataLoader(
        val_ds, batch_size=batch_size, shuffle=False,
        num_workers=4, pin_memory=True)

    n_patches = (img_size // patch_size) ** 2
    batches_per_epoch = len(train_loader)

    print(f"\n  ALEXANDRIA β€” The Library in Geometry")
    print(f"  Wikipedia β†’ UTF-8 bytes β†’ (3, {img_size}, {img_size}) β†’ PatchSVAE")
    print(f"  {n_patches} patches, ({V},{D}), hidden={hidden}, depth={depth}")
    print(f"  Batch={batch_size}, batches/epoch={batches_per_epoch}")
    print(f"  Bytes per sample: {3 * img_size * img_size:,}")
    print(f"  Text per sample: ~{3 * img_size * img_size // 5:,} words")
    print("=" * 100)
    print(f" {'ep':>3} {'batch':>7} | {'loss':>7} {'recon':>7} {'byteacc':>8} | "
          f"{'S0':>6} {'SD':>6} {'ratio':>5} {'erank':>5} | "
          f"{'row_cv':>7} {'prox':>5} | {'S_delta':>7}")
    print("-" * 100)

    best_recon = float('inf')
    global_batch = 0

    def save_checkpoint(path, epoch, test_mse, extra=None, upload=True):
        ckpt_out = {
            'epoch': epoch, 'test_mse': test_mse,
            'global_batch': global_batch,
            'model_state_dict': model.state_dict(),
            'optimizer_state_dict': opt.state_dict(),
            'scheduler_state_dict': sched.state_dict(),
            'config': {
                'V': V, 'D': D, 'patch_size': patch_size,
                'hidden': hidden, 'depth': depth,
                'n_cross_layers': n_cross_layers, 'target_cv': target_cv,
                'dataset': 'wikipedia_en', 'modality': 'text',
                'pretrained_from': pretrained_file,
                'img_size': img_size, 'lr': lr,
            },
        }
        if extra:
            ckpt_out.update(extra)
        torch.save(ckpt_out, path)
        size_mb = os.path.getsize(path) / (1024 * 1024)
        print(f"  πŸ’Ύ Saved: {path} ({size_mb:.1f}MB, ep{epoch}, MSE={test_mse:.6f})")
        if upload:
            upload_to_hf(path, os.path.basename(path))

    # ── Training Loop ──
    for epoch in range(1, epochs + 1):
        model.train()
        total_loss, total_recon, total_acc, n = 0, 0, 0, 0
        last_cv, last_prox, recon_w = target_cv, 1.0, 1.0 + boost
        t0 = time.time()

        pbar = tqdm(train_loader, desc=f"Ep {epoch}/{epochs}",
                    bar_format='{l_bar}{bar:20}{r_bar}')
        for batch_idx, (images, _) in enumerate(pbar):
            images = images.to(device)
            opt.zero_grad()
            out = model(images)
            recon_loss = F.mse_loss(out['recon'], images)

            with torch.no_grad():
                if batch_idx % 50 == 0:
                    current_cv = cv_of(out['svd']['M'][0, 0])
                    if current_cv > 0:
                        last_cv = current_cv
                    delta = last_cv - target_cv
                    last_prox = math.exp(-delta**2 / (2 * sigma**2))

                # Byte accuracy every 100 batches
                if batch_idx % 100 == 0:
                    batch_acc = byte_accuracy(out['recon'], images)
                    total_acc += batch_acc
                    pbar.set_postfix_str(
                        f"mse={recon_loss.item():.4f} bytes={batch_acc*100:.0f}% cv={last_cv:.3f}",
                        refresh=False)

            recon_w = 1.0 + boost * last_prox
            cv_pen = cv_weight * (1.0 - last_prox)
            cv_l = (last_cv - target_cv) ** 2
            loss = recon_w * recon_loss + cv_pen * cv_l
            loss.backward()

            torch.nn.utils.clip_grad_norm_(model.cross_attn.parameters(), max_norm=0.5)
            opt.step()

            total_loss += loss.item() * len(images)
            total_recon += recon_loss.item() * len(images)
            n += len(images)
            global_batch += 1

            # ── Readout ──
            if global_batch % report_every == 0:
                model.eval()
                with torch.no_grad():
                    test_imgs, _ = next(iter(test_loader))
                    test_imgs = test_imgs.to(device)
                    test_out = model(test_imgs)
                    test_mse = F.mse_loss(test_out['recon'], test_imgs).item()
                    test_acc = byte_accuracy(test_out['recon'], test_imgs)
                    S_mean = test_out['svd']['S'].mean(dim=(0, 1))
                    S_orig = test_out['svd']['S_orig'].mean(dim=(0, 1))
                    erank = model.effective_rank(
                        test_out['svd']['S'].reshape(-1, D)).mean().item()
                    s_delta = (S_mean - S_orig).abs().mean().item()
                    ratio = (S_mean[0] / (S_mean[-1] + 1e-8)).item()

                writer.add_scalar('train/recon', total_recon / n, global_batch)
                writer.add_scalar('test/recon_mse', test_mse, global_batch)
                writer.add_scalar('test/byte_accuracy', test_acc, global_batch)
                writer.add_scalar('geo/row_cv', last_cv, global_batch)
                writer.add_scalar('geo/ratio', ratio, global_batch)
                writer.add_scalar('geo/erank', erank, global_batch)
                writer.add_scalar('geo/S0', S_mean[0].item(), global_batch)
                writer.add_scalar('cross_attn/s_delta', s_delta, global_batch)

                print(f"\n {epoch:3d} {global_batch:7d} | "
                      f"{total_loss/n:7.4f} {total_recon/n:7.4f} {test_acc*100:7.1f}% | "
                      f"{S_mean[0]:6.3f} {S_mean[-1]:6.3f} {ratio:5.2f} {erank:5.2f} | "
                      f"{last_cv:7.4f} {last_prox:5.3f} | "
                      f"{s_delta:7.5f}")

                if test_mse < best_recon:
                    best_recon = test_mse
                    save_checkpoint(os.path.join(save_dir, 'best.pt'),
                                    epoch, test_mse,
                                    extra={'byte_accuracy': test_acc},
                                    upload=False)
                model.train()

        pbar.close()
        sched.step()
        epoch_time = time.time() - t0

        # ── Epoch eval ──
        model.eval()
        test_recon_total, test_acc_total, test_n = 0, 0, 0
        with torch.no_grad():
            for test_imgs, _ in test_loader:
                test_imgs = test_imgs.to(device)
                out = model(test_imgs)
                test_recon_total += F.mse_loss(out['recon'], test_imgs).item() * len(test_imgs)
                test_acc_total += byte_accuracy(out['recon'], test_imgs) * len(test_imgs)
                test_n += len(test_imgs)
        epoch_mse = test_recon_total / test_n
        epoch_acc = test_acc_total / test_n

        print(f"  Epoch {epoch}: {epoch_time:.1f}s, MSE={epoch_mse:.6f}, "
              f"bytes={epoch_acc*100:.1f}%, best={best_recon:.6f}")

        # Text samples every 10 epochs
        if epoch % 10 == 0 or epoch == 1:
            print(f"\n  ── Text Reconstruction Samples ──")
            sample_text_reconstruction(model, train_ds, device, n=3)

        if epoch_mse < best_recon:
            best_recon = epoch_mse
            save_checkpoint(os.path.join(save_dir, 'best.pt'),
                            epoch, epoch_mse,
                            extra={'byte_accuracy': epoch_acc},
                            upload=False)

        if epoch % save_every == 0:
            save_checkpoint(os.path.join(save_dir, f'epoch_{epoch:04d}.pt'),
                            epoch, epoch_mse,
                            extra={'byte_accuracy': epoch_acc})
            best_path = os.path.join(save_dir, 'best.pt')
            if os.path.exists(best_path):
                upload_to_hf(best_path, 'best.pt')
            writer.flush()
            if hf_enabled:
                try:
                    api.upload_folder(folder_path=tb_path,
                                      path_in_repo=f"{hf_version}/tensorboard/{run_name}",
                                      repo_id=hf_repo, repo_type="model")
                    print(f"  ☁️  TB synced")
                except:
                    pass

    writer.close()
    print(f"\n  ALEXANDRIA TRAINING COMPLETE")
    print(f"  Best MSE: {best_recon:.6f}")
    print(f"  Checkpoints: {save_dir}/")


if __name__ == "__main__":
    torch.set_float32_matmul_precision('high')
    train()