File size: 12,724 Bytes
2ea2835 | 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 | """Strict mmap-backed teacher bank for DINOv3 + V-JEPA2 features."""
from __future__ import annotations
import hashlib
import json
import os
from pathlib import Path
from typing import Any, Mapping, Sequence
import torch
from safetensors import safe_open
from torch.utils.data import Dataset
BANK_KIND = "dual_dinov3_vjepa2_matched_teacher_bank"
BANK_SCHEMA_VERSION = 1
BRANCH_SHAPES = {
"dinov3": (1024, 1024),
"vjepa2": (1024, 1024),
}
def _canonical_bytes(value: Any) -> bytes:
return json.dumps(
value,
sort_keys=True,
separators=(",", ":"),
ensure_ascii=False,
allow_nan=False,
).encode("utf-8")
def canonical_sha256(value: Any) -> str:
return hashlib.sha256(_canonical_bytes(value)).hexdigest()
def file_sha256(path: str | os.PathLike) -> str:
digest = hashlib.sha256()
with open(path, "rb") as handle:
while True:
block = handle.read(8 * 1024 * 1024)
if not block:
break
digest.update(block)
return digest.hexdigest()
def _strict_sha256(value: object, label: str) -> str:
normalized = str(value)
if (
len(normalized) != 64
or any(character not in "0123456789abcdef" for character in normalized)
):
raise ValueError(f"{label} must be a lowercase SHA-256")
return normalized
class DINOv3VJEPA2FeatureBank(Dataset):
"""Metadata dataset plus random-access frozen FP16 feature rows."""
def __init__(
self,
manifest_path: str | os.PathLike,
*,
expected_manifest_sha256: str,
expected_content_hash: str,
expected_source_manifest_sha256: str,
expected_source_csv_sha256: str,
expected_sample_identity_sha256: str,
expected_sample_count: int = 5000,
expected_vjepa_model_sha256: str | None = None,
expected_dinov3_checkpoint_sha256: str | None = None,
expected_dinov3_source_commit: str | None = None,
expected_shard_size: int = 32,
validate_all_shards: bool = True,
) -> None:
self.manifest_path = Path(manifest_path).expanduser().resolve()
if not self.manifest_path.is_file():
raise FileNotFoundError(self.manifest_path)
actual_manifest_sha = file_sha256(self.manifest_path)
if actual_manifest_sha != _strict_sha256(
expected_manifest_sha256, "expected_manifest_sha256"
):
raise ValueError("DINOv3+V-JEPA2 teacher bank manifest SHA-256 mismatch")
with self.manifest_path.open("r", encoding="utf-8") as handle:
manifest = json.load(handle)
if not isinstance(manifest, dict):
raise TypeError("DINOv3+V-JEPA2 teacher bank manifest must be an object")
required = {
"schema_version",
"kind",
"sample_count",
"source_manifest",
"branches",
"samples",
"shards",
"content_hash",
}
if set(manifest) != required:
raise ValueError(
"DINOv3+V-JEPA2 teacher bank manifest keys mismatch: "
f"{sorted(manifest)} != {sorted(required)}")
if (
manifest["schema_version"] != BANK_SCHEMA_VERSION
or manifest["kind"] != BANK_KIND
or int(manifest["sample_count"]) != int(expected_sample_count)
):
raise ValueError("DINOv3+V-JEPA2 teacher bank kind/schema/sample count mismatch")
payload = dict(manifest)
content_hash = payload.pop("content_hash")
if canonical_sha256(payload) != content_hash:
raise ValueError("DINOv3+V-JEPA2 teacher bank content_hash is invalid")
if content_hash != _strict_sha256(
expected_content_hash, "expected_content_hash"
):
raise ValueError("DINOv3+V-JEPA2 teacher bank content identity mismatch")
source = manifest["source_manifest"]
if source.get("sha256") != _strict_sha256(
expected_source_manifest_sha256,
"expected_source_manifest_sha256",
):
raise ValueError("DINOv3+V-JEPA2 bank source manifest SHA-256 mismatch")
if source.get("source_csv_sha256") != _strict_sha256(
expected_source_csv_sha256,
"expected_source_csv_sha256",
):
raise ValueError("DINOv3+V-JEPA2 bank source CSV SHA-256 mismatch")
if source.get("sample_identity_sha256") != _strict_sha256(
expected_sample_identity_sha256,
"expected_sample_identity_sha256",
):
raise ValueError("DINOv3+V-JEPA2 bank sample identity mismatch")
branches = manifest["branches"]
if set(branches) != set(BRANCH_SHAPES):
raise ValueError("DINOv3+V-JEPA2 teacher bank branch set mismatch")
for name, expected_shape in BRANCH_SHAPES.items():
branch = branches[name]
if (
tuple(branch.get("sample_shape", ())) != expected_shape
or branch.get("dtype") != "float16"
or not isinstance(branch.get("feature_contract"), Mapping)
or not isinstance(branch.get("model_identity"), Mapping)
):
raise ValueError(f"DINOv3+V-JEPA2 teacher branch {name!r} contract mismatch")
if expected_vjepa_model_sha256 is not None and (
branches["vjepa2"]["model_identity"].get("sha256")
!= _strict_sha256(
expected_vjepa_model_sha256,
"expected_vjepa_model_sha256",
)
):
raise ValueError("V-JEPA2 bank model SHA-256 mismatch")
if expected_dinov3_checkpoint_sha256 is not None and (
branches["dinov3"]["model_identity"].get("checkpoint_sha256")
!= _strict_sha256(
expected_dinov3_checkpoint_sha256,
"expected_dinov3_checkpoint_sha256",
)
):
raise ValueError("DINOv3 bank checkpoint SHA-256 mismatch")
if expected_dinov3_source_commit is not None and (
branches["dinov3"]["model_identity"].get("source_commit")
!= str(expected_dinov3_source_commit)
):
raise ValueError("DINOv3 bank source commit mismatch")
if type(expected_shard_size) is not int or expected_shard_size <= 0:
raise ValueError("expected_shard_size must be positive")
samples = manifest["samples"]
if len(samples) != int(expected_sample_count):
raise ValueError("DINOv3+V-JEPA2 bank sample table length mismatch")
self.samples = []
self._sample_index = {}
for expected_index, item in enumerate(samples):
if (
not isinstance(item, dict)
or set(item) != {"index", "sample_id", "prompt"}
or int(item["index"]) != expected_index
or not isinstance(item["sample_id"], str)
or not item["sample_id"]
or not isinstance(item["prompt"], str)
):
raise ValueError(
f"invalid DINOv3+V-JEPA2 bank sample record at {expected_index}")
if item["sample_id"] in self._sample_index:
raise ValueError("DINOv3+V-JEPA2 bank sample IDs must be unique")
self._sample_index[item["sample_id"]] = expected_index
self.samples.append(dict(item))
self.root = self.manifest_path.parent
self.shards = []
self._row_to_shard: list[tuple[int, int]] = [None] * len(samples)
expected_start = 0
for shard_index, item in enumerate(manifest["shards"]):
required_shard = {
"index", "path", "sample_start", "sample_end",
"sample_ids_sha256", "sha256", "content_hash",
}
if not isinstance(item, dict) or set(item) != required_shard:
raise ValueError(f"invalid shard record {shard_index}")
start = int(item["sample_start"])
end = int(item["sample_end"])
if (
int(item["index"]) != shard_index
or start != expected_start
or end <= start
or end > len(samples)
or end - start != min(expected_shard_size, len(samples) - start)
):
raise ValueError(
"DINOv3+V-JEPA2 bank shard ranges must be contiguous fixed-size "
f"chunks of {expected_shard_size}")
relative = Path(str(item["path"]))
if relative.is_absolute() or ".." in relative.parts:
raise ValueError("DINOv3+V-JEPA2 bank shard path must stay under bank root")
path = (self.root / relative).resolve()
if self.root not in path.parents or not path.is_file():
raise FileNotFoundError(path)
ids = [samples[row]["sample_id"] for row in range(start, end)]
if canonical_sha256(ids) != item["sample_ids_sha256"]:
raise ValueError("DINOv3+V-JEPA2 bank shard sample ID hash mismatch")
if validate_all_shards and file_sha256(path) != item["sha256"]:
raise ValueError(f"DINOv3+V-JEPA2 bank shard SHA-256 mismatch: {path}")
with safe_open(str(path), framework="pt", device="cpu") as handle:
metadata = handle.metadata() or {}
if metadata.get("content_hash") != item["content_hash"]:
raise ValueError("DINOv3+V-JEPA2 bank shard embedded content hash mismatch")
for name, shape in BRANCH_SHAPES.items():
tensor_shape = tuple(handle.get_slice(name).get_shape())
if tensor_shape != (end - start, *shape):
raise ValueError(
f"DINOv3+V-JEPA2 bank shard tensor {name!r} shape mismatch")
for row in range(start, end):
self._row_to_shard[row] = (shard_index, row - start)
self.shards.append({**item, "resolved_path": str(path)})
expected_start = end
if expected_start != len(samples):
raise ValueError("DINOv3+V-JEPA2 bank shards do not cover all samples")
self.manifest = manifest
self.content_hash = content_hash
self.manifest_sha256 = actual_manifest_sha
def __len__(self) -> int:
return len(self.samples)
def __getitem__(self, index: int) -> dict[str, object]:
item = self.samples[int(index)]
return {
"prompts": item["prompt"],
"sample_ids": item["sample_id"],
"teacher_bank_rows": int(item["index"]),
}
@staticmethod
def collate_fn(items: Sequence[Mapping[str, object]]) -> dict[str, object]:
return {
"prompts": [str(item["prompts"]) for item in items],
"sample_ids": [str(item["sample_ids"]) for item in items],
"teacher_bank_rows": torch.tensor([
int(item["teacher_bank_rows"]) for item in items
], dtype=torch.long),
}
def load_rows(
self,
sample_ids: Sequence[str],
) -> dict[str, torch.Tensor]:
"""Read only requested rows, preserving order and duplicate IDs."""
ids = tuple(str(value) for value in sample_ids)
if not ids or any(value not in self._sample_index for value in ids):
raise KeyError("DINOv3+V-JEPA2 bank request contains an unknown sample ID")
output = {name: [None] * len(ids) for name in BRANCH_SHAPES}
by_shard: dict[int, list[tuple[int, int]]] = {}
for output_index, sample_id in enumerate(ids):
row = self._sample_index[sample_id]
shard_index, offset = self._row_to_shard[row]
by_shard.setdefault(shard_index, []).append((output_index, offset))
for shard_index, requests in by_shard.items():
path = self.shards[shard_index]["resolved_path"]
with safe_open(path, framework="pt", device="cpu") as handle:
slices = {name: handle.get_slice(name) for name in BRANCH_SHAPES}
for output_index, offset in requests:
for name in BRANCH_SHAPES:
output[name][output_index] = slices[name][
offset:offset + 1
].squeeze(0)
return {
name: torch.stack(rows, dim=0)
for name, rows in output.items()
}
__all__ = [
"BANK_KIND",
"BANK_SCHEMA_VERSION",
"BRANCH_SHAPES",
"DINOv3VJEPA2FeatureBank",
"canonical_sha256",
"file_sha256",
]
|