File size: 25,884 Bytes
5a46e5d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
"""Versioned, hash-verified CPU dynamic-int8 checkpoints for BarunLM."""

from __future__ import annotations

import json
import platform
import shutil
import threading
from collections.abc import Mapping
from dataclasses import dataclass
from pathlib import Path
from types import MappingProxyType
from typing import Any, NoReturn

import torch
from tokenizers import Tokenizer
from torch import nn

from barunlm.config import BarunConfig
from barunlm.evaluation.generation import (
    REQUIRED_CHECKPOINT_FILES,
    load_verified_model,
    verify_checkpoint,
)
from barunlm.model import BarunLM
from barunlm.training.data import sha256_file

INT8_FORMAT_VERSION = "barun-cpu-dynamic-int8-v1"
INT8_MANIFEST_FILE = "quantization_manifest.json"
INT8_WEIGHTS_FILE = "model.int8.pt"
INT8_ARTIFACT_FILES = ("barun_config.json", INT8_WEIGHTS_FILE, "tokenizer.json")
MAX_INT8_WEIGHTS_BYTES = 8 * 1024**3
_QENGINE_LOCK = threading.RLock()


class QuantizationError(RuntimeError):
    """An int8 export or load request failed closed."""


@dataclass(frozen=True, slots=True)
class Int8CheckpointInfo:
    directory: Path
    manifest_sha256: str
    source_checkpoint_sha256: Mapping[str, str]
    artifact_sha256: Mapping[str, str]
    qengine: str
    torch_version: str
    quantized_modules: tuple[str, ...]
    float_linear_modules: tuple[str, ...]
    source_payload_bytes: int
    quantized_payload_bytes: int
    package_bytes: int
    reduction_fraction: float

    def to_dict(self) -> dict[str, Any]:
        return {
            "artifact_sha256": dict(sorted(self.artifact_sha256.items())),
            "directory": str(self.directory),
            "float_linear_modules": list(self.float_linear_modules),
            "format_version": INT8_FORMAT_VERSION,
            "manifest_sha256": self.manifest_sha256,
            "package_bytes": self.package_bytes,
            "qengine": self.qengine,
            "quantized_modules": list(self.quantized_modules),
            "quantized_payload_bytes": self.quantized_payload_bytes,
            "reduction_fraction": self.reduction_fraction,
            "source_checkpoint_sha256": dict(sorted(self.source_checkpoint_sha256.items())),
            "source_payload_bytes": self.source_payload_bytes,
            "torch_version": self.torch_version,
        }


def _is_sha256(value: Any) -> bool:
    return (
        isinstance(value, str)
        and len(value) == 64
        and all(character in "0123456789abcdef" for character in value)
    )


def _reject_pairs(pairs: list[tuple[str, Any]]) -> dict[str, Any]:
    value: dict[str, Any] = {}
    for key, child in pairs:
        if key in value:
            raise QuantizationError(f"duplicate JSON key in int8 manifest: {key!r}")
        value[key] = child
    return value


def _reject_constant(value: str) -> NoReturn:
    raise QuantizationError(f"non-finite JSON value in int8 manifest: {value!r}")


def _strict_json(path: Path) -> Mapping[str, Any]:
    try:
        value = json.loads(
            path.read_text(encoding="utf-8"),
            object_pairs_hook=_reject_pairs,
            parse_constant=_reject_constant,
        )
    except QuantizationError:
        raise
    except (OSError, UnicodeError, json.JSONDecodeError) as error:
        raise QuantizationError(f"invalid JSON artifact: {path}") from error
    if not isinstance(value, Mapping):
        raise QuantizationError(f"JSON artifact must be an object: {path}")
    return value


def _exact_fields(value: Mapping[str, Any], expected: set[str], *, path: str) -> None:
    missing = expected.difference(value)
    unknown = set(value).difference(expected)
    if missing or unknown:
        raise QuantizationError(
            f"{path} fields differ: missing={sorted(missing)!r}, unknown={sorted(unknown)!r}"
        )


def _hash_map(value: Any, *, expected_names: set[str], path: str) -> dict[str, str]:
    if not isinstance(value, Mapping):
        raise QuantizationError(f"{path} must be an object")
    if set(value) != expected_names:
        raise QuantizationError(
            f"{path} file set differs: expected={sorted(expected_names)!r}, "
            f"got={sorted(str(name) for name in value)!r}"
        )
    hashes: dict[str, str] = {}
    for raw_name, digest in value.items():
        if not isinstance(raw_name, str) or not _is_sha256(digest):
            raise QuantizationError(f"{path} contains an invalid file name or SHA-256")
        hashes[raw_name] = digest
    return hashes


def _name_list(value: Any, *, path: str) -> tuple[str, ...]:
    if not isinstance(value, list) or any(not isinstance(name, str) or not name for name in value):
        raise QuantizationError(f"{path} must be an array of non-empty strings")
    names = tuple(value)
    if names != tuple(sorted(set(names))):
        raise QuantizationError(f"{path} must be sorted and unique")
    return names


def _positive_int(value: Any, *, path: str) -> int:
    if type(value) is not int or value < 1:
        raise QuantizationError(f"{path} must be a positive integer")
    return value


def _nonnegative_int(value: Any, *, path: str) -> int:
    if type(value) is not int or value < 0:
        raise QuantizationError(f"{path} must be a non-negative integer")
    return value


def _module_plan(model: BarunLM) -> tuple[tuple[str, ...], tuple[str, ...]]:
    quantized: list[str] = []
    kept_float: list[str] = []
    embedding_weight = model.embedding.weight
    for name, module in model.named_modules():
        if not isinstance(module, nn.Linear):
            continue
        if module.weight is embedding_weight:
            kept_float.append(name)
        else:
            quantized.append(name)
    if not quantized:
        raise QuantizationError("BarunLM contains no eligible Linear modules")
    return tuple(sorted(quantized)), tuple(sorted(kept_float))


def _parameter_counts(
    model: BarunLM,
    quantized_modules: tuple[str, ...],
    float_linear_modules: tuple[str, ...],
) -> dict[str, int]:
    modules = dict(model.named_modules())
    return {
        "float_linear_weight_elements": sum(
            modules[name].weight.numel() for name in float_linear_modules
        ),
        "quantized_linear_weight_elements": sum(
            modules[name].weight.numel() for name in quantized_modules
        ),
        "total_unique_parameters": model.parameter_counts()["total"],
    }


def _set_qengine(qengine: str) -> None:
    if not isinstance(qengine, str) or not qengine or qengine == "none":
        raise QuantizationError("qengine must be explicitly named and cannot be 'none'")
    supported = tuple(torch.backends.quantized.supported_engines)
    if qengine not in supported:
        raise QuantizationError(
            f"qengine {qengine!r} is unavailable; supported engines are {supported!r}"
        )
    try:
        torch.backends.quantized.engine = qengine
    except RuntimeError as error:
        raise QuantizationError(f"failed to activate qengine {qengine!r}") from error
    if torch.backends.quantized.engine != qengine:
        raise QuantizationError(f"qengine activation did not select {qengine!r}")


def _quantize_structure(
    model: BarunLM,
    *,
    qengine: str,
    expected_modules: tuple[str, ...],
) -> BarunLM:
    with _QENGINE_LOCK:
        _set_qengine(qengine)
        qconfig_spec = {
            name: torch.ao.quantization.default_dynamic_qconfig for name in expected_modules
        }
        try:
            quantized = torch.ao.quantization.quantize_dynamic(
                model.cpu().eval(),
                qconfig_spec=qconfig_spec,
                dtype=torch.qint8,
                inplace=False,
            )
        except (RuntimeError, TypeError, ValueError) as error:
            raise QuantizationError(
                f"dynamic int8 conversion failed for qengine {qengine!r}"
            ) from error
    return quantized


def _verify_quantized_modules(
    model: BarunLM,
    *,
    expected_modules: tuple[str, ...],
    expected_weight_elements: int,
) -> None:
    modules = dict(model.named_modules())
    observed_elements = 0
    for name in expected_modules:
        module = modules.get(name)
        if not isinstance(module, torch.ao.nn.quantized.dynamic.Linear):
            raise QuantizationError(f"module {name!r} is not a dynamic quantized Linear")
        weight = module.weight()
        if weight.dtype is not torch.qint8:
            raise QuantizationError(f"module {name!r} does not contain qint8 weights")
        if weight.q_zero_point() != 0:
            raise QuantizationError(f"module {name!r} does not use symmetric zero-point weights")
        observed_elements += weight.numel()
    if observed_elements != expected_weight_elements:
        raise QuantizationError(
            "quantized weight element count differs from the versioned manifest"
        )


def _write_json(path: Path, value: Mapping[str, Any]) -> None:
    path.write_text(
        json.dumps(value, ensure_ascii=False, allow_nan=False, indent=2, sort_keys=True) + "\n",
        encoding="utf-8",
    )


def export_dynamic_int8_checkpoint(
    source_directory: str | Path,
    output_directory: str | Path,
    *,
    expected_source_sha256: Mapping[str, str],
    qengine: str,
) -> Int8CheckpointInfo:
    """Export a new CPU dynamic-int8 checkpoint without modifying its source."""

    source = Path(source_directory).resolve()
    output = Path(output_directory).resolve()
    if output.exists():
        raise QuantizationError(f"refusing to overwrite int8 checkpoint directory: {output}")
    with _QENGINE_LOCK:
        _set_qengine(qengine)
    source_hashes = verify_checkpoint(source, expected_sha256=expected_source_sha256)
    model, _, loaded_hashes = load_verified_model(source, expected_sha256=source_hashes)
    if loaded_hashes != source_hashes:  # pragma: no cover - defensive invariant
        raise QuantizationError("source checkpoint hashes changed between verification and load")
    source_model_hashes = {name: source_hashes[name] for name in REQUIRED_CHECKPOINT_FILES}
    quantized_modules, float_linear_modules = _module_plan(model)
    counts = _parameter_counts(model, quantized_modules, float_linear_modules)
    quantized = _quantize_structure(
        model,
        qengine=qengine,
        expected_modules=quantized_modules,
    )
    _verify_quantized_modules(
        quantized,
        expected_modules=quantized_modules,
        expected_weight_elements=counts["quantized_linear_weight_elements"],
    )

    output.parent.mkdir(parents=True, exist_ok=True)
    output.mkdir(exist_ok=False)
    try:
        shutil.copyfile(source / "barun_config.json", output / "barun_config.json")
        shutil.copyfile(source / "tokenizer.json", output / "tokenizer.json")
        torch.save(quantized.state_dict(), output / INT8_WEIGHTS_FILE)
        artifact_hashes = {name: sha256_file(output / name) for name in INT8_ARTIFACT_FILES}
        if artifact_hashes["barun_config.json"] != source_model_hashes["barun_config.json"]:
            raise QuantizationError("exported config hash differs from the source checkpoint")
        if artifact_hashes["tokenizer.json"] != source_model_hashes["tokenizer.json"]:
            raise QuantizationError("exported tokenizer hash differs from the source checkpoint")
        source_payload_bytes = sum(
            (source / name).stat().st_size for name in REQUIRED_CHECKPOINT_FILES
        )
        quantized_payload_bytes = sum(
            (output / name).stat().st_size for name in INT8_ARTIFACT_FILES
        )
        reduction_fraction = 1.0 - quantized_payload_bytes / source_payload_bytes
        manifest = {
            "algorithm": {
                "activation_quantization": "dynamic",
                "implementation": "torch.ao.quantization.quantize_dynamic",
                "qengine": qengine,
                "tied_embedding_linear_policy": "keep_float",
                "weight_dtype": "qint8",
                "weight_qscheme": "per_tensor_affine_symmetric_zero_point",
            },
            "architecture": "barunlm.model.BarunLM",
            "artifact_sha256": dict(sorted(artifact_hashes.items())),
            "float_linear_modules": list(float_linear_modules),
            "parameter_counts": counts,
            "quantized_modules": list(quantized_modules),
            "runtime": {
                "platform_machine": platform.machine(),
                "platform_system": platform.system(),
                "torch_version": torch.__version__,
            },
            "schema_version": INT8_FORMAT_VERSION,
            "size_bytes": {
                "quantized_payload": quantized_payload_bytes,
                "reduction_fraction": reduction_fraction,
                "source_payload": source_payload_bytes,
            },
            "source_checkpoint_sha256": dict(sorted(source_model_hashes.items())),
        }
        _write_json(output / INT8_MANIFEST_FILE, manifest)
        manifest_sha256 = sha256_file(output / INT8_MANIFEST_FILE)
        return verify_int8_checkpoint(output, expected_manifest_sha256=manifest_sha256)
    except Exception:
        shutil.rmtree(output)
        raise


def verify_int8_checkpoint(
    directory: str | Path,
    *,
    expected_manifest_sha256: str,
) -> Int8CheckpointInfo:
    """Verify a versioned int8 checkpoint and its externally pinned manifest hash."""

    checkpoint = Path(directory).resolve()
    if not _is_sha256(expected_manifest_sha256):
        raise QuantizationError("expected_manifest_sha256 must be a lowercase SHA-256")
    manifest_path = checkpoint / INT8_MANIFEST_FILE
    if not manifest_path.is_file() or manifest_path.is_symlink():
        raise QuantizationError(f"int8 checkpoint is missing {INT8_MANIFEST_FILE!r}")
    try:
        children = tuple(checkpoint.iterdir())
    except OSError as error:
        raise QuantizationError(
            f"cannot inventory int8 checkpoint directory: {checkpoint}"
        ) from error
    expected_children = set(INT8_ARTIFACT_FILES) | {INT8_MANIFEST_FILE}
    observed_children = {child.name for child in children}
    if observed_children != expected_children or any(not child.is_file() for child in children):
        raise QuantizationError(
            "int8 checkpoint file set differs: "
            f"expected={sorted(expected_children)!r}, got={sorted(observed_children)!r}"
        )
    actual_manifest_sha256 = sha256_file(manifest_path)
    if actual_manifest_sha256 != expected_manifest_sha256:
        raise QuantizationError(
            "int8 manifest SHA-256 mismatch: "
            f"expected {expected_manifest_sha256}, got {actual_manifest_sha256}"
        )
    manifest = _strict_json(manifest_path)
    _exact_fields(
        manifest,
        {
            "algorithm",
            "architecture",
            "artifact_sha256",
            "float_linear_modules",
            "parameter_counts",
            "quantized_modules",
            "runtime",
            "schema_version",
            "size_bytes",
            "source_checkpoint_sha256",
        },
        path="$",
    )
    if manifest["schema_version"] != INT8_FORMAT_VERSION:
        raise QuantizationError("unsupported int8 checkpoint format")
    if manifest["architecture"] != "barunlm.model.BarunLM":
        raise QuantizationError("unsupported int8 checkpoint architecture")

    algorithm = manifest["algorithm"]
    if not isinstance(algorithm, Mapping):
        raise QuantizationError("$.algorithm must be an object")
    expected_algorithm = {
        "activation_quantization": "dynamic",
        "implementation": "torch.ao.quantization.quantize_dynamic",
        "tied_embedding_linear_policy": "keep_float",
        "weight_dtype": "qint8",
        "weight_qscheme": "per_tensor_affine_symmetric_zero_point",
    }
    _exact_fields(algorithm, set(expected_algorithm) | {"qengine"}, path="$.algorithm")
    for name, expected in expected_algorithm.items():
        if algorithm[name] != expected:
            raise QuantizationError(f"unsupported int8 algorithm field {name!r}")
    qengine = algorithm["qengine"]
    if not isinstance(qengine, str) or not qengine:
        raise QuantizationError("$.algorithm.qengine must be a non-empty string")

    artifact_hashes = _hash_map(
        manifest["artifact_sha256"],
        expected_names=set(INT8_ARTIFACT_FILES),
        path="$.artifact_sha256",
    )
    source_hashes = _hash_map(
        manifest["source_checkpoint_sha256"],
        expected_names=set(REQUIRED_CHECKPOINT_FILES),
        path="$.source_checkpoint_sha256",
    )
    if artifact_hashes["barun_config.json"] != source_hashes["barun_config.json"]:
        raise QuantizationError("int8 config provenance differs from the source checkpoint")
    if artifact_hashes["tokenizer.json"] != source_hashes["tokenizer.json"]:
        raise QuantizationError("int8 tokenizer provenance differs from the source checkpoint")
    for name, expected in sorted(artifact_hashes.items()):
        path = checkpoint / name
        if not path.is_file() or path.is_symlink():
            raise QuantizationError(f"int8 checkpoint is missing hashed file {name!r}")
        actual = sha256_file(path)
        if actual != expected:
            raise QuantizationError(
                f"int8 checkpoint SHA-256 mismatch for {name}: expected {expected}, got {actual}"
            )

    runtime = manifest["runtime"]
    if not isinstance(runtime, Mapping):
        raise QuantizationError("$.runtime must be an object")
    _exact_fields(
        runtime,
        {"platform_machine", "platform_system", "torch_version"},
        path="$.runtime",
    )
    torch_version = runtime["torch_version"]
    if not isinstance(torch_version, str) or not torch_version:
        raise QuantizationError("$.runtime.torch_version must be a non-empty string")
    for name in ("platform_machine", "platform_system"):
        if not isinstance(runtime[name], str) or not runtime[name]:
            raise QuantizationError(f"$.runtime.{name} must be a non-empty string")

    counts = manifest["parameter_counts"]
    if not isinstance(counts, Mapping):
        raise QuantizationError("$.parameter_counts must be an object")
    _exact_fields(
        counts,
        {
            "float_linear_weight_elements",
            "quantized_linear_weight_elements",
            "total_unique_parameters",
        },
        path="$.parameter_counts",
    )
    _nonnegative_int(
        counts["float_linear_weight_elements"],
        path="$.parameter_counts.float_linear_weight_elements",
    )
    _positive_int(
        counts["quantized_linear_weight_elements"],
        path="$.parameter_counts.quantized_linear_weight_elements",
    )
    _positive_int(
        counts["total_unique_parameters"],
        path="$.parameter_counts.total_unique_parameters",
    )

    sizes = manifest["size_bytes"]
    if not isinstance(sizes, Mapping):
        raise QuantizationError("$.size_bytes must be an object")
    _exact_fields(
        sizes,
        {"quantized_payload", "reduction_fraction", "source_payload"},
        path="$.size_bytes",
    )
    source_payload_bytes = _positive_int(
        sizes["source_payload"], path="$.size_bytes.source_payload"
    )
    recorded_quantized_payload = _positive_int(
        sizes["quantized_payload"], path="$.size_bytes.quantized_payload"
    )
    observed_quantized_payload = sum(
        (checkpoint / name).stat().st_size for name in INT8_ARTIFACT_FILES
    )
    if recorded_quantized_payload != observed_quantized_payload:
        raise QuantizationError("recorded int8 payload size differs from the verified files")
    weight_bytes = (checkpoint / INT8_WEIGHTS_FILE).stat().st_size
    source_payload_bound = max(16 * 1024**2, 2 * source_payload_bytes)
    if weight_bytes > MAX_INT8_WEIGHTS_BYTES or weight_bytes > source_payload_bound:
        raise QuantizationError("verified int8 weight artifact exceeds the format size bound")
    expected_reduction = 1.0 - observed_quantized_payload / source_payload_bytes
    reduction = sizes["reduction_fraction"]
    if (
        not isinstance(reduction, (int, float))
        or not abs(float(reduction) - expected_reduction) < 1e-12
    ):
        raise QuantizationError("recorded int8 size reduction is inconsistent")

    quantized_modules = _name_list(manifest["quantized_modules"], path="$.quantized_modules")
    float_linear_modules = _name_list(
        manifest["float_linear_modules"], path="$.float_linear_modules"
    )
    package_bytes = observed_quantized_payload + manifest_path.stat().st_size
    return Int8CheckpointInfo(
        directory=checkpoint,
        manifest_sha256=actual_manifest_sha256,
        source_checkpoint_sha256=MappingProxyType(source_hashes),
        artifact_sha256=MappingProxyType(artifact_hashes),
        qengine=qengine,
        torch_version=torch_version,
        quantized_modules=quantized_modules,
        float_linear_modules=float_linear_modules,
        source_payload_bytes=source_payload_bytes,
        quantized_payload_bytes=observed_quantized_payload,
        package_bytes=package_bytes,
        reduction_fraction=expected_reduction,
    )


def load_verified_int8_model(
    directory: str | Path,
    *,
    expected_manifest_sha256: str,
) -> tuple[BarunLM, Tokenizer, Int8CheckpointInfo]:
    """Load only the explicitly requested, hash-verified dynamic-int8 representation."""

    info = verify_int8_checkpoint(
        directory,
        expected_manifest_sha256=expected_manifest_sha256,
    )
    if info.torch_version != torch.__version__:
        raise QuantizationError(
            "int8 artifact requires exact torch version "
            f"{info.torch_version!r}, current runtime is {torch.__version__!r}"
        )
    manifest = _strict_json(info.directory / INT8_MANIFEST_FILE)
    runtime = manifest["runtime"]
    if (
        runtime["platform_machine"] != platform.machine()
        or runtime["platform_system"] != platform.system()
    ):
        raise QuantizationError(
            "int8 artifact platform differs from the current CPU runtime; "
            "cross-platform packed-weight loading is not claimed by format v1"
        )
    config = BarunConfig.from_json(info.directory / "barun_config.json")
    tokenizer = Tokenizer.from_file(str(info.directory / "tokenizer.json"))
    if tokenizer.get_vocab_size(with_added_tokens=True) != config.vocab_size:
        raise QuantizationError("int8 tokenizer and model vocabulary sizes differ")
    float_model = BarunLM(config).cpu().eval()
    quantized_modules, float_linear_modules = _module_plan(float_model)
    if quantized_modules != info.quantized_modules:
        raise QuantizationError("int8 quantized module plan differs from the current architecture")
    if float_linear_modules != info.float_linear_modules:
        raise QuantizationError("int8 float-module plan differs from the current architecture")
    counts = _parameter_counts(float_model, quantized_modules, float_linear_modules)
    if counts != manifest["parameter_counts"]:
        raise QuantizationError("int8 parameter counts differ from the current architecture")
    with _QENGINE_LOCK:
        quantized_model = _quantize_structure(
            float_model,
            qengine=info.qengine,
            expected_modules=quantized_modules,
        )
        try:
            state = torch.load(
                info.directory / INT8_WEIGHTS_FILE,
                map_location="cpu",
                weights_only=True,
            )
        except (OSError, RuntimeError, TypeError, ValueError) as error:
            raise QuantizationError("failed to deserialize verified int8 weights") from error
        if not isinstance(state, Mapping):
            raise QuantizationError("verified int8 weight artifact is not a state mapping")
        if config.tie_embeddings:
            embedding_weight = state.get("embedding.weight")
            head_weight = state.get("lm_head.weight")
            if (
                not isinstance(embedding_weight, torch.Tensor)
                or not isinstance(head_weight, torch.Tensor)
                or embedding_weight.dtype is not torch.float32
                or head_weight.dtype is not torch.float32
                or not torch.equal(embedding_weight, head_weight)
            ):
                raise QuantizationError(
                    "verified int8 state does not preserve equal FP32 tied embedding/head weights"
                )
        try:
            missing, unexpected = quantized_model.load_state_dict(state, strict=True)
        except (RuntimeError, TypeError, ValueError) as error:
            raise QuantizationError("verified int8 weights do not match BarunLM") from error
        if missing or unexpected:
            raise QuantizationError(
                "verified int8 weights differ from BarunLM: "
                f"missing={missing}, unexpected={unexpected}"
            )
    if config.tie_embeddings and (
        not isinstance(quantized_model.lm_head, nn.Linear)
        or quantized_model.lm_head.weight is not quantized_model.embedding.weight
        or quantized_model.embedding.weight.dtype is not torch.float32
    ):
        raise QuantizationError("loaded int8 model did not preserve the FP32 tied output head")
    _verify_quantized_modules(
        quantized_model,
        expected_modules=quantized_modules,
        expected_weight_elements=counts["quantized_linear_weight_elements"],
    )
    return quantized_model.eval(), tokenizer, info


__all__ = [
    "INT8_ARTIFACT_FILES",
    "INT8_FORMAT_VERSION",
    "INT8_MANIFEST_FILE",
    "INT8_WEIGHTS_FILE",
    "MAX_INT8_WEIGHTS_BYTES",
    "Int8CheckpointInfo",
    "QuantizationError",
    "export_dynamic_int8_checkpoint",
    "load_verified_int8_model",
    "verify_int8_checkpoint",
]