| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Host runtime operations for local snapshots.""" |
| 4 | |
| 5 | import argparse |
| 6 | import ast |
| 7 | import hashlib |
| 8 | import importlib.metadata |
| 9 | import importlib.util |
| 10 | import json |
| 11 | import os |
| 12 | import platform |
| 13 | import select |
| 14 | import shutil |
| 15 | import signal |
| 16 | import socket |
| 17 | import stat |
| 18 | import struct |
| 19 | import subprocess |
| 20 | import sys |
| 21 | import tempfile |
| 22 | import time |
| 23 | import urllib.request |
| 24 | from collections.abc import Iterable |
| 25 | from contextlib import suppress |
| 26 | from dataclasses import dataclass |
| 27 | from datetime import datetime, timezone |
| 28 | from pathlib import Path |
| 29 | from typing import NoReturn |
| 30 | |
| 31 | import regex as re |
| 32 | from pydantic import ValidationError |
| 33 | |
| 34 | from vllm.snapshot.manifest import ( |
| 35 | ReadyMarker, |
| 36 | SnapshotManifest, |
| 37 | SnapshotRuntimeIdentity, |
| 38 | _fsync_directory, |
| 39 | _validation_path, |
| 40 | _write_json_atomic, |
| 41 | validate_artifact_root, |
| 42 | write_manifest_atomic, |
| 43 | ) |
| 44 | from vllm.snapshot.types import Oracle |
| 45 | |
| 46 | |
| 47 | class SnapshotCreateError(RuntimeError): |
| 48 | """Snapshot creation did not produce a complete artifact.""" |
| 49 | |
| 50 | |
| 51 | class SnapshotRestoreError(RuntimeError): |
| 52 | """Snapshot restore failed.""" |
| 53 | |
| 54 | |
| 55 | _COMMON_CRIU_OPTIONS = ( |
| 56 | "--shell-job", |
| 57 | "--ext-unix-sk", |
| 58 | "--tcp-established", |
| 59 | "--link-remap", |
| 60 | "--file-locks", |
| 61 | ) |
| 62 | _TCP_TABLES = ( |
| 63 | ("AF_INET", Path("/proc/net/tcp")), |
| 64 | ("AF_INET6", Path("/proc/net/tcp6")), |
| 65 | ) |
| 66 | |
| 67 | _CACHE_REMEDY = "recreate the snapshot or restore with the same cache state" |
| 68 | # Recorded instead of a digest for files whose content is volatile by design. |
| 69 | _CACHE_LOCKED = "locked" |
| 70 | # Recorded instead of a digest for a file capture found but could not read. |
| 71 | _CACHE_UNREADABLE = "unreadable" |
| 72 | |
| 73 | |
| 74 | def _generated_cache_prefixes() -> tuple[str, ...]: |
| 75 | """Path prefixes of the generated caches an engine process keeps open. |
| 76 | |
| 77 | vLLM redirects Inductor and Triton under ``VLLM_CACHE_ROOT`` when it |
| 78 | compiles, but a process can still hold their upstream defaults open. |
| 79 | A cache root is resolved and separator-terminated so that it still matches |
| 80 | the kernel-canonical descriptor targets under a symlinked root, and so that |
| 81 | a sibling directory sharing its name does not. Inductor's default location |
| 82 | is a bare prefix instead, because its directory carries a per-user suffix. |
| 83 | """ |
| 84 | import vllm.envs as envs |
| 85 | |
| 86 | inductor_root = os.environ.get("TORCHINDUCTOR_CACHE_DIR") |
| 87 | roots = tuple( |
| 88 | f"{os.path.realpath(os.path.expanduser(root)).rstrip(os.sep)}{os.sep}" |
| 89 | for root in ( |
| 90 | envs.VLLM_CACHE_ROOT, |
| 91 | os.environ.get("TRITON_CACHE_DIR") or "~/.triton/cache", |
| 92 | "~/.cache/flashinfer", |
| 93 | *((inductor_root,) if inductor_root else ()), |
| 94 | ) |
| 95 | ) |
| 96 | if inductor_root: |
| 97 | return roots |
| 98 | return roots + ( |
| 99 | os.path.join(os.path.realpath(tempfile.gettempdir()), "torchinductor_"), |
| 100 | ) |
| 101 | |
| 102 | |
| 103 | def _error_detail(error: BaseException) -> str: |
| 104 | message = str(error) |
| 105 | return f"{type(error).__name__}: {message}" if message else type(error).__name__ |
| 106 | |
| 107 | |
| 108 | def _close_pidfds(handles: Iterable[tuple[int, int]]) -> None: |
| 109 | for _pid, pidfd in handles: |
| 110 | with suppress(OSError): |
| 111 | os.close(pidfd) |
| 112 | |
| 113 | |
| 114 | @dataclass(frozen=True) |
| 115 | class ProcessInventory: |
| 116 | root_pid: int |
| 117 | process_tree: tuple[int, ...] |
| 118 | cuda_holders: tuple[int, ...] |
| 119 | gpu_uuid: str |
| 120 | external_cache_files: tuple[tuple[str, str], ...] = () |
| 121 | |
| 122 | |
| 123 | @dataclass(frozen=True) |
| 124 | class _TcpSocketRecord: |
| 125 | family: str |
| 126 | local_raw: str |
| 127 | remote_raw: str |
| 128 | inode: int |
| 129 | |
| 130 | |
| 131 | def _process_comm(pid: int) -> str: |
| 132 | try: |
| 133 | return (Path("/proc") / str(pid) / "comm").read_text().strip() |
| 134 | except OSError: |
| 135 | return "unknown" |
| 136 | |
| 137 | |
| 138 | def _decode_endpoint(family: str, raw: str) -> str: |
| 139 | """Render a ``/proc/net/tcp`` ``HEXADDR:HEXPORT`` field as ``ip:port``. |
| 140 | |
| 141 | The kernel prints the address as little-endian 32-bit words: one for IPv4, |
| 142 | four in order for IPv6. |
| 143 | """ |
| 144 | hex_address, _, hex_port = raw.partition(":") |
| 145 | port = int(hex_port, 16) |
| 146 | if family == "AF_INET": |
| 147 | packed = struct.pack("<I", int(hex_address, 16)) |
| 148 | return f"{socket.inet_ntop(socket.AF_INET, packed)}:{port}" |
| 149 | packed = struct.pack( |
| 150 | "<4I", *(int(hex_address[index : index + 8], 16) for index in (0, 8, 16, 24)) |
| 151 | ) |
| 152 | return f"[{socket.inet_ntop(socket.AF_INET6, packed)}]:{port}" |
| 153 | |
| 154 | |
| 155 | def _validate_tcp_connections( |
| 156 | records: tuple[_TcpSocketRecord, ...], socket_owners: dict[int, int] |
| 157 | ) -> None: |
| 158 | owned = tuple(record for record in records if record.inode in socket_owners) |
| 159 | endpoints = { |
| 160 | (record.family, record.local_raw, record.remote_raw) for record in owned |
| 161 | } |
| 162 | offenders = tuple( |
| 163 | record |
| 164 | for record in owned |
| 165 | if (record.family, record.remote_raw, record.local_raw) not in endpoints |
| 166 | ) |
| 167 | if not offenders: |
| 168 | return |
| 169 | named = "; ".join( |
| 170 | f"pid {socket_owners[record.inode]} " |
| 171 | f"({_process_comm(socket_owners[record.inode])}) " |
| 172 | f"{_decode_endpoint(record.family, record.local_raw)} -> " |
| 173 | f"{_decode_endpoint(record.family, record.remote_raw)}" |
| 174 | for record in offenders[:5] |
| 175 | ) |
| 176 | remaining = f" (+{len(offenders) - 5} more)" if len(offenders) > 5 else "" |
| 177 | raise SnapshotCreateError( |
| 178 | f"snapshot tree has an external established TCP connection: {named}{remaining}" |
| 179 | ) |
| 180 | |
| 181 | |
| 182 | class LocalSnapshotTools: |
| 183 | """Host-owned CRIU and cuda-checkpoint implementation.""" |
| 184 | |
| 185 | def __init__(self) -> None: |
| 186 | self.criu = shutil.which("criu") or "criu" |
| 187 | self.cuda_checkpoint = shutil.which("cuda-checkpoint") or "cuda-checkpoint" |
| 188 | self.nvidia_smi = shutil.which("nvidia-smi") or "nvidia-smi" |
| 189 | plugin_dir = os.environ.get("CRIU_CUDA_PLUGIN_DIR", "") |
| 190 | self.plugin_dir = Path(plugin_dir) if plugin_dir else None |
| 191 | self.shm_dir = Path("/dev/shm") |
| 192 | self.timeout_s = float(os.environ.get("VLLM_SNAPSHOT_TIMEOUT_S", "900")) |
| 193 | if not 0 < self.timeout_s < float("inf"): |
| 194 | raise ValueError("VLLM_SNAPSHOT_TIMEOUT_S must be a positive finite number") |
| 195 | self._children: dict[int, subprocess.Popen[bytes]] = {} |
| 196 | self._restored_processes: dict[int, tuple[tuple[int, int], ...]] = {} |
| 197 | |
| 198 | def _privileged(self) -> list[str]: |
| 199 | return [] if os.geteuid() == 0 else ["sudo", "-n"] |
| 200 | |
| 201 | def _run( |
| 202 | self, |
| 203 | command: list[str], |
| 204 | *, |
| 205 | timeout: float | None = None, |
| 206 | ) -> subprocess.CompletedProcess[str]: |
| 207 | return subprocess.run( |
| 208 | command, |
| 209 | check=True, |
| 210 | capture_output=True, |
| 211 | text=True, |
| 212 | timeout=timeout or self.timeout_s, |
| 213 | ) |
| 214 | |
| 215 | def preflight(self, action: str, artifact: Path) -> None: |
| 216 | if platform.system() != "Linux" or platform.machine() != "x86_64": |
| 217 | raise RuntimeError("snapshot requires Linux x86_64") |
| 218 | if action == "restore" and ( |
| 219 | not callable(getattr(os, "pidfd_open", None)) |
| 220 | or not callable(getattr(signal, "pidfd_send_signal", None)) |
| 221 | ): |
| 222 | raise RuntimeError("snapshot requires Linux pidfd support") |
| 223 | for command in (self.criu, self.cuda_checkpoint, self.nvidia_smi): |
| 224 | if shutil.which(command) is None: |
| 225 | raise RuntimeError(f"snapshot dependency not found: {command}") |
| 226 | if self.plugin_dir is None: |
| 227 | raise RuntimeError("CRIU_CUDA_PLUGIN_DIR is required") |
| 228 | if not (self.plugin_dir / "cuda_plugin.so").is_file(): |
| 229 | raise RuntimeError(f"CRIU CUDA plugin not found in {self.plugin_dir}") |
| 230 | if os.geteuid() != 0: |
| 231 | self._run(["sudo", "-n", "true"], timeout=5) |
| 232 | if action == "restore": |
| 233 | validate_artifact_root(artifact, creating=False) |
| 234 | else: |
| 235 | validate_artifact_root(artifact, creating=True) |
| 236 | |
| 237 | def launch_child( |
| 238 | self, |
| 239 | workdir: Path, |
| 240 | engine_argv: tuple[str, ...], |
| 241 | ) -> int: |
| 242 | command = [ |
| 243 | sys.executable, |
| 244 | "-m", |
| 245 | "vllm.snapshot.server", |
| 246 | "--ready-file", |
| 247 | str(workdir / "ready.json"), |
| 248 | "--release-file", |
| 249 | str(workdir / "release.json"), |
| 250 | "--release-timeout-s", |
| 251 | str(self.timeout_s), |
| 252 | "--", |
| 253 | *engine_argv, |
| 254 | ] |
| 255 | child_environment = os.environ.copy() |
| 256 | child_environment.setdefault("VLLM_WORKER_MULTIPROC_METHOD", "spawn") |
| 257 | with (workdir / "child.log").open("wb") as log_file: |
| 258 | process = subprocess.Popen( |
| 259 | command, |
| 260 | env=child_environment, |
| 261 | stdin=subprocess.DEVNULL, |
| 262 | stdout=log_file, |
| 263 | stderr=subprocess.STDOUT, |
| 264 | start_new_session=True, |
| 265 | ) |
| 266 | self._children[process.pid] = process |
| 267 | return process.pid |
| 268 | |
| 269 | def wait_ready(self, workdir: Path, root_pid: int) -> Oracle: |
| 270 | ready_file = workdir / "ready.json" |
| 271 | deadline = time.monotonic() + self.timeout_s |
| 272 | waitid = os.waitid # type: ignore[attr-defined] |
| 273 | while time.monotonic() < deadline: |
| 274 | if ready_file.is_file() and ready_file.stat().st_size: |
| 275 | try: |
| 276 | marker = ReadyMarker.model_validate_json(ready_file.read_text()) |
| 277 | except ValidationError as error: |
| 278 | raise SnapshotCreateError( |
| 279 | f"snapshot ready marker invalid: {_validation_path(error)}" |
| 280 | ) from error |
| 281 | return Oracle( |
| 282 | token_ids=marker.token_ids, |
| 283 | text=marker.text, |
| 284 | sampled_token_logprob=marker.sampled_token_logprob, |
| 285 | ) |
| 286 | process = self._children.get(root_pid) |
| 287 | if process is not None and waitid( |
| 288 | os.P_PID, |
| 289 | root_pid, |
| 290 | os.WEXITED | os.WNOHANG | os.WNOWAIT, |
| 291 | ): |
| 292 | log = (workdir / "child.log").read_text(errors="replace") |
| 293 | raise SnapshotCreateError(f"snapshot child exited before ready:\n{log}") |
| 294 | try: |
| 295 | os.kill(root_pid, 0) |
| 296 | except ProcessLookupError as error: |
| 297 | log = (workdir / "child.log").read_text(errors="replace") |
| 298 | raise SnapshotCreateError( |
| 299 | f"snapshot child exited before ready:\n{log}" |
| 300 | ) from error |
| 301 | time.sleep(0.05) |
| 302 | raise SnapshotCreateError("snapshot child did not become ready before timeout") |
| 303 | |
| 304 | def _tree_pids(self, root_pid: int) -> tuple[int, ...]: |
| 305 | rows = self._run(["ps", "-eo", "pid=,ppid="]).stdout.splitlines() |
| 306 | children: dict[int, list[int]] = {} |
| 307 | for row in rows: |
| 308 | pid, parent_pid = map(int, row.split()) |
| 309 | children.setdefault(parent_pid, []).append(pid) |
| 310 | stack = [root_pid] |
| 311 | found: list[int] = [] |
| 312 | while stack: |
| 313 | pid = stack.pop() |
| 314 | if pid in found: |
| 315 | continue |
| 316 | found.append(pid) |
| 317 | stack.extend(children.get(pid, [])) |
| 318 | return tuple(found) |
| 319 | |
| 320 | def _cuda_process_rows(self) -> tuple[str, ...]: |
| 321 | output = self._run( |
| 322 | [ |
| 323 | self.nvidia_smi, |
| 324 | "--query-compute-apps=pid,gpu_uuid", |
| 325 | "--format=csv,noheader,nounits", |
| 326 | ] |
| 327 | ).stdout |
| 328 | return tuple(line for line in output.splitlines() if line.strip()) |
| 329 | |
| 330 | def _cuda_pids(self, rows: tuple[str, ...]) -> tuple[int, ...]: |
| 331 | return tuple(sorted({int(line.split(",", 1)[0].strip()) for line in rows})) |
| 332 | |
| 333 | def _descriptor_targets(self, pid: int) -> tuple[str, ...]: |
| 334 | descriptor_dir = Path("/proc") / str(pid) / "fd" |
| 335 | try: |
| 336 | descriptors = tuple(descriptor_dir.iterdir()) |
| 337 | except FileNotFoundError as error: |
| 338 | raise SnapshotCreateError( |
| 339 | f"snapshot process exited during descriptor inventory: {pid}" |
| 340 | ) from error |
| 341 | targets: list[str] = [] |
| 342 | for descriptor in descriptors: |
| 343 | try: |
| 344 | targets.append(os.readlink(descriptor)) |
| 345 | except FileNotFoundError: |
| 346 | continue |
| 347 | return tuple(targets) |
| 348 | |
| 349 | def _descriptor_inventory( |
| 350 | self, process_tree: tuple[int, ...] |
| 351 | ) -> tuple[tuple[int, ...], dict[int, int]]: |
| 352 | """Collect io_uring holders and each socket inode's owning pid.""" |
| 353 | io_uring_pids: list[int] = [] |
| 354 | socket_owners: dict[int, int] = {} |
| 355 | for pid in process_tree: |
| 356 | targets = self._descriptor_targets(pid) |
| 357 | if "anon_inode:[io_uring]" in targets: |
| 358 | io_uring_pids.append(pid) |
| 359 | for target in targets: |
| 360 | if target.startswith("socket:[") and target.endswith("]"): |
| 361 | # Shared descriptors name one holder; any of them locates |
| 362 | # the connection for the operator. |
| 363 | socket_owners.setdefault(int(target[len("socket:[") : -1]), pid) |
| 364 | return tuple(io_uring_pids), socket_owners |
| 365 | |
| 366 | def _tcp_records(self) -> tuple[_TcpSocketRecord, ...]: |
| 367 | records: list[_TcpSocketRecord] = [] |
| 368 | for family, table in _TCP_TABLES: |
| 369 | if not table.exists(): |
| 370 | continue |
| 371 | for row in table.read_text().splitlines()[1:]: |
| 372 | fields = row.split() |
| 373 | if fields[3] != "01": |
| 374 | continue |
| 375 | records.append( |
| 376 | _TcpSocketRecord( |
| 377 | family=family, |
| 378 | local_raw=fields[1], |
| 379 | remote_raw=fields[2], |
| 380 | inode=int(fields[9]), |
| 381 | ) |
| 382 | ) |
| 383 | return tuple(records) |
| 384 | |
| 385 | def _cache_fingerprint(self, path: Path) -> str | None: |
| 386 | """Digest a cache file, or mark the ones whose content is volatile.""" |
| 387 | try: |
| 388 | if not stat.S_ISREG(path.stat().st_mode): |
| 389 | return None |
| 390 | except OSError: |
| 391 | return None |
| 392 | if path.name.endswith(".lock"): |
| 393 | return _CACHE_LOCKED |
| 394 | try: |
| 395 | return self._sha256(path) |
| 396 | except OSError: |
| 397 | # Restore can only check that this one still exists, so record the |
| 398 | # downgrade rather than passing it off as a volatile file. |
| 399 | return _CACHE_UNREADABLE |
| 400 | |
| 401 | def _external_cache_files( |
| 402 | self, process_tree: tuple[int, ...], artifact: Path |
| 403 | ) -> tuple[tuple[str, str], ...]: |
| 404 | """Fingerprint the generated-cache files the captured tree holds open. |
| 405 | |
| 406 | CRIU reopens these by path on restore, so a JIT cache that rotates |
| 407 | between capture and restore fails the restore permanently. Recording |
| 408 | them lets restore fail early and name the file that changed. Files |
| 409 | inside the artifact are excluded: they belong to the snapshot itself. |
| 410 | Deleted targets are excluded too, because CRIU remaps those links. |
| 411 | """ |
| 412 | prefixes = _generated_cache_prefixes() |
| 413 | artifact_prefix = f"{os.path.abspath(artifact)}{os.sep}" |
| 414 | candidates: set[str] = set() |
| 415 | for pid in process_tree: |
| 416 | for target in self._descriptor_targets(pid): |
| 417 | if target.endswith(" (deleted)") or target.startswith(artifact_prefix): |
| 418 | continue |
| 419 | if target.endswith(".lock") or target.startswith(prefixes): |
| 420 | candidates.add(target) |
| 421 | recorded = ( |
| 422 | (path, self._cache_fingerprint(Path(path))) for path in sorted(candidates) |
| 423 | ) |
| 424 | return tuple( |
| 425 | (path, fingerprint) |
| 426 | for path, fingerprint in recorded |
| 427 | if fingerprint is not None |
| 428 | ) |
| 429 | |
| 430 | def inventory(self, root_pid: int, artifact: Path) -> ProcessInventory: |
| 431 | process_tree = self._tree_pids(root_pid) |
| 432 | cuda_rows = self._cuda_process_rows() |
| 433 | cuda_holders = tuple( |
| 434 | pid for pid in self._cuda_pids(cuda_rows) if pid in process_tree |
| 435 | ) |
| 436 | if not cuda_holders: |
| 437 | raise SnapshotCreateError("snapshot tree has no CUDA-holding process") |
| 438 | io_uring_pids, socket_owners = self._descriptor_inventory(process_tree) |
| 439 | if io_uring_pids: |
| 440 | raise SnapshotCreateError( |
| 441 | "snapshot tree owns io_uring state that CRIU cannot dump; set " |
| 442 | "kernel.io_uring_disabled=1 before starting an unprivileged " |
| 443 | "vLLM process, or use =2 to disable it host-wide " |
| 444 | f"(pids: {', '.join(map(str, io_uring_pids))})" |
| 445 | ) |
| 446 | _validate_tcp_connections(self._tcp_records(), socket_owners) |
| 447 | gpu_uuid = self._gpu_uuid_for_pids(cuda_holders, cuda_rows) |
| 448 | return ProcessInventory( |
| 449 | root_pid=root_pid, |
| 450 | process_tree=process_tree, |
| 451 | cuda_holders=cuda_holders, |
| 452 | gpu_uuid=gpu_uuid, |
| 453 | external_cache_files=self._external_cache_files(process_tree, artifact), |
| 454 | ) |
| 455 | |
| 456 | def _gpu_uuid_for_pids(self, pids: tuple[int, ...], rows: tuple[str, ...]) -> str: |
| 457 | wanted = set(pids) |
| 458 | matches: set[str] = set() |
| 459 | try: |
| 460 | for line in rows: |
| 461 | pid, gpu_uuid = (item.strip() for item in line.split(",", 1)) |
| 462 | if int(pid) in wanted: |
| 463 | matches.add(gpu_uuid) |
| 464 | except (TypeError, ValueError) as error: |
| 465 | raise SnapshotCreateError( |
| 466 | "could not identify the GPU used by the snapshot process tree" |
| 467 | ) from error |
| 468 | if len(matches) != 1: |
| 469 | raise SnapshotCreateError( |
| 470 | "snapshot process tree must use exactly one identifiable GPU" |
| 471 | ) |
| 472 | return matches.pop() |
| 473 | |
| 474 | def _criu( |
| 475 | self, |
| 476 | action: str, |
| 477 | artifact: Path, |
| 478 | arguments: list[str], |
| 479 | ) -> None: |
| 480 | if self.plugin_dir is None: |
| 481 | raise RuntimeError("CRIU_CUDA_PLUGIN_DIR is required") |
| 482 | env = os.environ.copy() |
| 483 | env["PATH"] = f"{Path(self.cuda_checkpoint).parent}:{env['PATH']}" |
| 484 | env["VLLM_SNAPSHOT_IGNORED_FD_MAP"] = str(artifact / "ignored-nvidia-fds.tsv") |
| 485 | env["RIG_CRIU_IGNORED_FD_MAP"] = env["VLLM_SNAPSHOT_IGNORED_FD_MAP"] |
| 486 | command = [ |
| 487 | *self._privileged(), |
| 488 | "env", |
| 489 | f"PATH={env['PATH']}", |
| 490 | f"VLLM_SNAPSHOT_IGNORED_FD_MAP={env['VLLM_SNAPSHOT_IGNORED_FD_MAP']}", |
| 491 | f"RIG_CRIU_IGNORED_FD_MAP={env['RIG_CRIU_IGNORED_FD_MAP']}", |
| 492 | self.criu, |
| 493 | action, |
| 494 | "-L", |
| 495 | str(self.plugin_dir), |
| 496 | *arguments, |
| 497 | ] |
| 498 | self._run(command) |
| 499 | |
| 500 | def _record_child_log_size(self, artifact: Path) -> None: |
| 501 | child_log = artifact / "child.log" |
| 502 | size_path = artifact / "child.log.snapshot-size" |
| 503 | payload = f"{child_log.stat().st_size}\n".encode() |
| 504 | descriptor = os.open( |
| 505 | size_path, |
| 506 | os.O_WRONLY | os.O_CREAT | os.O_EXCL | getattr(os, "O_NOFOLLOW", 0), |
| 507 | 0o600, |
| 508 | ) |
| 509 | with os.fdopen(descriptor, "wb") as output: |
| 510 | output.write(payload) |
| 511 | output.flush() |
| 512 | os.fsync(output.fileno()) |
| 513 | |
| 514 | def _link_remap_names(self) -> set[str]: |
| 515 | return { |
| 516 | path.name |
| 517 | for path in self.shm_dir.glob("link_remap.*") |
| 518 | if path.name.removeprefix("link_remap.").isdigit() |
| 519 | } |
| 520 | |
| 521 | def _capture_link_remaps(self, artifact: Path, names: set[str]) -> None: |
| 522 | if not names: |
| 523 | return |
| 524 | destination_dir = artifact / "link-remaps" |
| 525 | destination_dir.mkdir(mode=0o700) |
| 526 | for name in sorted(names): |
| 527 | source = self.shm_dir / name |
| 528 | source_fd = os.open( |
| 529 | source, |
| 530 | os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0), |
| 531 | ) |
| 532 | try: |
| 533 | metadata = os.fstat(source_fd) |
| 534 | if not stat.S_ISREG(metadata.st_mode) or metadata.st_uid != os.getuid(): |
| 535 | raise SnapshotCreateError(f"invalid CRIU link remap: {source}") |
| 536 | destination_fd = os.open( |
| 537 | destination_dir / name, |
| 538 | os.O_WRONLY | os.O_CREAT | os.O_EXCL | getattr(os, "O_NOFOLLOW", 0), |
| 539 | 0o600, |
| 540 | ) |
| 541 | with os.fdopen(destination_fd, "wb") as destination: |
| 542 | with os.fdopen(os.dup(source_fd), "rb") as contents: |
| 543 | shutil.copyfileobj(contents, destination) |
| 544 | destination.flush() |
| 545 | os.fsync(destination.fileno()) |
| 546 | finally: |
| 547 | os.close(source_fd) |
| 548 | source.unlink() |
| 549 | |
| 550 | def _rollback_link_remaps(self, created: dict[str, tuple[int, int]]) -> None: |
| 551 | for name, identity in created.items(): |
| 552 | target = self.shm_dir / name |
| 553 | try: |
| 554 | metadata = target.lstat() |
| 555 | except FileNotFoundError: |
| 556 | continue |
| 557 | if (metadata.st_dev, metadata.st_ino) == identity: |
| 558 | target.unlink() |
| 559 | |
| 560 | def _restore_link_remaps(self, artifact: Path) -> dict[str, tuple[int, int]]: |
| 561 | source_dir = artifact / "link-remaps" |
| 562 | if not source_dir.exists(): |
| 563 | return {} |
| 564 | created: dict[str, tuple[int, int]] = {} |
| 565 | try: |
| 566 | for source in sorted(source_dir.iterdir()): |
| 567 | if ( |
| 568 | not source.name.removeprefix("link_remap.").isdigit() |
| 569 | or source.is_symlink() |
| 570 | or not source.is_file() |
| 571 | ): |
| 572 | raise SnapshotRestoreError(f"invalid saved link remap: {source}") |
| 573 | payload = source.read_bytes() |
| 574 | target = self.shm_dir / source.name |
| 575 | try: |
| 576 | descriptor = os.open( |
| 577 | target, |
| 578 | os.O_WRONLY |
| 579 | | os.O_CREAT |
| 580 | | os.O_EXCL |
| 581 | | getattr(os, "O_NOFOLLOW", 0), |
| 582 | 0o600, |
| 583 | ) |
| 584 | except FileExistsError as error: |
| 585 | try: |
| 586 | descriptor = os.open( |
| 587 | target, |
| 588 | os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0), |
| 589 | ) |
| 590 | except OSError: |
| 591 | raise SnapshotRestoreError( |
| 592 | f"conflicting CRIU link remap exists: {target}" |
| 593 | ) from error |
| 594 | with os.fdopen(descriptor, "rb") as existing: |
| 595 | metadata = os.fstat(existing.fileno()) |
| 596 | if ( |
| 597 | not stat.S_ISREG(metadata.st_mode) |
| 598 | or metadata.st_uid != os.getuid() |
| 599 | or existing.read() != payload |
| 600 | ): |
| 601 | raise SnapshotRestoreError( |
| 602 | f"conflicting CRIU link remap exists: {target}" |
| 603 | ) from error |
| 604 | continue |
| 605 | metadata = os.fstat(descriptor) |
| 606 | created[source.name] = (metadata.st_dev, metadata.st_ino) |
| 607 | with os.fdopen(descriptor, "wb") as output: |
| 608 | output.write(payload) |
| 609 | output.flush() |
| 610 | os.fsync(output.fileno()) |
| 611 | except BaseException: |
| 612 | self._rollback_link_remaps(created) |
| 613 | raise |
| 614 | return created |
| 615 | |
| 616 | def _reset_child_log(self, artifact: Path) -> None: |
| 617 | size_path = artifact / "child.log.snapshot-size" |
| 618 | descriptor = os.open( |
| 619 | size_path, |
| 620 | os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0), |
| 621 | ) |
| 622 | with os.fdopen(descriptor, "r") as source: |
| 623 | size = int(source.read().strip()) |
| 624 | if size < 0: |
| 625 | raise SnapshotRestoreError("captured child log size is invalid") |
| 626 | |
| 627 | child_log = artifact / "child.log" |
| 628 | descriptor = os.open( |
| 629 | child_log, |
| 630 | os.O_WRONLY | getattr(os, "O_NOFOLLOW", 0), |
| 631 | ) |
| 632 | try: |
| 633 | metadata = os.fstat(descriptor) |
| 634 | if not stat.S_ISREG(metadata.st_mode) or metadata.st_size < size: |
| 635 | raise SnapshotRestoreError("captured child log is invalid") |
| 636 | os.ftruncate(descriptor, size) |
| 637 | os.fsync(descriptor) |
| 638 | finally: |
| 639 | os.close(descriptor) |
| 640 | |
| 641 | def dump(self, workdir: Path, inventory: ProcessInventory) -> None: |
| 642 | images = workdir / "images" |
| 643 | images.mkdir(mode=0o700) |
| 644 | fd_map = workdir / "ignored-nvidia-fds.tsv" |
| 645 | fd_map.touch(mode=0o600) |
| 646 | remaps_before = self._link_remap_names() |
| 647 | try: |
| 648 | self._criu( |
| 649 | "dump", |
| 650 | workdir, |
| 651 | [ |
| 652 | "--tree", |
| 653 | str(inventory.root_pid), |
| 654 | "--images-dir", |
| 655 | str(images), |
| 656 | "--log-file", |
| 657 | "dump.log", |
| 658 | *_COMMON_CRIU_OPTIONS, |
| 659 | ], |
| 660 | ) |
| 661 | self._capture_link_remaps( |
| 662 | workdir, |
| 663 | self._link_remap_names() - remaps_before, |
| 664 | ) |
| 665 | except BaseException: |
| 666 | for name in self._link_remap_names() - remaps_before: |
| 667 | (self.shm_dir / name).unlink(missing_ok=True) |
| 668 | raise |
| 669 | self._record_child_log_size(workdir) |
| 670 | |
| 671 | def verify_dead(self, inventory: ProcessInventory) -> None: |
| 672 | process = self._children.get(inventory.root_pid) |
| 673 | if process is not None: |
| 674 | process.wait(timeout=10) |
| 675 | for pid in inventory.process_tree: |
| 676 | try: |
| 677 | os.kill(pid, 0) |
| 678 | except ProcessLookupError: |
| 679 | continue |
| 680 | raise SnapshotCreateError(f"process survived CRIU dump: {pid}") |
| 681 | self._children.pop(inventory.root_pid, None) |
| 682 | |
| 683 | def _version(self, command: list[str]) -> str: |
| 684 | try: |
| 685 | result = self._run(command, timeout=10) |
| 686 | except subprocess.CalledProcessError as error: |
| 687 | return (error.stdout or error.stderr).splitlines()[0] |
| 688 | return result.stdout.splitlines()[0] if result.stdout else "unknown" |
| 689 | |
| 690 | def _sha256(self, path: Path) -> str: |
| 691 | digest = hashlib.sha256() |
| 692 | with path.open("rb") as source: |
| 693 | for chunk in iter(lambda: source.read(1024 * 1024), b""): |
| 694 | digest.update(chunk) |
| 695 | return digest.hexdigest() |
| 696 | |
| 697 | def _plugin_identity(self) -> tuple[tuple[str, str], ...]: |
| 698 | if self.plugin_dir is None: |
| 699 | return () |
| 700 | return tuple( |
| 701 | (f"snapshot_plugin_sha256:{path.name}", self._sha256(path)) |
| 702 | for path in sorted(self.plugin_dir.glob("*.so")) |
| 703 | ) |
| 704 | |
| 705 | def _torch_identity(self) -> tuple[str, str]: |
| 706 | torch_version = importlib.metadata.version("torch") |
| 707 | torch_spec = importlib.util.find_spec("torch") |
| 708 | locations = torch_spec and torch_spec.submodule_search_locations |
| 709 | if not locations: |
| 710 | raise RuntimeError("installed torch package could not be located") |
| 711 | version_path = Path(next(iter(locations))) / "version.py" |
| 712 | # Parse the version file rather than importing torch: snapshot create's |
| 713 | # CLI path stays import-light (vllm/entrypoints/cli/snapshot.py). |
| 714 | syntax = ast.parse(version_path.read_text(), filename=str(version_path)) |
| 715 | cuda_runtime = "" |
| 716 | for statement in syntax.body: |
| 717 | target = None |
| 718 | value = None |
| 719 | if isinstance(statement, ast.Assign) and len(statement.targets) == 1: |
| 720 | target = statement.targets[0] |
| 721 | value = statement.value |
| 722 | elif isinstance(statement, ast.AnnAssign): |
| 723 | target = statement.target |
| 724 | value = statement.value |
| 725 | if isinstance(target, ast.Name) and target.id == "cuda" and value: |
| 726 | parsed = ast.literal_eval(value) |
| 727 | cuda_runtime = str(parsed or "") |
| 728 | break |
| 729 | return torch_version, cuda_runtime |
| 730 | |
| 731 | def _gpu_identity(self, gpu_uuid: str) -> tuple[str, str, str]: |
| 732 | output = self._run( |
| 733 | [ |
| 734 | self.nvidia_smi, |
| 735 | "--query-gpu=name,uuid,driver_version", |
| 736 | "--format=csv,noheader,nounits", |
| 737 | f"--id={gpu_uuid}", |
| 738 | ] |
| 739 | ).stdout.strip() |
| 740 | name, uuid, driver = (item.strip() for item in output.split(",", 2)) |
| 741 | return name, uuid, driver |
| 742 | |
| 743 | def _artifact_bytes(self, workdir: Path) -> int: |
| 744 | return sum(path.stat().st_size for path in workdir.rglob("*") if path.is_file()) |
| 745 | |
| 746 | def _environment_identity(self) -> tuple[tuple[str, str], ...]: |
| 747 | prefixes = ("VLLM_", "CUDA_", "NCCL_", "TORCH_", "TRITON_") |
| 748 | environment = os.environ.copy() |
| 749 | environment.setdefault("VLLM_WORKER_MULTIPROC_METHOD", "spawn") |
| 750 | selected = tuple( |
| 751 | sorted( |
| 752 | (key, hashlib.sha256(key.encode() + b"\0" + value.encode()).hexdigest()) |
| 753 | for key, value in environment.items() |
| 754 | if key.startswith(prefixes) |
| 755 | and key not in {"VLLM_API_KEY", "VLLM_SNAPSHOT_TIMEOUT_S"} |
| 756 | ) |
| 757 | ) |
| 758 | return selected + self._plugin_identity() |
| 759 | |
| 760 | def make_manifest( |
| 761 | self, |
| 762 | args: argparse.Namespace, |
| 763 | engine_argv: tuple[str, ...], |
| 764 | inventory: ProcessInventory, |
| 765 | oracle: Oracle, |
| 766 | workdir: Path, |
| 767 | ) -> SnapshotManifest: |
| 768 | identity = self.current_identity(inventory.gpu_uuid) |
| 769 | redacted_argv = list(engine_argv) |
| 770 | redact_api_keys = False |
| 771 | for index, argument in enumerate(redacted_argv): |
| 772 | option, separator, _ = argument.partition("=") |
| 773 | normalized_option = option.replace("_", "-") |
| 774 | # Parsed argv can contain unambiguous long-option abbreviations. |
| 775 | if len(normalized_option) > 2 and "--api-key".startswith(normalized_option): |
| 776 | redact_api_keys = not separator |
| 777 | if separator: |
| 778 | redacted_argv[index] = f"{option}=***" |
| 779 | elif len(normalized_option) > 2 and "--hf-token".startswith( |
| 780 | normalized_option |
| 781 | ): |
| 782 | if separator: |
| 783 | redacted_argv[index] = f"{option}=***" |
| 784 | elif index + 1 < len(redacted_argv): |
| 785 | redacted_argv[index + 1] = "***" |
| 786 | redact_api_keys = False |
| 787 | elif argument.startswith("-"): |
| 788 | redact_api_keys = False |
| 789 | elif redact_api_keys: |
| 790 | redacted_argv[index] = "***" |
| 791 | engine_argv = tuple(redacted_argv) |
| 792 | revision = str(getattr(args, "revision", None) or "") |
| 793 | tokenizer_revision = str(getattr(args, "tokenizer_revision", None) or revision) |
| 794 | source_model = str(getattr(args, "model_tag", None) or args.model) |
| 795 | from vllm.config.model import get_served_model_name |
| 796 | |
| 797 | served_model_name = get_served_model_name( |
| 798 | source_model, getattr(args, "served_model_name", None) |
| 799 | ) |
| 800 | return SnapshotManifest( |
| 801 | schema_version=1, |
| 802 | boundary="post-engine-init-reloadable-state-released", |
| 803 | created_at=datetime.now(timezone.utc).isoformat(), |
| 804 | artifact_bytes=self._artifact_bytes(workdir), |
| 805 | **identity.model_dump(), |
| 806 | model=source_model, |
| 807 | served_model_name=str(served_model_name), |
| 808 | model_revision=revision, |
| 809 | tokenizer_revision=tokenizer_revision, |
| 810 | engine_argv=engine_argv, |
| 811 | process_tree=inventory.process_tree, |
| 812 | cuda_holders=inventory.cuda_holders, |
| 813 | external_cache_files=inventory.external_cache_files, |
| 814 | oracle_token_ids=oracle.token_ids, |
| 815 | oracle_text=oracle.text, |
| 816 | oracle_sampled_token_logprob=oracle.sampled_token_logprob, |
| 817 | ) |
| 818 | |
| 819 | def publish(self, workdir: Path, manifest: SnapshotManifest) -> None: |
| 820 | write_manifest_atomic(workdir, manifest) |
| 821 | _fsync_directory(workdir.parent) |
| 822 | |
| 823 | def current_identity(self, gpu_uuid: str) -> SnapshotRuntimeIdentity: |
| 824 | gpu_name, gpu_uuid, driver_version = self._gpu_identity(gpu_uuid) |
| 825 | torch_version, cuda_runtime = self._torch_identity() |
| 826 | host_id = Path("/etc/machine-id").read_text().strip() |
| 827 | return SnapshotRuntimeIdentity( |
| 828 | vllm_version=importlib.metadata.version("vllm"), |
| 829 | python_version=platform.python_version(), |
| 830 | torch_version=torch_version, |
| 831 | cuda_runtime=cuda_runtime, |
| 832 | driver_version=driver_version, |
| 833 | criu_version=self._version([self.criu, "--version"]), |
| 834 | cuda_checkpoint_sha256=self._sha256(Path(self.cuda_checkpoint)), |
| 835 | kernel_release=platform.release(), |
| 836 | host_id=host_id, |
| 837 | gpu_name=gpu_name, |
| 838 | gpu_uuid=gpu_uuid, |
| 839 | environment=self._environment_identity(), |
| 840 | ) |
| 841 | |
| 842 | def _read_restored_pid(self, pidfile: Path) -> int: |
| 843 | try: |
| 844 | payload = self._run([*self._privileged(), "cat", str(pidfile)]).stdout |
| 845 | except Exception as error: |
| 846 | raise SnapshotRestoreError("restored PID file is missing") from error |
| 847 | match = re.fullmatch(r"([1-9][0-9]{0,9})\n?", payload) |
| 848 | if match is None: |
| 849 | raise SnapshotRestoreError("restored PID file is invalid") |
| 850 | restored_pid = int(match.group(1)) |
| 851 | if restored_pid > 2**31 - 1: |
| 852 | raise SnapshotRestoreError("restored PID file is invalid") |
| 853 | return restored_pid |
| 854 | |
| 855 | def _process_state(self, pid: int) -> tuple[int, int, int, int]: |
| 856 | payload = (Path("/proc") / str(pid) / "stat").read_text() |
| 857 | command_end = payload.rfind(")") |
| 858 | fields = payload[command_end + 2 :].split() |
| 859 | if command_end < 0 or len(fields) < 20: |
| 860 | raise ValueError(f"invalid process state for PID {pid}") |
| 861 | return (int(fields[1]), int(fields[2]), int(fields[3]), int(fields[19])) |
| 862 | |
| 863 | def _process_states(self) -> dict[int, tuple[int, int, int, int]]: |
| 864 | states: dict[int, tuple[int, int, int, int]] = {} |
| 865 | for entry in Path("/proc").iterdir(): |
| 866 | if not entry.name.isdigit(): |
| 867 | continue |
| 868 | pid = int(entry.name) |
| 869 | try: |
| 870 | states[pid] = self._process_state(pid) |
| 871 | except (OSError, ValueError): |
| 872 | continue |
| 873 | return states |
| 874 | |
| 875 | def _process_command(self, pid: int) -> tuple[str, ...]: |
| 876 | payload = (Path("/proc") / str(pid) / "cmdline").read_bytes() |
| 877 | return tuple( |
| 878 | os.fsdecode(argument) for argument in payload.split(b"\0") if argument |
| 879 | ) |
| 880 | |
| 881 | def _pidfd_open(self, pid: int) -> int: |
| 882 | pidfd_open = getattr(os, "pidfd_open", None) |
| 883 | if not callable(pidfd_open): |
| 884 | raise SnapshotRestoreError("snapshot requires Linux pidfd support") |
| 885 | return pidfd_open(pid) |
| 886 | |
| 887 | def _pidfd_exited(self, pidfd: int) -> bool: |
| 888 | poller = select.poll() |
| 889 | poller.register(pidfd, select.POLLIN | select.POLLHUP | select.POLLERR) |
| 890 | return any(self._pidfd_event_is_exit(event) for _fd, event in poller.poll(0)) |
| 891 | |
| 892 | def _pidfd_event_is_exit(self, event: int) -> bool: |
| 893 | if event & (select.POLLERR | select.POLLNVAL): |
| 894 | raise SnapshotRestoreError("restored process pidfd poll failed") |
| 895 | return bool(event & (select.POLLIN | select.POLLHUP)) |
| 896 | |
| 897 | def _pin_restored_tree( |
| 898 | self, |
| 899 | artifact: Path, |
| 900 | process_tree: tuple[int, ...], |
| 901 | cuda_holders: tuple[int, ...], |
| 902 | ) -> None: |
| 903 | root_pid = process_tree[0] |
| 904 | expected_pids = set(process_tree) |
| 905 | expected_cuda_holders = set(cuda_holders) |
| 906 | if not expected_cuda_holders or not expected_cuda_holders.issubset( |
| 907 | expected_pids |
| 908 | ): |
| 909 | raise SnapshotRestoreError( |
| 910 | "captured CUDA holders are outside the captured process tree" |
| 911 | ) |
| 912 | if root_pid in self._restored_processes: |
| 913 | raise SnapshotRestoreError("restored PID already has a pinned transaction") |
| 914 | handles: list[tuple[int, int]] = [] |
| 915 | try: |
| 916 | for pid in process_tree: |
| 917 | handles.append((pid, self._pidfd_open(pid))) |
| 918 | handles_by_pid = dict(handles) |
| 919 | if any(self._pidfd_exited(pidfd) for _pid, pidfd in handles): |
| 920 | raise SnapshotRestoreError( |
| 921 | "restored process exited during verification" |
| 922 | ) |
| 923 | |
| 924 | root_state = self._process_state(root_pid) |
| 925 | if root_state[1:3] != (root_pid, root_pid): |
| 926 | raise SnapshotRestoreError( |
| 927 | "restored root is not the session and process-group leader" |
| 928 | ) |
| 929 | command = self._process_command(root_pid) |
| 930 | release_file = str(artifact / "release.json") |
| 931 | if not any( |
| 932 | command[index : index + 2] == ("-m", "vllm.snapshot.server") |
| 933 | for index in range(len(command) - 1) |
| 934 | ) or not any( |
| 935 | command[index : index + 2] == ("--release-file", release_file) |
| 936 | for index in range(len(command) - 1) |
| 937 | ): |
| 938 | raise SnapshotRestoreError( |
| 939 | "restored root command does not match the snapshot barrier" |
| 940 | ) |
| 941 | if self._process_state(root_pid) != root_state: |
| 942 | raise SnapshotRestoreError("restored PID changed during verification") |
| 943 | if self._pidfd_exited(handles[0][1]): |
| 944 | raise SnapshotRestoreError("restored root exited during verification") |
| 945 | |
| 946 | first_states = self._process_states() |
| 947 | live_session = { |
| 948 | pid for pid, state in first_states.items() if state[2] == root_pid |
| 949 | } |
| 950 | if not expected_cuda_holders.issubset(live_session): |
| 951 | raise SnapshotRestoreError( |
| 952 | "restored session is missing captured CUDA holders" |
| 953 | ) |
| 954 | if live_session != expected_pids: |
| 955 | raise SnapshotRestoreError( |
| 956 | "restored session does not match the captured process tree" |
| 957 | ) |
| 958 | if first_states[root_pid] != root_state: |
| 959 | raise SnapshotRestoreError("restored PID changed during verification") |
| 960 | for pid in process_tree[1:]: |
| 961 | state = first_states[pid] |
| 962 | pidfd = handles_by_pid[pid] |
| 963 | if self._process_state(pid) != state or self._pidfd_exited(pidfd): |
| 964 | raise SnapshotRestoreError( |
| 965 | "restored process tree changed during verification" |
| 966 | ) |
| 967 | |
| 968 | second_states = self._process_states() |
| 969 | second_session = { |
| 970 | pid for pid, state in second_states.items() if state[2] == root_pid |
| 971 | } |
| 972 | if second_session != expected_pids or any( |
| 973 | second_states[pid] != first_states[pid] for pid in process_tree |
| 974 | ): |
| 975 | raise SnapshotRestoreError( |
| 976 | "restored process tree changed during verification" |
| 977 | ) |
| 978 | if any(self._pidfd_exited(pidfd) for _pid, pidfd in handles): |
| 979 | raise SnapshotRestoreError( |
| 980 | "restored process exited during verification" |
| 981 | ) |
| 982 | except BaseException as error: |
| 983 | _close_pidfds(handles) |
| 984 | if isinstance(error, SnapshotRestoreError): |
| 985 | raise |
| 986 | if not isinstance(error, Exception): |
| 987 | raise |
| 988 | raise SnapshotRestoreError( |
| 989 | "restored process tree could not be pinned" |
| 990 | ) from error |
| 991 | self._restored_processes[root_pid] = tuple(handles) |
| 992 | |
| 993 | def _abort_failed_restore( |
| 994 | self, |
| 995 | artifact: Path, |
| 996 | root_pid: int, |
| 997 | staged_remaps: dict[str, tuple[int, int]], |
| 998 | primary: BaseException, |
| 999 | ) -> NoReturn: |
| 1000 | secondary: list[str] = [] |
| 1001 | try: |
| 1002 | _write_json_atomic( |
| 1003 | artifact / "release.json", {"release": False}, overwrite=True |
| 1004 | ) |
| 1005 | except BaseException as error: |
| 1006 | secondary.append(f"abort marker failed: {_error_detail(error)}") |
| 1007 | # The marker above is only a request. Terminate and wait for the tree we |
| 1008 | # own, before releasing the shared state it may still be using. A tree |
| 1009 | # that could not be pinned is not ours to signal. |
| 1010 | if root_pid in self._restored_processes: |
| 1011 | try: |
| 1012 | self.cleanup(root_pid) |
| 1013 | except BaseException as error: |
| 1014 | secondary.append(f"restore cleanup failed: {_error_detail(error)}") |
| 1015 | try: |
| 1016 | self._rollback_link_remaps(staged_remaps) |
| 1017 | except BaseException as error: |
| 1018 | secondary.append(f"remap rollback failed: {_error_detail(error)}") |
| 1019 | if secondary: |
| 1020 | raise SnapshotRestoreError( |
| 1021 | f"snapshot restore failed: {_error_detail(primary)}; " |
| 1022 | + "; ".join(secondary) |
| 1023 | ) from primary |
| 1024 | raise primary |
| 1025 | |
| 1026 | def _verify_external_cache_files(self, manifest: SnapshotManifest) -> None: |
| 1027 | """Reject a rotated generated cache before CRIU reopens it by path. |
| 1028 | |
| 1029 | Like the PID probe above this reports a stale host early rather than |
| 1030 | guarding it: CRIU stays authoritative for a cache replaced afterwards. |
| 1031 | """ |
| 1032 | for path, fingerprint in manifest.external_cache_files: |
| 1033 | target = Path(path) |
| 1034 | if not target.is_file(): |
| 1035 | raise SnapshotRestoreError( |
| 1036 | f"external cache file vanished since capture: {path}; " |
| 1037 | f"{_CACHE_REMEDY}" |
| 1038 | ) |
| 1039 | if fingerprint in (_CACHE_LOCKED, _CACHE_UNREADABLE): |
| 1040 | continue |
| 1041 | try: |
| 1042 | current = self._sha256(target) |
| 1043 | except OSError as error: |
| 1044 | raise SnapshotRestoreError( |
| 1045 | f"external cache file is unreadable since capture: {path}; " |
| 1046 | f"{_CACHE_REMEDY}" |
| 1047 | ) from error |
| 1048 | if current != fingerprint: |
| 1049 | raise SnapshotRestoreError( |
| 1050 | f"external cache file changed since capture: {path}; " |
| 1051 | f"{_CACHE_REMEDY}" |
| 1052 | ) |
| 1053 | |
| 1054 | def restore(self, artifact: Path, manifest: SnapshotManifest) -> int: |
| 1055 | # This is necessarily a best-effort TOCTOU check. CRIU restore and the |
| 1056 | # rollback below remain authoritative if a PID is claimed after it. |
| 1057 | for pid in manifest.process_tree: |
| 1058 | try: |
| 1059 | os.kill(pid, 0) |
| 1060 | except ProcessLookupError: |
| 1061 | continue |
| 1062 | except PermissionError: |
| 1063 | pass |
| 1064 | except Exception as error: |
| 1065 | raise SnapshotRestoreError( |
| 1066 | f"captured PID availability probe failed: {pid}" |
| 1067 | ) from error |
| 1068 | raise SnapshotRestoreError(f"captured PID is already occupied: {pid}") |
| 1069 | self._verify_external_cache_files(manifest) |
| 1070 | expected_root_pid = manifest.process_tree[0] |
| 1071 | (artifact / "release.json").unlink(missing_ok=True) |
| 1072 | pidfile = artifact / "restored.pid" |
| 1073 | pidfile.unlink(missing_ok=True) |
| 1074 | self._reset_child_log(artifact) |
| 1075 | staged_remaps = self._restore_link_remaps(artifact) |
| 1076 | try: |
| 1077 | self._criu( |
| 1078 | "restore", |
| 1079 | artifact, |
| 1080 | [ |
| 1081 | "--images-dir", |
| 1082 | str(artifact / "images"), |
| 1083 | "--log-file", |
| 1084 | "restore.log", |
| 1085 | *_COMMON_CRIU_OPTIONS, |
| 1086 | "--restore-detached", |
| 1087 | "--pidfile", |
| 1088 | str(pidfile), |
| 1089 | ], |
| 1090 | ) |
| 1091 | # Pin before reading the PID file so that a failure between restore |
| 1092 | # and the read cannot orphan the tree. |
| 1093 | self._pin_restored_tree( |
| 1094 | artifact, manifest.process_tree, manifest.cuda_holders |
| 1095 | ) |
| 1096 | restored_pid = self._read_restored_pid(pidfile) |
| 1097 | if restored_pid != expected_root_pid: |
| 1098 | raise SnapshotRestoreError( |
| 1099 | "restored PID does not match the captured process tree" |
| 1100 | ) |
| 1101 | except BaseException as error: |
| 1102 | self._abort_failed_restore( |
| 1103 | artifact, expected_root_pid, staged_remaps, error |
| 1104 | ) |
| 1105 | return restored_pid |
| 1106 | |
| 1107 | def release(self, artifact: Path, host: str | None, port: int) -> None: |
| 1108 | if not 1 <= port <= 65535: |
| 1109 | raise SnapshotRestoreError(f"snapshot restore port is invalid: {port}") |
| 1110 | family = socket.AF_INET6 if host and ":" in host else socket.AF_INET |
| 1111 | with socket.socket(family=family, type=socket.SOCK_STREAM) as listener_probe: |
| 1112 | listener_probe.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1) |
| 1113 | try: |
| 1114 | listener_probe.bind((host or "", port)) |
| 1115 | except OSError as error: |
| 1116 | address = f"{host or '0.0.0.0'}:{port}" |
| 1117 | raise SnapshotRestoreError( |
| 1118 | f"snapshot restore address is already in use: {address}" |
| 1119 | ) from error |
| 1120 | |
| 1121 | _write_json_atomic( |
| 1122 | artifact / "release.json", |
| 1123 | {"release": True, "host": host, "port": port}, |
| 1124 | overwrite=True, |
| 1125 | ) |
| 1126 | |
| 1127 | def _connect_host(self, host: str | None) -> str: |
| 1128 | if not host or host == "0.0.0.0": |
| 1129 | return "127.0.0.1" |
| 1130 | if host == "::": |
| 1131 | return "::1" |
| 1132 | return host |
| 1133 | |
| 1134 | def wait_listener(self, root_pid: int, host: str | None, port: int) -> None: |
| 1135 | deadline = time.monotonic() + self.timeout_s |
| 1136 | connect_host = self._connect_host(host) |
| 1137 | while time.monotonic() < deadline: |
| 1138 | try: |
| 1139 | with socket.create_connection((connect_host, port), timeout=1): |
| 1140 | return |
| 1141 | except OSError as error: |
| 1142 | handles = self._restored_processes.get(root_pid) |
| 1143 | if handles is None or self._pidfd_exited(handles[0][1]): |
| 1144 | raise SnapshotRestoreError( |
| 1145 | "restored process exited before binding HTTP" |
| 1146 | ) from error |
| 1147 | time.sleep(0.1) |
| 1148 | raise SnapshotRestoreError("restored HTTP listener did not become ready") |
| 1149 | |
| 1150 | def request_oracle( |
| 1151 | self, host: str | None, port: int, manifest: SnapshotManifest |
| 1152 | ) -> Oracle: |
| 1153 | connect_host = self._connect_host(host) |
| 1154 | url_host = f"[{connect_host}]" if ":" in connect_host else connect_host |
| 1155 | payload = json.dumps( |
| 1156 | { |
| 1157 | "model": manifest.served_model_name, |
| 1158 | "prompt": "The capital of France is", |
| 1159 | "min_tokens": 1, |
| 1160 | "max_tokens": 1, |
| 1161 | "temperature": 0, |
| 1162 | "seed": 0, |
| 1163 | "logprobs": 0, |
| 1164 | "return_token_ids": True, |
| 1165 | } |
| 1166 | ).encode() |
| 1167 | request = urllib.request.Request( |
| 1168 | f"http://{url_host}:{port}/v1/completions", |
| 1169 | data=payload, |
| 1170 | headers={"Content-Type": "application/json"}, |
| 1171 | ) |
| 1172 | with urllib.request.urlopen(request, timeout=120) as response: |
| 1173 | body = json.load(response) |
| 1174 | choice = body["choices"][0] |
| 1175 | try: |
| 1176 | return Oracle( |
| 1177 | token_ids=tuple(choice["token_ids"]), |
| 1178 | text=choice["text"], |
| 1179 | sampled_token_logprob=choice["logprobs"]["token_logprobs"][0], |
| 1180 | ) |
| 1181 | except ValueError as error: |
| 1182 | raise SnapshotRestoreError( |
| 1183 | "snapshot HTTP canary must return exactly one finite token logprob" |
| 1184 | ) from error |
| 1185 | except (IndexError, KeyError, TypeError) as error: |
| 1186 | raise SnapshotRestoreError( |
| 1187 | "snapshot HTTP canary response is missing sampled token logprob" |
| 1188 | ) from error |
| 1189 | |
| 1190 | def cleanup(self, root_pid: int) -> None: |
| 1191 | handles = self._restored_processes.get(root_pid) |
| 1192 | if handles is None: |
| 1193 | raise SnapshotRestoreError("no pinned restore transaction for cleanup") |
| 1194 | pidfd_send_signal = getattr(signal, "pidfd_send_signal", None) |
| 1195 | if not callable(pidfd_send_signal): |
| 1196 | raise SnapshotRestoreError("restored process cleanup is incomplete") |
| 1197 | try: |
| 1198 | for _pid, pidfd in handles: |
| 1199 | with suppress(ProcessLookupError): |
| 1200 | pidfd_send_signal(pidfd, signal.SIGTERM, None, 0) |
| 1201 | time.sleep(min(0.5, self.timeout_s)) |
| 1202 | for _pid, pidfd in handles: |
| 1203 | with suppress(ProcessLookupError): |
| 1204 | pidfd_send_signal(pidfd, signal.SIGKILL, None, 0) |
| 1205 | except BaseException as error: |
| 1206 | raise SnapshotRestoreError( |
| 1207 | f"restored process cleanup is incomplete: {_error_detail(error)}" |
| 1208 | ) from error |
| 1209 | |
| 1210 | poller = select.poll() |
| 1211 | for _pid, pidfd in handles: |
| 1212 | poller.register(pidfd, select.POLLIN | select.POLLHUP | select.POLLERR) |
| 1213 | pending = {pidfd for _pid, pidfd in handles} |
| 1214 | deadline = time.monotonic() + self.timeout_s |
| 1215 | while pending: |
| 1216 | remaining = deadline - time.monotonic() |
| 1217 | if remaining <= 0: |
| 1218 | raise SnapshotRestoreError("restored process cleanup is incomplete") |
| 1219 | for pidfd, event in poller.poll(max(1, int(remaining * 1000))): |
| 1220 | if not self._pidfd_event_is_exit(event): |
| 1221 | continue |
| 1222 | pending.discard(pidfd) |
| 1223 | with suppress(KeyError, OSError): |
| 1224 | poller.unregister(pidfd) |
| 1225 | self._restored_processes.pop(root_pid) |
| 1226 | _close_pidfds(handles) |
| 1227 | |
| 1228 | def complete_restore(self, root_pid: int) -> None: |
| 1229 | handles = self._restored_processes.pop(root_pid, None) |
| 1230 | if handles is None: |
| 1231 | raise SnapshotRestoreError("no pinned restore transaction to complete") |
| 1232 | _close_pidfds(handles) |
| 1233 | |
| 1234 | def abort_create(self, root_pid: int) -> None: |
| 1235 | process = self._children.get(root_pid) |
| 1236 | if process is None: |
| 1237 | return |
| 1238 | with suppress(ProcessLookupError): |
| 1239 | os.killpg(root_pid, signal.SIGKILL) |
| 1240 | try: |
| 1241 | process.wait(timeout=10) |
| 1242 | except subprocess.TimeoutExpired as error: |
| 1243 | raise SnapshotCreateError( |
| 1244 | f"snapshot child process group {root_pid} survived SIGKILL" |
| 1245 | ) from error |
| 1246 | self._children.pop(root_pid, None) |