File size: 17,850 Bytes
49ad2ef | 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 | """Split the existing pilot graphs into one encoder and two unchanged heads.
This is a graph partition, not retraining. Historical bundles are read-only.
The selected graph cuts are specific to the checked Vons pilot exports.
"""
from __future__ import annotations
import argparse
import copy
import hashlib
import json
import math
import shutil
from collections import Counter
from pathlib import Path
from typing import Any
MODEL_LIMIT_BYTES = 64 * 1024 * 1024
ENCODER_INPUTS = ["input_ids", "attention_mask", "token_type_ids", "option_mask"]
EMBEDDINGS = "view_19"
POOLED = "div"
def digest(path: Path) -> str:
value = hashlib.sha256()
with path.open("rb") as source:
for block in iter(lambda: source.read(1024 * 1024), b""):
value.update(block)
return value.hexdigest()
def _onnx() -> Any:
import onnx
return onnx
def _rename(model: Any, names: dict[str, str]) -> None:
for node in model.graph.node:
for fields in (node.input, node.output):
for index, name in enumerate(fields):
fields[index] = names.get(name, name)
for values in (model.graph.input, model.graph.output, model.graph.value_info,
model.graph.initializer):
for value in values:
value.name = names.get(value.name, value.name)
def _validate_external_paths(path: Path) -> None:
"""Check external tensor locations before loading any external bytes."""
onnx = _onnx()
model = onnx.load(path, load_external_data=False)
for tensor in model.graph.initializer:
if tensor.data_location != onnx.TensorProto.EXTERNAL:
continue
metadata = {item.key: item.value for item in tensor.external_data}
location = metadata.get("location", "")
relative = Path(location)
if not location or relative.is_absolute() or ".." in relative.parts:
raise ValueError(f"unsafe external tensor location: {location}")
target = path.parent / relative
if target.is_symlink() or not target.resolve(strict=True).is_relative_to(path.parent.resolve()):
raise ValueError(f"external tensor escapes graph directory: {location}")
def _tensor_hash(tensor: Any) -> str:
onnx = _onnx()
array = onnx.numpy_helper.to_array(tensor)
header = json.dumps([int(tensor.data_type), list(tensor.dims)]).encode()
return hashlib.sha256(header + array.tobytes()).hexdigest()
def encoder_parameters(model: Any) -> Counter[str]:
"""Include named encoder tensors and exporter-folded parameter matrices.
Exporter-folded matmul weights lose their original names. The pilot's
large floating point initializers are parameter matrices; positional/index
constants use integral dtypes. Small arithmetic constants are not weights.
"""
onnx = _onnx()
return Counter(
_tensor_hash(tensor)
for tensor in model.graph.initializer
if tensor.data_type == onnx.TensorProto.FLOAT
and (tensor.name.startswith("encoder.") or math.prod(tensor.dims) >= 4096)
)
def partition_graphs(direct: Any, diffusion: Any) -> dict[str, Any]:
onnx = _onnx()
from onnx.utils import Extractor
encoder = Extractor(direct).extract_model(ENCODER_INPUTS, [EMBEDDINGS, POOLED])
other_encoder = Extractor(diffusion).extract_model(ENCODER_INPUTS, [EMBEDDINGS, POOLED])
if not encoder_parameters(encoder) or encoder_parameters(encoder) != encoder_parameters(other_encoder):
raise ValueError("encoder parameter tensors differ; sharing would change the model")
direct_head = Extractor(direct).extract_model(
[EMBEDDINGS, POOLED, "option_mask"], ["logits", "answerability"],
)
diffusion_head = Extractor(diffusion).extract_model(
[POOLED, "option_mask", "initial_noise"], ["scores", "answerability"],
)
# The original fixed graph obtains its batch size from token_type_ids.
# The pooled representation has precisely the same leading batch axis.
for node in diffusion_head.graph.node:
if "token_type_ids" in node.input:
attributes = {item.name: onnx.helper.get_attribute_value(item) for item in node.attribute}
if node.op_type != "Shape" or attributes.get("start", 0) != 0 or attributes.get("end") != 1:
raise ValueError("unsupported token dependency in diffusion head")
for index, name in enumerate(node.input):
if name == "token_type_ids":
node.input[index] = POOLED
result = {"encoder": encoder, "direct": direct_head, "diffusion": diffusion_head}
for graph in result.values():
_rename(graph, {EMBEDDINGS: "candidate_embeddings", POOLED: "pooled"})
onnx.checker.check_model(graph)
return result
def export_bundle(direct_path: Path, diffusion_path: Path, tokenizer: Path, output: Path,
calibration: Path | None = None) -> dict[str, Any]:
onnx = _onnx()
if output.exists() and any(output.iterdir()):
raise ValueError("output directory must be new or empty; historical bundles are never overwritten")
for path in (direct_path, diffusion_path):
_validate_external_paths(path)
direct = onnx.load(direct_path)
diffusion = onnx.load(diffusion_path)
parts = partition_graphs(direct, diffusion)
output.mkdir(parents=True, exist_ok=True)
records = []
for name, graph in parts.items():
path = output / f"{name}.onnx"
onnx.save_model(copy.deepcopy(graph), path, save_as_external_data=True,
all_tensors_to_one_file=True, location=f"{name}.onnx.data",
size_threshold=1024, convert_attribute=False)
onnx.checker.check_model(str(path))
records.append({"path": path.name, "role": f"{name}_graph"})
external = path.with_suffix(".onnx.data")
if external.exists():
records.append({"path": external.name, "role": f"{name}_weights"})
token_dir = output / "tokenizer"
token_dir.mkdir()
for filename in ("tokenizer.json", "tokenizer_config.json"):
source = tokenizer / filename
if source.is_symlink() or not source.is_file():
raise ValueError(f"required tokenizer asset missing or symlinked: {filename}")
shutil.copyfile(source, token_dir / filename)
records.append({"path": f"tokenizer/{filename}", "role": "tokenizer"})
calibration_data: dict[str, Any] = {"direct": None, "diffusion": None}
calibration_source = None
if calibration is not None:
raw = json.loads(calibration.read_text())
fit = raw.get("calibration", raw.get("config", {}).get("calibration"))
if not isinstance(fit, dict):
raise ValueError("direct calibration input has no fitted calibration parameters")
calibration_data["direct"] = fit
calibration_source = {"sha256": digest(calibration), "scope": "direct only"}
(output / "calibration.json").write_text(json.dumps(calibration_data, indent=2) + "\n")
records.append({"path": "calibration.json", "role": "calibration"})
settings = {
"max_tokens": 512, "max_options": 32, "max_questions": 8,
"diffusion_inference_steps": 4, "diffusion_training_steps": 32,
"diffusion_candidate_slots": 32, "initial_noise": "host-supplied standard normal float32",
"default_abstain_threshold": 0.55, "default_answerability_threshold": 0.5,
"calibration": "calibration.json", "calibration_source": calibration_source,
}
(output / "config.json").write_text(json.dumps(settings, indent=2) + "\n")
records.append({"path": "config.json", "role": "config"})
for record in records:
path = output / record["path"]
record.update(bytes=path.stat().st_size, sha256=digest(path))
total = sum(record["bytes"] for record in records)
source_assets = []
for label, source in (("direct", direct_path), ("diffusion", diffusion_path)):
source_names = {source.name}
source_model = onnx.load(source, load_external_data=False)
for tensor in source_model.graph.initializer:
source_names.update(item.value for item in tensor.external_data if item.key == "location")
source_assets.extend({"backend": label, "path": filename,
"bytes": (source.parent / filename).stat().st_size,
"sha256": digest(source.parent / filename)}
for filename in sorted(source_names))
manifest = {
"schema": "vons.shared-bundle/v1", "model_id": "vons-pilot-shared-v1",
"transformation": "partition existing graphs; no training or parameter changes",
"encoder": {"graph": "encoder.onnx", "outputs": ["candidate_embeddings", "pooled"]},
"heads": {
"direct": {"graph": "direct.onnx", "inputs": ["candidate_embeddings", "pooled", "option_mask"]},
"diffusion": {"graph": "diffusion.onnx", "inputs": ["pooled", "option_mask", "initial_noise"]},
},
"files": records, "model_asset_bytes": total, "limit_bytes": MODEL_LIMIT_BYTES,
"size_pass": total <= MODEL_LIMIT_BYTES,
"source_graphs": {"direct_sha256": digest(direct_path), "diffusion_sha256": digest(diffusion_path)},
"source_assets": source_assets,
"source_tool_sha256": digest(Path(__file__)),
"release_status": "research_candidate; parity and notices required before release",
"runtime_bytes": None, "runtime_bytes_reason": "runtime is not included in this model-only bundle",
}
manifest_path = output / "manifest.json"
manifest_path.write_text(json.dumps(manifest, indent=2, sort_keys=True) + "\n")
(output / "manifest.sha256").write_text(f"{digest(manifest_path)} manifest.json\n")
verify_manifest(manifest_path)
return manifest
def verify_manifest(path: Path, *, expected_digest: str | None = None) -> dict[str, Any]:
manifest = json.loads(path.read_text())
if manifest.get("schema") != "vons.shared-bundle/v1":
raise ValueError("unsupported shared bundle manifest")
expected = expected_digest or path.with_suffix(".sha256").read_text().split()[0]
if digest(path) != expected:
raise ValueError("release manifest digest mismatch")
root = path.parent.resolve()
names = set()
total = 0
for item in manifest["files"]:
relative = Path(item["path"])
target = root / relative
if relative.is_absolute() or ".." in relative.parts or target.is_symlink():
raise ValueError("unsafe manifest asset path")
if not target.resolve(strict=True).is_relative_to(root):
raise ValueError("manifest asset escapes bundle")
if relative.as_posix() in names:
raise ValueError("duplicate manifest asset")
names.add(relative.as_posix())
if target.stat().st_size != item["bytes"] or digest(target) != item["sha256"]:
raise ValueError(f"asset digest or byte count mismatch: {relative}")
total += target.stat().st_size
if total != manifest["model_asset_bytes"] or manifest["size_pass"] != (total <= MODEL_LIMIT_BYTES):
raise ValueError("bundle byte total mismatch")
required = {"encoder.onnx", "direct.onnx", "diffusion.onnx", "config.json",
"calibration.json", "tokenizer/tokenizer.json", "tokenizer/tokenizer_config.json"}
if not required.issubset(names):
raise ValueError("manifest omits required model assets")
onnx = _onnx()
for filename in ("encoder.onnx", "direct.onnx", "diffusion.onnx"):
_validate_external_paths(root / filename)
graph = onnx.load(root / filename, load_external_data=False)
for tensor in graph.graph.initializer:
for item in tensor.external_data:
if item.key == "location" and item.value not in names:
raise ValueError("manifest omits referenced external tensor data")
return {"pass": True, "model_asset_bytes": total, "size_pass": total <= MODEL_LIMIT_BYTES,
"manifest_sha256": digest(path)}
def verify_parity(bundle: Path, direct_path: Path, diffusion_path: Path,
*, cases: int = 20) -> dict[str, Any]:
"""Compare graph partitions with historical graphs using identical tensors.
This verifies graph transformation only. It is not task-quality evaluation
or a latency benchmark, and does not validate browser tokenization.
"""
import numpy as np
import onnxruntime as ort
verified = verify_manifest(bundle / "manifest.json")
options = ort.SessionOptions()
options.intra_op_num_threads = 2
options.inter_op_num_threads = 1
options.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL
def session(path: Path) -> Any:
return ort.InferenceSession(str(path), options, providers=["CPUExecutionProvider"])
encoder = session(bundle / "encoder.onnx")
heads = {name: session(bundle / f"{name}.onnx") for name in ("direct", "diffusion")}
full = {"direct": session(direct_path), "diffusion": session(diffusion_path)}
samples = []
for case in range(cases):
count = (2, 4, 8, 16, 32)[case % 5]
batch = 2 if case % 4 == 3 else 1
length = (8, 31, 64, 127)[case % 4]
# Token IDs are valid BERT vocabulary entries. Fixed local arrays avoid
# downloads and isolate graph partitioning from tokenizer differences.
random = np.random.default_rng(700 + case)
tokens = random.integers(100, 2000, (batch, count, length), dtype=np.int64)
mask = np.zeros((batch, 32), dtype=np.bool_)
mask[:, :count] = True
ids = np.zeros((batch, 32, 512), np.int64)
attention = np.zeros_like(ids)
ids[:, :count, :length] = tokens
attention[:, :count, :length] = 1
feed = {"input_ids": ids, "attention_mask": attention,
"token_type_ids": np.zeros_like(ids), "option_mask": mask}
representation = encoder.run(None, feed)
for backend in ("direct", "diffusion"):
for seed in ((7, 17, 27) if backend == "diffusion" else (7,)):
noise = np.random.default_rng(seed + case).standard_normal((batch, 32)).astype(np.float32)
shared_feed = {"pooled": representation[1], "option_mask": mask}
original_feed = dict(feed)
if backend == "direct":
shared_feed["candidate_embeddings"] = representation[0]
else:
shared_feed["initial_noise"] = noise
original_feed["initial_noise"] = noise
expected = full[backend].run(None, original_feed)
actual = heads[backend].run(None, shared_feed)
finite = np.broadcast_to(mask, expected[0].shape)
scores_finite = bool(np.isfinite(expected[0][finite]).all()
and np.isfinite(actual[0][finite]).all())
masked = bool(np.isneginf(expected[0][~finite]).all()
and np.isneginf(actual[0][~finite]).all())
answers_finite = bool(np.isfinite(expected[1]).all() and np.isfinite(actual[1]).all())
passed = bool(scores_finite and masked and answers_finite
and np.allclose(expected[0][finite], actual[0][finite], atol=1e-5, rtol=1e-5)
and np.allclose(expected[1], actual[1], atol=1e-5, rtol=1e-5))
samples.append({
"case": case, "backend": backend, "seed": seed, "candidates": count,
"batch": batch, "shape": list(ids.shape), "active_tokens": length,
"input_sha256": hashlib.sha256(ids.tobytes() + attention.tobytes() + mask.tobytes()).hexdigest(),
"noise_sha256": hashlib.sha256(noise.tobytes()).hexdigest() if backend == "diffusion" else None,
"max_abs_scores": float(np.max(np.abs(expected[0][finite] - actual[0][finite]))),
"max_abs_answerability": float(np.max(np.abs(expected[1] - actual[1]))),
"pass": passed,
})
return {"schema": "vons.shared-bundle-parity/v1", "manifest_sha256": verified["manifest_sha256"],
"provider": "CPUExecutionProvider", "ort_version": ort.__version__,
"intra_op_threads": 2, "atol": 1e-5, "rtol": 1e-5, "cases": cases,
"scope": "graph partition only; no quality, browser, calibration or latency claim",
"samples": samples, "pass": all(sample["pass"] for sample in samples)}
def main() -> None:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--direct", type=Path, required=True)
parser.add_argument("--diffusion", type=Path, required=True)
parser.add_argument("--tokenizer", type=Path, required=True)
parser.add_argument("--output", type=Path, required=True)
parser.add_argument("--direct-calibration", type=Path)
parser.add_argument("--parity-report", type=Path)
args = parser.parse_args()
report = export_bundle(args.direct, args.diffusion, args.tokenizer, args.output,
args.direct_calibration)
print(json.dumps({key: report[key] for key in ("model_asset_bytes", "size_pass", "release_status")}))
if args.parity_report is not None:
parity = verify_parity(args.output, args.direct, args.diffusion)
args.parity_report.parent.mkdir(parents=True, exist_ok=True)
args.parity_report.write_text(json.dumps(parity, indent=2, allow_nan=False) + "\n")
print(json.dumps({"parity_pass": parity["pass"], "samples": len(parity["samples"])}))
if not parity["pass"]:
raise SystemExit(1)
if __name__ == "__main__":
main()
|