File size: 19,513 Bytes
349eba5 | 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 | #!/usr/bin/env python3
"""Retokenize Morphosis logical-order traces for strict DSV4 replay.
The input actor rows remain byte-preserved in the sealed source artifact. This
materializer creates one tar.gz bundle per root program, replacing only the GLM
prompt/output token IDs and output length with values produced by the frozen
DeepSeek-V4 serving encoder and tokenizer.
"""
from __future__ import annotations
import argparse
import copy
import hashlib
import importlib.util
import json
import lzma
import os
import re
import shutil
import tarfile
import tempfile
import time
from collections import Counter
from pathlib import Path
from types import ModuleType
from typing import Any, Iterable
from transformers import AutoTokenizer
ROOT_ID_RE = re.compile(r"root-\d{3}-[A-Za-z0-9_.-]+")
def sha256_file(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as handle:
while chunk := handle.read(8 << 20):
digest.update(chunk)
return digest.hexdigest()
def sha256_json(value: Any) -> str:
return hashlib.sha256(
json.dumps(value, ensure_ascii=False, separators=(",", ":")).encode("utf-8")
).hexdigest()
def load_encoder(path: Path) -> ModuleType:
spec = importlib.util.spec_from_file_location("frozen_encoding_dsv4", path)
if spec is None or spec.loader is None:
raise ValueError(f"cannot load DSV4 encoder: {path}")
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
if not hasattr(module, "encode_messages"):
raise ValueError(f"DSV4 encoder lacks encode_messages: {path}")
return module
def flatten_content(content: Any, *, location: str, audit: Counter[str]) -> str:
if content is None:
return ""
if isinstance(content, str):
return content
if not isinstance(content, list):
raise TypeError(f"{location}: unsupported content type {type(content).__name__}")
parts: list[str] = []
for index, part in enumerate(content):
if isinstance(part, str):
audit["content_part:string"] += 1
parts.append(part)
continue
if not isinstance(part, dict):
raise TypeError(f"{location}[{index}]: invalid content part")
kind = str(part.get("type") or "")
if kind in {"text", "input_text"} and isinstance(part.get("text"), str):
audit[f"content_part:{kind}"] += 1
parts.append(part["text"])
continue
raise ValueError(f"{location}[{index}]: unsupported content part {kind!r}")
return " ".join(parts)
def prepare_messages(
messages: list[dict[str, Any]],
*,
tools: list[dict[str, Any]],
location: str,
audit: Counter[str],
) -> list[dict[str, Any]]:
prepared = copy.deepcopy(messages)
for index, message in enumerate(prepared):
if not isinstance(message, dict):
raise TypeError(f"{location}.messages[{index}] is not an object")
role = message.get("role")
if not isinstance(role, str) or not role:
raise ValueError(f"{location}.messages[{index}] has invalid role")
if "content" in message:
message["content"] = flatten_content(
message["content"],
location=f"{location}.messages[{index}].content",
audit=audit,
)
elif role == "assistant":
message["content"] = ""
for call_index, tool_call in enumerate(message.get("tool_calls") or []):
function = tool_call.get("function") if isinstance(tool_call, dict) else None
if not isinstance(function, dict):
raise TypeError(
f"{location}.messages[{index}].tool_calls[{call_index}] is invalid"
)
arguments = function.get("arguments")
if isinstance(arguments, str):
arguments = json.loads(arguments)
function["arguments"] = arguments
if arguments is not None and not isinstance(arguments, dict):
raise TypeError(
f"{location}.messages[{index}].tool_calls[{call_index}] "
"arguments must be an object"
)
if not prepared or prepared[0].get("role") != "system":
prepared.insert(0, {"role": "system", "content": ""})
prepared[0]["tools"] = copy.deepcopy(tools)
return prepared
def converted_actor(
row: dict[str, Any],
*,
tokenizer: Any,
encoder: ModuleType,
vocab_size: int,
context_limit: int,
audit: Counter[str],
) -> tuple[dict[str, Any], dict[str, int], list[dict[str, Any]]]:
trace_id = row.get("trace_id")
if row.get("schema_version") != 2 or not isinstance(trace_id, str):
raise ValueError("input is not a schema-v2 ReplayTrace")
metadata = row.get("metadata")
if not isinstance(metadata, dict) or not isinstance(metadata.get("multiagent"), dict):
raise ValueError(f"{trace_id}: missing multiagent metadata")
output = copy.deepcopy(row)
messages = copy.deepcopy(row.get("initial_messages") or [])
tools = copy.deepcopy(row.get("tools") or [])
source_requests: list[dict[str, Any]] = []
totals: Counter[str] = Counter()
generate_index = 0
for event_index, event in enumerate(output["events"]):
kind = event.get("type")
if kind == "append_message":
messages.append(copy.deepcopy(event["message"]))
continue
if kind == "replace_context":
messages = copy.deepcopy(event["messages"])
continue
if kind == "wait":
continue
if kind != "generate":
raise ValueError(f"{trace_id}: unknown event type {kind!r}")
assistant = event.get("recorded_assistant")
source_prompt_ids = event.get("recorded_prompt_token_ids")
source_output_ids = event.get("recorded_output_token_ids")
if (
not isinstance(assistant, dict)
or not isinstance(source_prompt_ids, list)
or not source_prompt_ids
or not isinstance(source_output_ids, list)
or not source_output_ids
):
raise ValueError(
f"{trace_id}: request {generate_index} lacks complete source contract"
)
location = f"{trace_id}/request-{generate_index}"
prompt_messages = prepare_messages(
messages,
tools=tools,
location=location,
audit=audit,
)
completed_messages = prepare_messages(
messages + [assistant],
tools=tools,
location=location,
audit=audit,
)
prompt_text = encoder.encode_messages(
prompt_messages,
thinking_mode="thinking",
reasoning_effort="high",
reasoning_effort_profile="preview",
)
completed_text = encoder.encode_messages(
completed_messages,
thinking_mode="thinking",
reasoning_effort="high",
reasoning_effort_profile="preview",
)
if not completed_text.startswith(prompt_text):
raise ValueError(f"{location}: assistant is not a text continuation")
prompt_ids = tokenizer.encode(prompt_text)
completed_ids = tokenizer.encode(completed_text)
if completed_ids[: len(prompt_ids)] != prompt_ids:
raise ValueError(f"{location}: completed tokens do not extend prompt tokens")
output_ids = completed_ids[len(prompt_ids) :]
if not prompt_ids or prompt_ids[0] != tokenizer.bos_token_id:
raise ValueError(f"{location}: converted prompt lacks DSV4 BOS")
if not output_ids or output_ids[-1] != tokenizer.eos_token_id:
raise ValueError(f"{location}: converted output lacks DSV4 EOS")
if max(prompt_ids) >= vocab_size or max(output_ids) >= vocab_size:
raise ValueError(f"{location}: converted token exceeds vocab size")
if len(prompt_ids) + len(output_ids) > context_limit:
raise ValueError(
f"{location}: context {len(prompt_ids) + len(output_ids)} "
f"exceeds {context_limit}"
)
target_prompt_sha256 = sha256_json(prompt_ids)
target_output_sha256 = sha256_json(output_ids)
source_requests.append(
{
"event_index": event_index,
"generate_index": generate_index,
"source_model": event.get("source_model"),
"source_prompt_tokens": len(source_prompt_ids),
"source_output_tokens": len(source_output_ids),
"source_prompt_sha256": sha256_json(source_prompt_ids),
"source_output_sha256": sha256_json(source_output_ids),
"target_prompt_tokens": len(prompt_ids),
"target_output_tokens": len(output_ids),
"target_prompt_sha256": target_prompt_sha256,
"target_output_sha256": target_output_sha256,
}
)
event["recorded_prompt_token_ids"] = prompt_ids
event["recorded_prompt_token_count"] = len(prompt_ids)
event["recorded_prompt_token_ids_sha256"] = target_prompt_sha256
event["recorded_output_token_ids"] = output_ids
event["max_tokens"] = len(output_ids)
totals["requests"] += 1
totals["source_prompt_tokens"] += len(source_prompt_ids)
totals["source_output_tokens"] += len(source_output_ids)
totals["target_prompt_tokens"] += len(prompt_ids)
totals["target_output_tokens"] += len(output_ids)
totals["target_max_prompt_tokens"] = max(
totals["target_max_prompt_tokens"], len(prompt_ids)
)
totals["target_max_output_tokens"] = max(
totals["target_max_output_tokens"], len(output_ids)
)
messages.append(copy.deepcopy(assistant))
generate_index += 1
return output, dict(totals), source_requests
def write_root_bundle(
*,
output_dir: Path,
root_id: str,
actor_rows: list[dict[str, Any]],
actor_totals: list[dict[str, int]],
) -> dict[str, Any]:
if ROOT_ID_RE.fullmatch(root_id) is None:
raise ValueError(f"unsafe root id: {root_id!r}")
archive_path = output_dir / f"{root_id}.tar.gz"
if archive_path.exists():
raise FileExistsError(archive_path)
totals: Counter[str] = Counter()
for row in actor_totals:
for key, value in row.items():
if key.startswith("target_max_"):
totals[key] = max(totals[key], value)
else:
totals[key] += value
manifest = {
"schema_version": 1,
"root_id": root_id,
"actors": len(actor_rows),
"lead_actors": sum(
row["metadata"]["multiagent"].get("parent_trace_id") is None
for row in actor_rows
),
**dict(totals),
}
with tempfile.TemporaryDirectory(prefix=f"{root_id}-", dir=output_dir) as temp:
root = Path(temp) / root_id
root.mkdir()
actors_path = root / "actors.jsonl"
with actors_path.open("w", encoding="utf-8") as handle:
for row in actor_rows:
handle.write(
json.dumps(row, ensure_ascii=False, separators=(",", ":")) + "\n"
)
(root / "manifest.json").write_text(
json.dumps(manifest, indent=2, sort_keys=True) + "\n",
encoding="utf-8",
)
temp_archive = archive_path.with_suffix(archive_path.suffix + ".tmp")
with tarfile.open(temp_archive, "w:gz", compresslevel=1) as archive:
archive.add(root, arcname=root_id)
os.replace(temp_archive, archive_path)
manifest["archive"] = archive_path.name
manifest["archive_bytes"] = archive_path.stat().st_size
manifest["archive_sha256"] = sha256_file(archive_path)
return manifest
def rows(path: Path) -> Iterable[dict[str, Any]]:
with lzma.open(path, "rt", encoding="utf-8") as handle:
for line_number, line in enumerate(handle, 1):
try:
yield json.loads(line)
except json.JSONDecodeError as error:
raise ValueError(f"invalid JSONL row {line_number}") from error
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--input-xz", type=Path, required=True)
parser.add_argument("--output-dir", type=Path, required=True)
parser.add_argument("--target-model", type=Path, required=True)
parser.add_argument("--dsv4-encoder", type=Path, required=True)
parser.add_argument("--source-artifact-sha256", required=True)
parser.add_argument("--max-roots", type=int)
parser.add_argument("--shard-count", type=int, default=1)
parser.add_argument("--shard-index", type=int, default=0)
args = parser.parse_args()
if args.output_dir.exists():
raise FileExistsError(args.output_dir)
if args.max_roots is not None and args.max_roots <= 0:
raise ValueError("--max-roots must be positive")
if args.shard_count <= 0 or not 0 <= args.shard_index < args.shard_count:
raise ValueError("shard index must be in [0, shard count)")
args.output_dir.mkdir(parents=True)
tokenizer = AutoTokenizer.from_pretrained(
args.target_model,
trust_remote_code=True,
local_files_only=True,
)
encoder = load_encoder(args.dsv4_encoder)
model_config = json.loads((args.target_model / "config.json").read_text())
vocab_size = int(model_config["vocab_size"])
context_limit = int(model_config["max_position_embeddings"])
if not isinstance(tokenizer.eos_token_id, int):
raise TypeError("target tokenizer lacks EOS token")
audit: Counter[str] = Counter()
archives: list[dict[str, Any]] = []
current_root: str | None = None
actor_rows: list[dict[str, Any]] = []
actor_totals: list[dict[str, int]] = []
seen_roots: set[str] = set()
seen_input_roots: set[str] = set()
input_root: str | None = None
input_root_index = -1
selected = False
request_audit_rows = 0
def flush() -> bool:
nonlocal actor_rows, actor_totals, current_root
if current_root is None:
return False
archives.append(
write_root_bundle(
output_dir=args.output_dir,
root_id=current_root,
actor_rows=actor_rows,
actor_totals=actor_totals,
)
)
print(json.dumps(archives[-1], sort_keys=True), flush=True)
seen_roots.add(current_root)
actor_rows = []
actor_totals = []
current_root = None
return args.max_roots is not None and len(archives) >= args.max_roots
started = time.time()
request_audit_path = args.output_dir / "request-token-audit.jsonl.xz"
with lzma.open(
request_audit_path,
"wt",
encoding="utf-8",
preset=1,
) as request_audit:
for source_row in rows(args.input_xz):
root_id = source_row["metadata"]["multiagent"]["root_task_id"]
if root_id != input_root:
if current_root is not None and flush():
break
if root_id in seen_input_roots:
raise ValueError(f"non-contiguous root block: {root_id}")
seen_input_roots.add(root_id)
input_root = root_id
input_root_index += 1
selected = input_root_index % args.shard_count == args.shard_index
if selected:
if root_id in seen_roots:
raise ValueError(f"selected root repeated: {root_id}")
current_root = root_id
if not selected:
continue
output_row, totals, request_rows = converted_actor(
source_row,
tokenizer=tokenizer,
encoder=encoder,
vocab_size=vocab_size,
context_limit=context_limit,
audit=audit,
)
actor_rows.append(output_row)
actor_totals.append(totals)
for request_row in request_rows:
request_audit.write(
json.dumps(
{
"root_id": root_id,
"trace_id": output_row["trace_id"],
**request_row,
},
separators=(",", ":"),
sort_keys=True,
)
+ "\n"
)
request_audit_rows += 1
if current_root is not None and (
args.max_roots is None or len(archives) < args.max_roots
):
flush()
global_totals: Counter[str] = Counter()
for item in archives:
for key in (
"actors",
"lead_actors",
"requests",
"source_prompt_tokens",
"source_output_tokens",
"target_prompt_tokens",
"target_output_tokens",
):
global_totals[key] += int(item[key])
for key in ("target_max_prompt_tokens", "target_max_output_tokens"):
global_totals[key] = max(global_totals[key], int(item[key]))
seal = {
"schema_version": 1,
"kind": "morphosis-dsv4-materialization",
"created_at_s": time.time(),
"elapsed_s": time.time() - started,
"source": {
"artifact_sha256": args.source_artifact_sha256,
"logical_order_xz": str(args.input_xz),
"logical_order_xz_sha256": sha256_file(args.input_xz),
},
"target": {
"model": "deepseek-v4-flash",
"model_path": str(args.target_model),
"vocab_size": vocab_size,
"context_limit": context_limit,
"eos_token_id": tokenizer.eos_token_id,
"tokenizer_sha256": sha256_file(args.target_model / "tokenizer.json"),
"tokenizer_config_sha256": sha256_file(
args.target_model / "tokenizer_config.json"
),
"dsv4_encoder": str(args.dsv4_encoder),
"dsv4_encoder_sha256": sha256_file(args.dsv4_encoder),
"thinking_mode": "thinking",
"reasoning_effort": "high",
"reasoning_effort_profile": "preview",
},
"shard": {
"count": args.shard_count,
"index": args.shard_index,
"input_roots_seen": len(seen_input_roots),
},
"roots": len(archives),
"totals": dict(global_totals),
"content_audit": dict(sorted(audit.items())),
"request_token_audit": {
"path": request_audit_path.name,
"rows": request_audit_rows,
"sha256": sha256_file(request_audit_path),
},
"archives": archives,
}
(args.output_dir / "materialization-seal.json").write_text(
json.dumps(seal, indent=2, sort_keys=True) + "\n",
encoding="utf-8",
)
with (args.output_dir / "archives.sha256").open("w") as handle:
for item in archives:
handle.write(f"{item['archive_sha256']} {item['archive']}\n")
print(json.dumps({"seal": seal}, sort_keys=True), flush=True)
if __name__ == "__main__":
main()
|