File size: 8,270 Bytes
fa30c0a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
68f349e
 
fa30c0a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
68f349e
fa30c0a
 
 
 
 
 
 
 
 
68f349e
fa30c0a
 
 
 
 
68f349e
 
 
 
 
 
fa30c0a
 
 
68f349e
 
 
 
 
 
 
 
 
 
 
 
 
fa30c0a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Pure contract and provenance helpers for the Fable expert-router campaign.

This module intentionally has no Torch dependency.  It is shared by local
unit tests, the Colab/Kaggle entry points, and the GPU structural smoke.
"""
from __future__ import annotations

import hashlib
import json
import struct
from dataclasses import dataclass
from pathlib import Path
from typing import Any, Iterable


EXPECTED_PROJECTIONS = {
    "gate_proj.weight": (512, 2048),
    "up_proj.weight": (512, 2048),
    "down_proj.weight": (2048, 512),
}

DTYPE_BYTES = {"BF16": 2, "F16": 2}


def sha256(path: Path, chunk_size: int = 8 * 1024 * 1024) -> str:
    digest = hashlib.sha256()
    with path.open("rb") as handle:
        while chunk := handle.read(chunk_size):
            digest.update(chunk)
    return digest.hexdigest()


def canonical_json_sha256(value: Any) -> str:
    encoded = json.dumps(value, sort_keys=True, separators=(",", ":")).encode("utf-8")
    return hashlib.sha256(encoded).hexdigest()


def read_json(path: Path) -> dict[str, Any]:
    payload = json.loads(path.read_text(encoding="utf-8"))
    if not isinstance(payload, dict):
        raise ValueError(f"expected a JSON object: {path}")
    return payload


def safetensors_header(path: Path) -> tuple[int, dict[str, Any]]:
    with path.open("rb") as handle:
        raw_size = handle.read(8)
        if len(raw_size) != 8:
            raise ValueError(f"truncated Safetensors size header: {path}")
        size = struct.unpack("<Q", raw_size)[0]
        if size < 2 or size > 256 * 1024 * 1024:
            raise ValueError(f"implausible Safetensors header size {size}: {path}")
        raw_header = handle.read(size)
        if len(raw_header) != size:
            raise ValueError(f"truncated Safetensors JSON header: {path}")
    header = json.loads(raw_header)
    if not isinstance(header, dict):
        raise ValueError(f"Safetensors header is not an object: {path}")
    return 8 + size, header


@dataclass(frozen=True)
class BankValidation:
    bank_id: str
    tensor_count: int
    payload_bytes: int
    dtypes: tuple[str, ...]
    layers: tuple[int, ...]

    def as_dict(self) -> dict[str, Any]:
        return {
            "bankId": self.bank_id,
            "tensorCount": self.tensor_count,
            "payloadBytes": self.payload_bytes,
            "dtypes": list(self.dtypes),
            "layers": list(self.layers),
        }


def selected_expert_ids(bank: dict[str, Any], layer: int) -> list[int]:
    raw = bank.get("selectedExperts", {}).get(str(layer))
    if not isinstance(raw, list):
        raise ValueError(f"bank {bank.get('id')} has no selected expert list for layer {layer}")
    ids = [int(item) for item in raw]
    if len(ids) != 32 or len(set(ids)) != 32:
        raise ValueError(f"bank {bank.get('id')} layer {layer} must select 32 unique experts")
    return ids


def expected_tensor_names(bank: dict[str, Any]) -> Iterable[tuple[int, int, str, tuple[int, int]]]:
    for layer in range(30):
        for expert in selected_expert_ids(bank, layer):
            for projection, shape in EXPECTED_PROJECTIONS.items():
                yield layer, expert, f"model.layers.{layer}.mlp.experts.{expert}.{projection}", shape


def validate_bank_header_entries(
    header: dict[str, Any], file_payload: int, bank: dict[str, Any]
) -> BankValidation:
    tensor_entries = {key: value for key, value in header.items() if key != "__metadata__"}
    expected = list(expected_tensor_names(bank))
    expected_names = {row[2] for row in expected}
    actual_names = set(tensor_entries)
    missing = sorted(expected_names - actual_names)
    extra = sorted(actual_names - expected_names)
    if missing or extra:
        raise ValueError(
            f"bank tensor identity mismatch: missing={missing[:3]} ({len(missing)}), "
            f"extra={extra[:3]} ({len(extra)})"
        )

    dtypes: set[str] = set()
    payload_bytes = 0
    intervals: list[tuple[int, int, str]] = []
    for layer, expert, name, expected_shape in expected:
        entry = tensor_entries[name]
        shape = tuple(int(item) for item in entry.get("shape", []))
        if shape != expected_shape:
            raise ValueError(
                f"{name} has shape {shape}, expected {expected_shape} "
                f"(layer={layer}, expert={expert})"
            )
        dtype = str(entry.get("dtype"))
        if dtype not in DTYPE_BYTES:
            raise ValueError(f"{name} has unsupported training dtype {dtype}")
        offsets = entry.get("data_offsets")
        if not isinstance(offsets, list) or len(offsets) != 2:
            raise ValueError(f"{name} has malformed data offsets")
        start, end = map(int, offsets)
        expected_bytes = expected_shape[0] * expected_shape[1] * DTYPE_BYTES[dtype]
        if start < 0 or end <= start or end - start != expected_bytes:
            raise ValueError(
                f"{name} has invalid data offsets {offsets}; expected {expected_bytes} bytes"
            )
        intervals.append((start, end, name))
        payload_bytes += end - start
        dtypes.add(dtype)

    # Safetensors does not require tensor entries (or our selected-expert
    # manifest) to be ordered by their payload offsets.  Validate the physical
    # layout independently from semantic tensor identity.
    previous_end = 0
    for start, end, name in sorted(intervals):
        if start != previous_end:
            relation = "overlap" if start < previous_end else "gap"
            raise ValueError(
                f"bank payload has a {relation} before {name}: "
                f"expected offset {previous_end}, found {start}"
            )
        previous_end = end

    if previous_end != file_payload or payload_bytes != file_payload:
        raise ValueError(
            f"bank payload coverage mismatch: last={previous_end}, summed={payload_bytes}, "
            f"file={file_payload}"
        )
    return BankValidation(
        bank_id=str(bank["id"]),
        tensor_count=len(expected),
        payload_bytes=payload_bytes,
        dtypes=tuple(sorted(dtypes)),
        layers=tuple(range(30)),
    )


def validate_bank_header(bank_path: Path, bank: dict[str, Any]) -> BankValidation:
    data_start, header = safetensors_header(bank_path)
    return validate_bank_header_entries(header, bank_path.stat().st_size - data_start, bank)


def validate_curriculum_row(row: dict[str, Any], expected_split: str) -> None:
    if row.get("schema") != "AutonomaFableRouterRecord.v2":
        raise ValueError(f"unexpected curriculum schema: {row.get('schema')}")
    if row.get("split") != expected_split:
        raise ValueError(f"row split {row.get('split')} does not match {expected_split}")
    if row.get("lane") not in {
        "host_preservation",
        "verified_expert",
        "interaction_pattern",
    }:
        raise ValueError(f"unexpected curriculum lane: {row.get('lane')}")
    messages = row.get("messages")
    if not isinstance(messages, list) or not messages:
        raise ValueError("curriculum row has no messages")
    if not any(message.get("role") == "assistant" for message in messages if isinstance(message, dict)):
        raise ValueError("curriculum row has no assistant target")


def iter_jsonl(path: Path) -> Iterable[dict[str, Any]]:
    with path.open("r", encoding="utf-8") as handle:
        for line_number, line in enumerate(handle, 1):
            try:
                payload = json.loads(line)
            except json.JSONDecodeError as exc:
                raise ValueError(f"invalid JSONL at {path}:{line_number}: {exc}") from exc
            if not isinstance(payload, dict):
                raise ValueError(f"non-object JSONL row at {path}:{line_number}")
            yield payload


def verify_file(path: Path, expected_bytes: int, expected_sha256: str) -> dict[str, Any]:
    actual_bytes = path.stat().st_size
    actual_sha256 = sha256(path)
    if actual_bytes != expected_bytes or actual_sha256 != expected_sha256:
        raise ValueError(
            f"artifact mismatch for {path}: bytes={actual_bytes}/{expected_bytes}, "
            f"sha256={actual_sha256}/{expected_sha256}"
        )
    return {"path": str(path), "bytes": actual_bytes, "sha256": actual_sha256}