File size: 24,970 Bytes
6e408ce
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""
Chimera 5.2 — CPU-first training script.

Highlights vs the previous version:

* MeZO optimiser uses a single deterministic seed per step, samples each
  parameter's perturbation direction *on demand* via per-parameter seeds and
  drops the heavy direction cache.  This brings the memory cost of MeZO back
  down to "1× model" exactly as advertised.
* AdamW path uses fused parameter groups and shares the same loss closure as
  MeZO so accumulation and logging are identical between modes.
* Logging never references an undefined ``lr`` (the previous draft printed it
  before the AdamW step ran on the first accumulator boundary).
* Gradient checkpointing falls back to ``use_reentrant=False`` (the modern,
  faster path).
* Tokeniser/dataset loading is unchanged but the Python loops are skipped
  entirely for ``max_tokens=0``.

Recommended commands::

    # MeZO smoke test on TinyStories
    python train.py --scale tiny --seq_len 64 --max_steps 20 --optimizer mezo

    # AdamW with grad checkpointing + bf16
    python train.py --scale small --seq_len 256 --max_steps 1000 \\
                   --optimizer adamw --grad_checkpoint --bf16
"""

from __future__ import annotations

import argparse
import json
import math
import os
import sys
import time

# CPU threading must be configured *before* importing torch.
def _setup_cpu_runtime() -> None:
    n_cpus = os.cpu_count() or 4
    os.environ.setdefault("OMP_NUM_THREADS", str(n_cpus))
    os.environ.setdefault("MKL_NUM_THREADS", str(n_cpus))
    os.environ.setdefault("KMP_AFFINITY", "granularity=fine,compact,1,0")
    os.environ.setdefault("KMP_BLOCKTIME", "1")
    os.environ.setdefault("MALLOC_CONF", "background_thread:true,metadata_thp:auto")


_setup_cpu_runtime()


import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader, Dataset

sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))

from chimera import Chimera51ForCausalLM
from chimera.quantization import BitLinear


torch.set_num_threads(int(os.environ.get("OMP_NUM_THREADS", os.cpu_count() or 4)))
try:
    torch.set_num_interop_threads(int(os.environ.get("CHIMERA_INTEROP_THREADS", "1")))
except RuntimeError:
    pass


# Optional Intel Extension for PyTorch.
HAS_IPEX = False
try:  # pragma: no cover - optional dependency.
    import intel_extension_for_pytorch as ipex  # noqa: F401
    HAS_IPEX = True
except Exception:
    pass


# ---------------------------------------------------------------------------
# MeZO optimiser
# ---------------------------------------------------------------------------

class MeZOOptimizer:
    """Memory-Efficient Zeroth-Order optimiser (Princeton MeZO).

    Each step runs *two* forward passes around ``θ`` and uses the resulting
    loss difference to estimate a projected gradient.  No backward pass and
    no per-parameter optimiser state — memory cost is exactly ``1× model``.

    For BitLinear layers we mask perturbations to currently non-zero ternary
    positions, so ``~1/3`` of the weights skip both perturbation and update.
    """

    def __init__(self, model: nn.Module, lr: float = 1e-4, eps: float = 1e-3,
                 weight_decay: float = 0.0, momentum: float = 0.0,
                 direction: str = "rademacher"):
        self.model = model
        self.lr = float(lr)
        self.eps = float(eps)
        self.wd = float(weight_decay)
        self.momentum = float(momentum)
        if direction not in ("rademacher", "gaussian"):
            raise ValueError(f"unknown direction: {direction!r}")
        self.direction = direction

        # Collect trainable parameters once and deduplicate tied weights.
        self._bitlinear_modules: list[tuple[str, BitLinear]] = []
        self._dense_params: list[tuple[str, torch.Tensor]] = []
        seen: set[int] = set()

        for name, module in model.named_modules():
            if isinstance(module, BitLinear):
                self._bitlinear_modules.append((name, module))
                seen.add(id(module.weight))
                if module.bias is not None:
                    seen.add(id(module.bias))

        for name, p in model.named_parameters():
            if p.requires_grad and id(p) not in seen:
                self._dense_params.append((name, p))
                seen.add(id(p))

        # Optional momentum buffer — only allocated when momentum > 0.
        self._momentum: dict[int, torch.Tensor] = {}
        if self.momentum > 0:
            for _, p in self._dense_params:
                self._momentum[id(p)] = torch.zeros_like(p.data)
            for _, m in self._bitlinear_modules:
                self._momentum[id(m.weight)] = torch.zeros_like(m.weight.data)

        # Snapshot ternary non-zero masks once per step.
        self._step_masks: dict[int, torch.Tensor] = {}

    # ------------------------------------------------------------------
    # Direction sampling — deterministic per (step seed, parameter index).
    # ------------------------------------------------------------------

    def _direction(self, p: torch.Tensor, seed: int) -> torch.Tensor:
        gen = torch.Generator(device="cpu")
        gen.manual_seed(int(seed) & 0x7FFF_FFFF_FFFF_FFFF)
        if self.direction == "gaussian":
            return torch.randn(p.shape, dtype=p.dtype, device="cpu",
                               generator=gen).to(p.device)
        z = torch.empty(p.shape, dtype=p.dtype, device="cpu")
        z.bernoulli_(0.5, generator=gen).mul_(2).sub_(1)
        return z.to(p.device)

    def _walk_params(self):
        """Yield ``(seed_offset, param, mask_or_None)`` for every trainable tensor."""
        offset = 0
        for _, module in self._bitlinear_modules:
            yield offset, module.weight.data, self._step_masks.get(id(module.weight))
            offset += 1
            if module.bias is not None:
                yield offset, module.bias.data, None
                offset += 1
        for _, p in self._dense_params:
            yield offset, p.data, None
            offset += 1

    def _perturb(self, base_seed: int, scale: float) -> None:
        for off, p, mask in self._walk_params():
            z = self._direction(p, base_seed + off * 1_000_003)
            if mask is not None:
                z = z * mask.to(dtype=z.dtype, device=z.device)
            p.add_(z, alpha=scale)
        # Mark BitLinear caches stale.
        for _, m in self._bitlinear_modules:
            m.invalidate_packed()

    def _update(self, base_seed: int, projected_grad: float) -> None:
        for off, p, mask in self._walk_params():
            z = self._direction(p, base_seed + off * 1_000_003)
            if mask is not None:
                z = z * mask.to(dtype=z.dtype, device=z.device)
            buf = self._momentum.get(id(p))
            if buf is not None:
                buf.mul_(self.momentum).add_(z, alpha=projected_grad)
                p.add_(buf, alpha=-self.lr)
            else:
                p.add_(z, alpha=-self.lr * projected_grad)
            if self.wd > 0:
                p.mul_(1 - self.lr * self.wd)
        for _, m in self._bitlinear_modules:
            m.invalidate_packed()

    @torch.no_grad()
    def step(self, loss_fn, batch) -> float:
        """Run one MeZO step (two forward passes) and return the mean loss."""
        seed = int(torch.randint(0, 2**31, (1,)).item())

        # Snapshot ternary non-zero masks once for this step.
        self._step_masks = {
            id(m.weight): m.ternary_nonzero_mask().detach()
            for _, m in self._bitlinear_modules
        }

        # Forward at θ + εz.
        self._perturb(seed, +self.eps)
        loss_pos = float(loss_fn(batch).item())

        # Net displacement: θ + εz - 2εz = θ - εz.
        self._perturb(seed, -2.0 * self.eps)
        loss_neg = float(loss_fn(batch).item())

        # Restore θ.
        self._perturb(seed, +self.eps)

        projected_grad = (loss_pos - loss_neg) / (2.0 * self.eps)
        self._update(seed, projected_grad)
        self._step_masks = {}

        return 0.5 * (loss_pos + loss_neg)


# ---------------------------------------------------------------------------
# Dataset & tokenisation helpers.
# ---------------------------------------------------------------------------

class TokenDataset(Dataset):
    def __init__(self, chunks: torch.Tensor):
        self.chunks = chunks

    def __len__(self) -> int:
        return self.chunks.size(0)

    def __getitem__(self, idx: int) -> dict:
        c = self.chunks[idx]
        return {"input_ids": c, "labels": c}


def _matches_category_filter(ex: dict, filters: list) -> bool:
    cat = ex.get("category", "") or ""
    if not cat:
        return False
    cat_lower = cat.lower()
    return any(f.lower() in cat_lower for f in filters)


def _format_example(ex: dict, tok, text_column: str = "auto",
                    include_reasoning: bool = False) -> str:
    if text_column == "auto":
        for cand in ("messages", "text", "content", "conversation"):
            if cand in ex:
                text_column = cand
                break
        else:
            text_column = ""

    if text_column == "messages" and "messages" in ex:
        msgs = ex["messages"]
        if include_reasoning and isinstance(msgs, list):
            new_msgs = []
            for m in msgs:
                if isinstance(m, dict) and m.get("role") == "assistant" and "reasoning" in m:
                    new_msgs.append({
                        "role": "assistant",
                        "content": (f"<|thinking|>\n{m['reasoning']}\n<|/thinking|>\n"
                                    f"{m.get('content', '')}"),
                    })
                else:
                    new_msgs.append(m)
            msgs = new_msgs
        return tok.apply_chat_template(msgs)

    if text_column and text_column in ex:
        val = ex[text_column]
        if isinstance(val, str):
            return val
        if isinstance(val, list) and val and isinstance(val[0], dict):
            return tok.apply_chat_template(val)
        return str(val)
    return str(ex)


def build_dataset(seq_len: int, max_samples=None, max_tokens=None,
                  split: str = "train",
                  dataset_name: str = "roneneldan/TinyStories",
                  dataset_config: str = None, text_column: str = "auto",
                  category_filter: str = None,
                  include_reasoning: bool = False):
    from datasets import load_dataset
    from chimera import ChimeraTokenizer

    print(f"[DATA] Loading {dataset_name} ({split})...")
    load_kwargs = {"split": split, "streaming": True}
    if dataset_config:
        load_kwargs["name"] = dataset_config
    ds = load_dataset(dataset_name, **load_kwargs)
    tok = ChimeraTokenizer(pretrained="o200k_base")

    cat_filters = ([c.strip() for c in category_filter.split(",") if c.strip()]
                   if category_filter else None)
    if cat_filters:
        print(f"[DATA] Filtering categories: {cat_filters}")

    if max_tokens is not None:
        token_budget = int(max_tokens)
    elif max_samples is not None:
        token_budget = int(max_samples) * (seq_len + 1)
    else:
        token_budget = None

    if token_budget is None or token_budget <= 0:
        # Fallback: list-based collection.
        all_ids: list[int] = []
        target = (max_samples * (seq_len + 1)) if max_samples else float("inf")
        for ex in ds:
            if cat_filters and not _matches_category_filter(ex, cat_filters):
                continue
            text = _format_example(ex, tok, text_column, include_reasoning)
            if not text or not text.strip():
                continue
            ids = tok.encode(text, add_special_tokens=False)
            ids.append(tok.eos_token_id)
            all_ids.extend(ids)
            if len(all_ids) >= target:
                break
        all_ids = torch.tensor(all_ids, dtype=torch.long)
    else:
        # Pre-allocated token buffer.
        buffer = torch.empty(token_budget, dtype=torch.long)
        buf_idx = 0
        processed = skipped = 0
        for ex in ds:
            if cat_filters and not _matches_category_filter(ex, cat_filters):
                skipped += 1
                continue
            text = _format_example(ex, tok, text_column, include_reasoning)
            if not text or not text.strip():
                skipped += 1
                continue
            ids = tok.encode(text, add_special_tokens=False)
            ids.append(tok.eos_token_id)
            n = len(ids)
            if buf_idx + n > token_budget:
                n = token_budget - buf_idx
                if n <= 0:
                    break
                ids = ids[:n]
            if n > 0:
                buffer[buf_idx:buf_idx + n] = torch.tensor(ids, dtype=torch.long)
                buf_idx += n
            processed += 1
            if buf_idx >= token_budget:
                break
            if (processed % 10_000) == 0:
                print(f"  {processed:,} examples, {buf_idx:,} tokens...")
        all_ids = buffer[:buf_idx]
        print(f"[DATA] Processed {processed:,} examples, skipped {skipped:,}.")

    if all_ids.numel() == 0:
        raise ValueError("No data matched filters.")

    n = all_ids.numel() // (seq_len + 1)
    if max_samples:
        n = min(n, max_samples)
    chunks = all_ids[:n * (seq_len + 1)].view(n, seq_len + 1)
    print(f"[DATA] {n:,} chunks × {seq_len} tokens = {n * seq_len:,} total")
    return TokenDataset(chunks), tok


# ---------------------------------------------------------------------------
# Learning-rate schedule.
# ---------------------------------------------------------------------------

def cosine_lr(step: int, warmup: int, total: int, max_lr: float, min_lr: float
              ) -> float:
    if warmup > 0 and step < warmup:
        return max_lr * (step + 1) / warmup
    if step >= total:
        return min_lr
    p = (step - warmup) / max(1, total - warmup)
    return min_lr + 0.5 * (max_lr - min_lr) * (1.0 + math.cos(math.pi * p))


# ---------------------------------------------------------------------------
# Main loop.
# ---------------------------------------------------------------------------

_SCALE_PRESETS = {
    "tiny":   dict(hidden_size=256,  intermediate_size=512,  num_heads=4, head_dim=48),
    "small":  dict(hidden_size=512,  intermediate_size=1024, num_heads=8, head_dim=48),
    "medium": dict(hidden_size=1024, intermediate_size=2048, num_heads=8, head_dim=96),
}


def train(args) -> None:
    with open(args.config) as f:
        config = json.load(f)

    if args.scale in _SCALE_PRESETS:
        config.update(_SCALE_PRESETS[args.scale])
    config["num_hidden_layers"] = int(config.get("num_hidden_layers", 28))

    config["vocab_size"] = config.get("vocab_size", 200073)
    config.setdefault("gated_deltanet", {})["chunk_size"] = min(args.seq_len, 64)
    config.setdefault("xlstm", {})["memory_size_per_head"] = [config["head_dim"], config["head_dim"]]
    config.setdefault("titans", {}).update({
        "memory_depth": 2, "persistent_memory_slots": 16,
        "local_window_size": min(args.seq_len, 256),
    })
    moe_cfg = config.setdefault("backbone", {}).setdefault("moe", {})
    moe_cfg.setdefault("layers", [3, 7, 11, 15, 19, 23, 27])
    moe_cfg.setdefault("moe_intermediate_size", config["intermediate_size"] // 4)
    moe_cfg.setdefault("n_routed_experts", 8)
    moe_cfg.setdefault("n_shared_experts", 1)
    moe_cfg.setdefault("num_experts_per_tok", 2)
    config.setdefault("looping", {}).update({
        "enabled": True, "prelude": [0, 3], "loop": [4, 23], "coda": [24, 27],
        "loop_range": [1, 3], "loop_default": 2,
    })
    config.setdefault("span_inference", {})["enabled"] = True
    config.setdefault("grammar", {})["enabled"] = True
    config.setdefault("entropy_valve", {})["enabled"] = True
    config.setdefault("debt_ledger", {})["enabled"] = True
    config.setdefault("multimodal", {})["enabled"] = False

    use_mezo = (args.optimizer == "mezo")
    use_bf16 = bool(args.bf16)
    use_compile = bool(args.compile)

    print("=" * 60)
    print(f"CHIMERA 5.2 TRAINING — scale={args.scale}, "
          f"optimizer={'MeZO' if use_mezo else 'AdamW'}, bf16={use_bf16}")
    print(f"Layers={config['num_hidden_layers']}  hidden={config['hidden_size']}  "
          f"vocab={config['vocab_size']}  seq_len={args.seq_len}  steps={args.max_steps}")
    print(f"Threads: {torch.get_num_threads()}  IPEX={HAS_IPEX}")
    print("=" * 60)

    model = Chimera51ForCausalLM(config)
    counts = model.count_parameters()
    print(f"Params: total={counts['total']:,} ternary={counts['ternary']:,}")

    if args.grad_checkpoint and not use_mezo:
        model.enable_gradient_checkpointing()
        print("[OPT] Gradient checkpointing ON")

    if HAS_IPEX and not use_mezo:
        adamw = torch.optim.AdamW(model.parameters(), lr=args.lr)
        model, adamw = ipex.optimize(
            model, optimizer=adamw,
            dtype=torch.bfloat16 if use_bf16 else torch.float32, level="O1")
        print("[OPT] IPEX optimisation applied (level O1)")
    else:
        adamw = None

    if use_compile:
        print("[OPT] Compiling model with torch.compile (inductor)...")
        model = torch.compile(model, backend="inductor", mode="default", dynamic=True)

    dataset, tok = build_dataset(
        args.seq_len, max_samples=args.max_samples, max_tokens=args.max_tokens,
        split=args.dataset_split, dataset_name=args.dataset_name,
        dataset_config=args.dataset_config, text_column=args.text_column,
        category_filter=args.category_filter,
        include_reasoning=args.include_reasoning,
    )
    loader = DataLoader(
        dataset, batch_size=args.batch_size, shuffle=True,
        num_workers=args.num_workers, drop_last=True,
        persistent_workers=args.num_workers > 0,
        prefetch_factor=2 if args.num_workers > 0 else None,
    )

    if use_mezo:
        optimizer = MeZOOptimizer(
            model, lr=args.lr * 0.01, eps=1e-3,
            weight_decay=0.1, momentum=0.9, direction=args.mezo_direction,
        )
    else:
        no_decay = {"A_log", "dt_bias", "norm", "bias", "embed", "energy_weights"}
        decay_params, no_decay_params = [], []
        for n, p in model.named_parameters():
            if not p.requires_grad:
                continue
            if any(tag in n for tag in no_decay):
                no_decay_params.append(p)
            else:
                decay_params.append(p)
        if adamw is None:
            optimizer = torch.optim.AdamW(
                [{"params": decay_params,    "weight_decay": 0.1},
                 {"params": no_decay_params, "weight_decay": 0.0}],
                lr=args.lr, betas=(0.9, 0.95))
        else:
            optimizer = adamw

    def compute_loss(batch) -> torch.Tensor:
        ids = batch["input_ids"][:, :-1]
        labels = batch["labels"][:, 1:]
        if use_bf16:
            with torch.autocast(device_type="cpu", dtype=torch.bfloat16):
                out = model(ids, labels=labels)
        else:
            out = model(ids, labels=labels)
        return out.loss

    os.makedirs(args.output_dir, exist_ok=True)
    log_path = os.path.join(args.output_dir, "log.jsonl")
    log_f = open(log_path, "w", encoding="utf-8")

    model.train()
    step = 0
    cur_lr = args.lr
    total_loss = 0.0
    best_loss = float("inf")
    toks = 0
    t0 = time.time()
    data_iter = iter(loader)
    warmup = min(args.warmup, max(1, args.max_steps // 10))

    if not use_mezo:
        optimizer.zero_grad(set_to_none=True)

    print(f"\n{'=' * 60}\nTraining starts\n{'=' * 60}\n")

    while step < args.max_steps:
        try:
            batch = next(data_iter)
        except StopIteration:
            data_iter = iter(loader)
            batch = next(data_iter)

        if use_mezo:
            cur_lr = cosine_lr(step, warmup, args.max_steps,
                               args.lr * 0.01, args.lr * 0.001)
            optimizer.lr = cur_lr
            loss_val = optimizer.step(compute_loss, batch)
            total_loss += loss_val
        else:
            loss = compute_loss(batch)
            (loss / args.grad_accum).backward()
            total_loss += float(loss.item())
            if (step + 1) % args.grad_accum == 0:
                torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
                cur_lr = cosine_lr(step, warmup, args.max_steps,
                                   args.lr, args.lr * 0.1)
                for pg in optimizer.param_groups:
                    pg["lr"] = cur_lr
                optimizer.step()
                optimizer.zero_grad(set_to_none=True)

        toks += batch["input_ids"][:, :-1].numel()
        step += 1

        if step % args.log_every == 0:
            dt = time.time() - t0
            avg = total_loss / args.log_every
            ppl = math.exp(min(avg, 20))
            tps = toks / dt if dt > 0 else 0
            eta_h = (args.max_steps - step) / (step / dt) / 3600 if dt > 0 else 0.0
            log_f.write(json.dumps({
                "step": step, "loss": round(avg, 4), "ppl": round(ppl, 2),
                "lr": cur_lr, "tok/s": round(tps),
                "optimizer": "mezo" if use_mezo else "adamw",
            }) + "\n")
            log_f.flush()
            print(f"  step {step:>6}/{args.max_steps} | loss {avg:.4f} | "
                  f"ppl {ppl:>8.2f} | lr {cur_lr:.2e} | "
                  f"{tps:.0f} tok/s | ETA {eta_h:.1f}h")
            best_loss = min(best_loss, avg)
            total_loss = 0.0
            toks = 0
            t0 = time.time()

        if step % args.save_every == 0:
            ckpt_dir = os.path.join(args.output_dir, f"ckpt-{step}")
            os.makedirs(ckpt_dir, exist_ok=True)
            raw = getattr(model, "_orig_mod", model)
            torch.save({
                "model": raw.state_dict(), "config": config,
                "step": step, "optimizer": args.optimizer,
            }, os.path.join(ckpt_dir, "ckpt.pt"))
            print(f"  [SAVE] {ckpt_dir}")

    final_dir = os.path.join(args.output_dir, "final")
    os.makedirs(final_dir, exist_ok=True)
    raw = getattr(model, "_orig_mod", model)
    torch.save({
        "model": raw.state_dict(), "config": config,
        "step": step, "best_loss": best_loss,
    }, os.path.join(final_dir, "model.pt"))
    with open(os.path.join(final_dir, "config.json"), "w", encoding="utf-8") as fh:
        json.dump(config, fh, indent=2)
    log_f.close()

    print(f"\n{'=' * 60}")
    print(f"DONE — best loss {best_loss:.4f}, ppl {math.exp(min(best_loss, 20)):.2f}")
    print(f"Saved to {final_dir}")


# ---------------------------------------------------------------------------
# CLI
# ---------------------------------------------------------------------------

def _build_argparser() -> argparse.ArgumentParser:
    p = argparse.ArgumentParser(description="Chimera 5.2 CPU-first training")
    p.add_argument("--config", default="config.json")
    p.add_argument("--scale", default="tiny", choices=["tiny", "small", "medium", "full"])
    p.add_argument("--seq_len", type=int, default=256)
    p.add_argument("--optimizer", default="mezo", choices=["mezo", "adamw"])
    p.add_argument("--batch_size", type=int, default=2)
    p.add_argument("--grad_accum", type=int, default=8)
    p.add_argument("--lr", type=float, default=1e-3)
    p.add_argument("--warmup", type=int, default=200)
    p.add_argument("--max_steps", type=int, default=5000)
    p.add_argument("--max_samples", type=int, default=None)
    p.add_argument("--max_tokens", type=int, default=None)
    p.add_argument("--bf16", action="store_true", default=True)
    p.add_argument("--no-bf16", dest="bf16", action="store_false")
    p.add_argument("--compile", action="store_true", default=False)
    p.add_argument("--grad_checkpoint", action="store_true", default=True)
    p.add_argument("--no-grad-checkpoint", dest="grad_checkpoint", action="store_false")
    p.add_argument("--mezo_direction", choices=["rademacher", "gaussian"],
                   default="rademacher")
    p.add_argument("--dataset_name", default="roneneldan/TinyStories")
    p.add_argument("--dataset_config", default=None)
    p.add_argument("--dataset_split", default="train")
    p.add_argument("--text_column", default="auto")
    p.add_argument("--category_filter", default=None)
    p.add_argument("--include_reasoning", action="store_true", default=False)
    p.add_argument("--num_workers", type=int, default=2)
    p.add_argument("--log_every", type=int, default=10)
    p.add_argument("--save_every", type=int, default=1000)
    p.add_argument("--output_dir", default="./chimera_output")
    return p


if __name__ == "__main__":
    args = _build_argparser().parse_args()
    train(args)