| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import contextlib |
| 5 | import os |
| 6 | import threading |
| 7 | import weakref |
| 8 | from collections.abc import Callable, Iterator, Sequence |
| 9 | from dataclasses import dataclass |
| 10 | from enum import Enum, auto |
| 11 | from multiprocessing import connection |
| 12 | from multiprocessing.process import BaseProcess |
| 13 | from multiprocessing.queues import Queue |
| 14 | from typing import TYPE_CHECKING, cast |
| 15 | |
| 16 | import msgspec |
| 17 | import zmq |
| 18 | |
| 19 | from vllm import envs |
| 20 | from vllm.config import CacheConfig, ParallelConfig, VllmConfig |
| 21 | from vllm.logger import init_logger |
| 22 | from vllm.platforms import current_platform |
| 23 | from vllm.ray.ray_env import get_env_vars_to_copy |
| 24 | from vllm.utils import numa_utils |
| 25 | from vllm.utils.network_utils import ( |
| 26 | get_open_port, |
| 27 | get_open_zmq_ipc_path, |
| 28 | get_tcp_uri, |
| 29 | zmq_socket_ctx, |
| 30 | ) |
| 31 | from vllm.utils.system_utils import get_mp_context, set_env_var |
| 32 | from vllm.v1.engine.coordinator import DPCoordinator |
| 33 | from vllm.v1.executor import Executor, UniProcExecutor |
| 34 | from vllm.v1.executor.ray_utils import WORKER_SPECIFIC_ENV_VARS |
| 35 | from vllm.v1.utils import _SubprocessWrapper, get_engine_client_zmq_addr, shutdown |
| 36 | |
| 37 | if TYPE_CHECKING: |
| 38 | from ray.util.placement_group import PlacementGroup |
| 39 | |
| 40 | logger = init_logger(__name__) |
| 41 | |
| 42 | STARTUP_POLL_PERIOD_MS = 10000 |
| 43 | ROCM_ENGINE_PROCESS_SHUTDOWN_TIMEOUT_S = 15.0 |
| 44 | |
| 45 | |
| 46 | def get_engine_process_shutdown_timeout( |
| 47 | request_timeout: float | None, |
| 48 | process_timeout: float | None, |
| 49 | ) -> float | None: |
| 50 | """Return the EngineCore process-manager shutdown timeout. |
| 51 | |
| 52 | ``VllmConfig.shutdown_timeout`` controls how long in-flight requests may |
| 53 | drain. A value of zero therefore tells EngineCore to abort requests as soon |
| 54 | as it receives SIGTERM. The parent process manager still needs a separate |
| 55 | window in which the EngineCore can release device resources before it is |
| 56 | force-killed. ROCm teardown can take longer than the generic best-effort |
| 57 | window, and force-killing during teardown can leave VRAM resident. |
| 58 | |
| 59 | ``process_timeout`` may be a remaining budget computed by an outer process |
| 60 | manager. Keep it unchanged unless both values are zero: a zero remaining |
| 61 | budget for a positive request timeout must not receive a fresh grace period |
| 62 | because EngineCore relies on that deadline to enforce request draining. |
| 63 | """ |
| 64 | if request_timeout == 0 and process_timeout == 0 and current_platform.is_rocm(): |
| 65 | return ROCM_ENGINE_PROCESS_SHUTDOWN_TIMEOUT_S |
| 66 | return process_timeout |
| 67 | |
| 68 | |
| 69 | class CoreEngineState(Enum): |
| 70 | NEW = auto() |
| 71 | CONNECTED = auto() |
| 72 | READY = auto() |
| 73 | |
| 74 | |
| 75 | class CoreEngine: |
| 76 | """One per data parallel rank, used to track state during handshaking.""" |
| 77 | |
| 78 | def __init__(self, index: int = 0, local: bool = True): |
| 79 | self.local = local |
| 80 | self.identity = index.to_bytes(2, "little") |
| 81 | |
| 82 | self.state = CoreEngineState.NEW |
| 83 | |
| 84 | |
| 85 | @dataclass |
| 86 | class EngineZmqAddresses: |
| 87 | # ZMQ input socket addresses for each front-end client (requests) |
| 88 | inputs: list[str] |
| 89 | # ZMQ output socket addresses for each front-end client (responses) |
| 90 | outputs: list[str] |
| 91 | # ZMQ input socket address of DP coordinator if applicable |
| 92 | coordinator_input: str | None = None |
| 93 | # ZMQ output socket address of DP coordinator if applicable |
| 94 | coordinator_output: str | None = None |
| 95 | # ZMQ socket for front-end to connect to DP coordinator. |
| 96 | # Not used by engine, just relayed to front-end in handshake response. |
| 97 | # Only required for external DP LB case. |
| 98 | frontend_stats_publish_address: str | None = None |
| 99 | |
| 100 | |
| 101 | @dataclass |
| 102 | class EngineHandshakeMetadata: |
| 103 | """Metadata sent to each engine process during startup handshake, |
| 104 | including addresses of the front-end ZMQ queues that they should |
| 105 | connect to. |
| 106 | """ |
| 107 | |
| 108 | addresses: EngineZmqAddresses |
| 109 | parallel_config: dict[str, int | str | list[int]] |
| 110 | |
| 111 | |
| 112 | def _make_control_bundle(node_ip: str) -> dict[str, float]: |
| 113 | # The engine actor is scheduled on the final CPU-only bundle. Keep that |
| 114 | # bundle colocated with the group's first GPU bundle so the actor does not |
| 115 | # float to an unrelated node and reorder worker ranks away from the |
| 116 | # advertised DP bootstrap host. |
| 117 | return {"CPU": 1.0, "node:" + node_ip: 0.001} |
| 118 | |
| 119 | |
| 120 | def _get_bundle_node_ip(bundle: dict[str, float]) -> str: |
| 121 | for key in bundle: |
| 122 | if key.startswith("node:"): |
| 123 | return key.split(":", 1)[1] |
| 124 | raise ValueError(f"Missing node affinity in placement bundle: {bundle}") |
| 125 | |
| 126 | |
| 127 | def _dp_nodes_master_first( |
| 128 | available_resources: dict[str, dict], dp_master_ip: str |
| 129 | ) -> list[tuple[str, dict]]: |
| 130 | """Live Ray nodes as ``(node_id, resources)`` with the DP master first. |
| 131 | |
| 132 | The master is found through the ``node:<ip>`` resource Ray sets on every |
| 133 | node, so this works with the resource maps from ``ray._private.state``. |
| 134 | Do not replace it with ``ray.util.state.list_nodes()``: that API goes |
| 135 | through the dashboard HTTP server and only exists with ``ray[default]`` |
| 136 | (see #23822). |
| 137 | """ |
| 138 | master_key = f"node:{dp_master_ip}" |
| 139 | nodes = sorted( |
| 140 | available_resources.items(), key=lambda item: master_key not in item[1] |
| 141 | ) |
| 142 | assert len(nodes) > 0, "No nodes with resources found in Ray cluster." |
| 143 | assert master_key in nodes[0][1], ( |
| 144 | f"The DP master node (ip: {dp_master_ip}) is missing or dead" |
| 145 | ) |
| 146 | assert len(nodes) == 1 or master_key not in nodes[1][1], ( |
| 147 | "There can only be one head node" |
| 148 | ) |
| 149 | return nodes |
| 150 | |
| 151 | |
| 152 | def _node_ip_from_resources(node_resources: dict) -> str | None: |
| 153 | """Return the node IP encoded in a Ray per-node resource dict, or None. |
| 154 | |
| 155 | Ray advertises each node's IP as a ``node:<ip>`` resource key. The head node |
| 156 | also carries ``node:__internal_head__``, and placement groups add |
| 157 | ``..._group_...`` keys; both are ignored. |
| 158 | """ |
| 159 | for key in node_resources: |
| 160 | if ( |
| 161 | key.startswith("node:") |
| 162 | and key != "node:__internal_head__" |
| 163 | and "_group_" not in key |
| 164 | ): |
| 165 | return key.split(":", 1)[1] |
| 166 | return None |
| 167 | |
| 168 | |
| 169 | @contextlib.contextmanager |
| 170 | def _configure_uniproc_startup_threads( |
| 171 | executor_class: type[Executor], local_engine_count: int |
| 172 | ) -> Iterator[None]: |
| 173 | if ( |
| 174 | not issubclass(executor_class, UniProcExecutor) |
| 175 | or current_platform.is_cpu() |
| 176 | or "OMP_NUM_THREADS" in os.environ |
| 177 | ): |
| 178 | yield |
| 179 | return |
| 180 | |
| 181 | import torch |
| 182 | |
| 183 | from vllm.utils.torch_utils import ( |
| 184 | OMP_NUM_THREADS_SET_BY_VLLM, |
| 185 | set_default_torch_num_threads, |
| 186 | startup_omp_num_threads, |
| 187 | ) |
| 188 | |
| 189 | original_threads = torch.get_num_threads() |
| 190 | num_threads = min(original_threads, startup_omp_num_threads(local_engine_count)) |
| 191 | torch_threads = ( |
| 192 | set_default_torch_num_threads(num_threads) |
| 193 | if num_threads < original_threads |
| 194 | else contextlib.nullcontext() |
| 195 | ) |
| 196 | # Fork inherits Torch's setting; spawn reads OMP_NUM_THREADS on import. |
| 197 | with ( |
| 198 | set_env_var("OMP_NUM_THREADS", str(num_threads)), |
| 199 | set_env_var(OMP_NUM_THREADS_SET_BY_VLLM, "1"), |
| 200 | torch_threads, |
| 201 | ): |
| 202 | yield |
| 203 | |
| 204 | |
| 205 | class CoreEngineProcManager: |
| 206 | """Utility class to handle creation, readiness, and shutdown |
| 207 | of background processes used by the AsyncLLM and LLMEngine. |
| 208 | """ |
| 209 | |
| 210 | def __init__( |
| 211 | self, |
| 212 | local_engine_count: int, |
| 213 | start_index: int, |
| 214 | local_start_index: int, |
| 215 | vllm_config: VllmConfig, |
| 216 | local_client: bool, |
| 217 | handshake_address: str, |
| 218 | executor_class: type[Executor], |
| 219 | log_stats: bool, |
| 220 | client_handshake_address: str | None = None, |
| 221 | tensor_queue: Queue | None = None, |
| 222 | ): |
| 223 | self._request_shutdown_timeout = vllm_config.shutdown_timeout |
| 224 | context = get_mp_context() |
| 225 | common_kwargs = { |
| 226 | "vllm_config": vllm_config, |
| 227 | "local_client": local_client, |
| 228 | "handshake_address": handshake_address, |
| 229 | "executor_class": executor_class, |
| 230 | "log_stats": log_stats, |
| 231 | "tensor_queue": tensor_queue, |
| 232 | } |
| 233 | |
| 234 | if client_handshake_address: |
| 235 | common_kwargs["client_handshake_address"] = client_handshake_address |
| 236 | |
| 237 | is_dp = vllm_config.parallel_config.data_parallel_size > 1 |
| 238 | |
| 239 | from vllm.v1.engine.core import EngineCoreProc |
| 240 | |
| 241 | self.processes: list[BaseProcess] = [] |
| 242 | local_dp_ranks = [] |
| 243 | for index in range(local_engine_count): |
| 244 | local_index = local_start_index + index |
| 245 | global_index = start_index + index |
| 246 | |
| 247 | # Start EngineCore in background process. |
| 248 | local_dp_ranks.append(local_index) |
| 249 | self.processes.append( |
| 250 | context.Process( |
| 251 | target=EngineCoreProc.run_engine_core, |
| 252 | name=f"EngineCore_DP{global_index}" if is_dp else "EngineCore", |
| 253 | kwargs=common_kwargs |
| 254 | | {"dp_rank": global_index, "local_dp_rank": local_index}, |
| 255 | ) |
| 256 | ) |
| 257 | |
| 258 | self._finalizer = weakref.finalize(self, shutdown, self.processes) |
| 259 | self.manager_stopped = threading.Event() |
| 260 | self.failed_proc_name: str | None = None |
| 261 | |
| 262 | # All ranks share this config object: capture the user-provided |
| 263 | # --device-ids list before the per-rank shard overwrites it. Mutating |
| 264 | # the config before each proc.start() works because the spawn method |
| 265 | # pickles process args at start() time, sequentially per rank. |
| 266 | user_assigned_gpu_ids = vllm_config.parallel_config.assigned_physical_gpu_ids |
| 267 | try: |
| 268 | with _configure_uniproc_startup_threads(executor_class, local_engine_count): |
| 269 | for proc, local_dp_rank in zip(self.processes, local_dp_ranks): |
| 270 | # Populate the logical-to-physical GPU mapping in DP for |
| 271 | # platforms that cannot rely on |
| 272 | # torch.accelerator.set_device_index(), and for Ray. |
| 273 | needs_device_env_isolation = not ( |
| 274 | current_platform.is_cuda_alike() or current_platform.is_xpu() |
| 275 | ) |
| 276 | if is_dp and ( |
| 277 | needs_device_env_isolation |
| 278 | or vllm_config.parallel_config.use_ray |
| 279 | ): |
| 280 | set_assigned_physical_gpu_ids_for_dp_rank( |
| 281 | vllm_config, local_dp_rank, user_assigned_gpu_ids |
| 282 | ) |
| 283 | |
| 284 | with numa_utils.configure_subprocess( |
| 285 | # EngineCore itself does not have a TP/PP-local rank. |
| 286 | # set_assigned_physical_gpu_ids_for_dp_rank() populates |
| 287 | # the logical-to-physical mapping for this DP |
| 288 | # shard, so local_rank=0 means "the first local GPU in |
| 289 | # this shard". The actual TP/PP worker processes spawned |
| 290 | # by the executor are bound separately with their own |
| 291 | # local_rank values. |
| 292 | vllm_config, |
| 293 | local_rank=0, |
| 294 | dp_local_rank=local_dp_rank, |
| 295 | process_kind="EngineCore", |
| 296 | ): |
| 297 | proc.start() |
| 298 | finally: |
| 299 | # Kill other procs if not all are running. |
| 300 | if self.finished_procs(): |
| 301 | self.shutdown() |
| 302 | |
| 303 | def shutdown(self, timeout: float | None = None) -> None: |
| 304 | """Shutdown engine core processes with configurable timeout.""" |
| 305 | self.manager_stopped.set() |
| 306 | if self._finalizer.detach() is not None: |
| 307 | process_timeout = get_engine_process_shutdown_timeout( |
| 308 | self._request_shutdown_timeout, timeout |
| 309 | ) |
| 310 | if process_timeout != timeout: |
| 311 | logger.info( |
| 312 | "[shutdown] EngineCore process manager: using %ss ROCm " |
| 313 | "cleanup grace after immediate request abort", |
| 314 | process_timeout, |
| 315 | ) |
| 316 | shutdown(self.processes, timeout=process_timeout) |
| 317 | |
| 318 | def monitor_engine_liveness(self) -> None: |
| 319 | """Monitor engine core process liveness.""" |
| 320 | sentinel_to_proc = {proc.sentinel: proc for proc in self.processes} |
| 321 | sentinels = set(sentinel_to_proc.keys()) |
| 322 | |
| 323 | while sentinels and not self.manager_stopped.is_set(): |
| 324 | died_sentinels = connection.wait(sentinels, timeout=1) |
| 325 | |
| 326 | for sentinel in died_sentinels: |
| 327 | proc = sentinel_to_proc.pop(cast(int, sentinel)) |
| 328 | exitcode = proc.exitcode |
| 329 | if exitcode != 0 and not self.manager_stopped.is_set(): |
| 330 | self.failed_proc_name = proc.name |
| 331 | if died_sentinels: |
| 332 | break |
| 333 | |
| 334 | self.shutdown() |
| 335 | |
| 336 | def sentinels(self) -> list: |
| 337 | return [proc.sentinel for proc in self.processes] |
| 338 | |
| 339 | def finished_procs(self) -> dict[str, int]: |
| 340 | """Returns dict of proc name -> exit code for any finished procs.""" |
| 341 | return { |
| 342 | proc.name: proc.exitcode |
| 343 | for proc in self.processes |
| 344 | if proc.exitcode is not None |
| 345 | } |
| 346 | |
| 347 | |
| 348 | class SignalCallback: |
| 349 | """Safely trigger a callback from signal handler context via a dedicated thread.""" |
| 350 | |
| 351 | def __init__(self, callback: Callable[[], None]): |
| 352 | self._callback = callback |
| 353 | self._event = threading.Event() |
| 354 | self._stopped = False |
| 355 | self._thread = threading.Thread( |
| 356 | target=self._run, |
| 357 | daemon=True, |
| 358 | name="signal-callback", |
| 359 | ) |
| 360 | self._thread.start() |
| 361 | |
| 362 | def _run(self): |
| 363 | self._event.wait() |
| 364 | if not self._stopped: |
| 365 | self._callback() |
| 366 | |
| 367 | def trigger(self): |
| 368 | self._event.set() |
| 369 | |
| 370 | def stop(self): |
| 371 | self._stopped = True |
| 372 | self._event.set() |
| 373 | |
| 374 | |
| 375 | def set_assigned_physical_gpu_ids_for_dp_rank( |
| 376 | vllm_config: VllmConfig, |
| 377 | local_dp_rank: int, |
| 378 | user_assigned_gpu_ids: list[int] | None = None, |
| 379 | ) -> None: |
| 380 | """Populate assigned_physical_gpu_ids on the config for the given DP rank. |
| 381 | |
| 382 | user_assigned_gpu_ids is the full (un-sharded) --device-ids list, if the |
| 383 | user provided one; this DP rank's shard is sliced from it. It is passed |
| 384 | explicitly rather than read from the config because callers may reuse |
| 385 | one config object across DP ranks, overwriting the field each time. |
| 386 | """ |
| 387 | world_size = vllm_config.parallel_config.world_size |
| 388 | local_world_size = vllm_config.parallel_config.local_world_size |
| 389 | evar = current_platform.device_control_env_var |
| 390 | |
| 391 | physical_gpu_ids = get_physical_gpu_ids_for_local_dp_rank( |
| 392 | evar, |
| 393 | local_dp_rank, |
| 394 | world_size, |
| 395 | local_world_size, |
| 396 | user_assigned_gpu_ids=user_assigned_gpu_ids, |
| 397 | ) |
| 398 | vllm_config.parallel_config.assigned_physical_gpu_ids = physical_gpu_ids |
| 399 | |
| 400 | |
| 401 | def get_physical_gpu_ids_for_local_dp_rank( |
| 402 | device_control_env_var: str, |
| 403 | local_dp_rank: int, |
| 404 | world_size: int, |
| 405 | local_world_size: int | None = None, |
| 406 | user_assigned_gpu_ids: list[int] | None = None, |
| 407 | ) -> list[int]: |
| 408 | """Returns list of physical GPU IDs for the specified |
| 409 | data parallel rank. |
| 410 | |
| 411 | For example, if world_size=2 and local_dp_rank=1, and there are 4 devices, |
| 412 | this will return [2, 3] for local_dp_rank=1. |
| 413 | |
| 414 | If user_assigned_gpu_ids is provided (e.g. from --device-ids), this DP |
| 415 | rank's shard is sliced from it instead of being derived from the |
| 416 | device-control env var. |
| 417 | """ |
| 418 | if local_world_size is None: |
| 419 | local_world_size = world_size |
| 420 | if user_assigned_gpu_ids is not None: |
| 421 | start = local_dp_rank * world_size |
| 422 | stop = start + local_world_size |
| 423 | if stop > len(user_assigned_gpu_ids): |
| 424 | raise ValueError( |
| 425 | f"--device-ids provides {len(user_assigned_gpu_ids)} devices, " |
| 426 | f"but DP rank {local_dp_rank} needs devices [{start}, {stop})" |
| 427 | ) |
| 428 | return user_assigned_gpu_ids[start:stop] |
| 429 | try: |
| 430 | return [ |
| 431 | current_platform.device_id_to_physical_device_id(i) |
| 432 | for i in range( |
| 433 | local_dp_rank * world_size, |
| 434 | local_dp_rank * world_size + local_world_size, |
| 435 | ) |
| 436 | ] |
| 437 | except IndexError as e: |
| 438 | raise Exception( |
| 439 | f"Error computing device indices for " |
| 440 | f"{device_control_env_var}: " |
| 441 | f"local range: [{local_dp_rank * world_size}, " |
| 442 | f"{(local_dp_rank + 1) * world_size}) " |
| 443 | "base value: " |
| 444 | f'"{os.getenv(device_control_env_var)}"' |
| 445 | ) from e |
| 446 | |
| 447 | |
| 448 | def _apply_dp_identity_suffix(dp_vllm_config, dp_rank: int) -> None: |
| 449 | # Ray actor names (RayExecutorV2) and KV-connector engine_ids must |
| 450 | # be unique across sibling DP engines or registration collides. |
| 451 | # Use the global DP rank, not a node-local rank, since sibling DP |
| 452 | # engines can span multiple nodes. |
| 453 | dp_vllm_config.instance_id = f"{dp_vllm_config.instance_id}_dp{dp_rank}" |
| 454 | if dp_vllm_config.kv_transfer_config is not None: |
| 455 | dp_vllm_config.kv_transfer_config.engine_id = ( |
| 456 | f"{dp_vllm_config.kv_transfer_config.engine_id}_dp{dp_rank}" |
| 457 | ) |
| 458 | |
| 459 | |
| 460 | class CoreEngineActorManager: |
| 461 | """Utility class to handle creation, readiness, and shutdown |
| 462 | of core engine Ray actors used by the AsyncLLM and LLMEngine. |
| 463 | |
| 464 | Different from CoreEngineProcManager, this class manages |
| 465 | core engines for both local and remote nodes. |
| 466 | """ |
| 467 | |
| 468 | def __init__( |
| 469 | self, |
| 470 | vllm_config: VllmConfig, |
| 471 | addresses: EngineZmqAddresses, |
| 472 | executor_class: type[Executor], |
| 473 | log_stats: bool, |
| 474 | placement_groups: list["PlacementGroup"] | None = None, |
| 475 | local_dp_ranks: list[int] | None = None, |
| 476 | ): |
| 477 | import copy |
| 478 | |
| 479 | import ray |
| 480 | from ray.runtime_env import RuntimeEnv |
| 481 | from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy |
| 482 | |
| 483 | from vllm.v1.engine.core import DPMoEEngineCoreActor, EngineCoreActor |
| 484 | |
| 485 | dp_size = vllm_config.parallel_config.data_parallel_size |
| 486 | actor_class = ( |
| 487 | DPMoEEngineCoreActor |
| 488 | if dp_size > 1 and vllm_config.model_config.is_moe |
| 489 | else EngineCoreActor |
| 490 | ) |
| 491 | |
| 492 | self.local_engine_actors: list[ray.ActorHandle] = [] |
| 493 | self.remote_engine_actors: list[ray.ActorHandle] = [] |
| 494 | |
| 495 | env_vars_list = get_env_vars_to_copy( |
| 496 | destination=actor_class.__name__, |
| 497 | exclude_vars=WORKER_SPECIFIC_ENV_VARS, |
| 498 | ) |
| 499 | self.env_vars_dict = { |
| 500 | name: os.environ[name] for name in env_vars_list if name in os.environ |
| 501 | } |
| 502 | runtime_env = RuntimeEnv(env_vars=self.env_vars_dict) |
| 503 | |
| 504 | self.addresses = addresses |
| 505 | self.executor_class = executor_class |
| 506 | self.log_stats = log_stats |
| 507 | local_engine_count = vllm_config.parallel_config.data_parallel_size_local |
| 508 | world_size = vllm_config.parallel_config.world_size |
| 509 | self.manager_stopped = threading.Event() |
| 510 | self.failed_proc_name: str | None = None |
| 511 | |
| 512 | if ray.is_initialized(): |
| 513 | logger.info("Ray is already initialized. Skipping Ray initialization.") |
| 514 | else: |
| 515 | ray.init() |
| 516 | |
| 517 | parallel_config = vllm_config.parallel_config |
| 518 | if parallel_config.enable_elastic_ep: |
| 519 | from vllm.distributed.utils import create_tcp_store |
| 520 | |
| 521 | ip = parallel_config.data_parallel_master_ip |
| 522 | store = create_tcp_store( |
| 523 | ip, |
| 524 | 0, |
| 525 | is_master=True, |
| 526 | world_size=-1, |
| 527 | wait_for_workers=False, |
| 528 | ) |
| 529 | parallel_config._coord_store_port = store.port |
| 530 | self._coord_store = store |
| 531 | |
| 532 | if placement_groups is not None: |
| 533 | assert local_dp_ranks is not None, ( |
| 534 | "local_dp_ranks must be provided if placement_groups is provided" |
| 535 | ) |
| 536 | assert len(placement_groups) == len(local_dp_ranks), ( |
| 537 | "placement_groups and local_dp_ranks must have the same length" |
| 538 | ) |
| 539 | logger.info("Using provided placement groups") |
| 540 | # TODO(rui): validate passed-in placement groups |
| 541 | self.created_placement_groups = [] |
| 542 | else: |
| 543 | placement_groups, local_dp_ranks = ( |
| 544 | CoreEngineActorManager.create_dp_placement_groups(vllm_config) |
| 545 | ) |
| 546 | self.created_placement_groups = placement_groups |
| 547 | assert len(placement_groups) == dp_size, ( |
| 548 | "Number of placement groups must match data parallel size" |
| 549 | ) |
| 550 | |
| 551 | self.placement_group_is_local = [] |
| 552 | refs = [] |
| 553 | for index, local_index, pg in zip( |
| 554 | range(dp_size), local_dp_ranks, placement_groups |
| 555 | ): |
| 556 | dp_vllm_config = copy.deepcopy(vllm_config) |
| 557 | if dp_size > 1: |
| 558 | _apply_dp_identity_suffix(dp_vllm_config, index) |
| 559 | dp_vllm_config.parallel_config.placement_group = pg |
| 560 | local_client = index < local_engine_count |
| 561 | |
| 562 | # Ray XPU known issue: dpctl initializes the GPU runtime early, so |
| 563 | # setting device env vars in Ray actor's initialization method |
| 564 | # will not affect device selection. See: |
| 565 | # https://github.com/ray-project/ray/blob/master/python/ray/_private/accelerators/intel_gpu.py#L56 # noqa: E501 |
| 566 | if current_platform.is_xpu(): |
| 567 | device_evar = current_platform.device_control_env_var |
| 568 | physical_gpu_ids = get_physical_gpu_ids_for_local_dp_rank( |
| 569 | device_evar, local_index, world_size |
| 570 | ) |
| 571 | actor_env_vars = self.env_vars_dict.copy() |
| 572 | actor_env_vars[device_evar] = ",".join(str(d) for d in physical_gpu_ids) |
| 573 | runtime_env = RuntimeEnv(env_vars=actor_env_vars) |
| 574 | |
| 575 | actor = ( |
| 576 | ray.remote(actor_class) |
| 577 | .options( |
| 578 | scheduling_strategy=PlacementGroupSchedulingStrategy( |
| 579 | placement_group=pg, |
| 580 | placement_group_bundle_index=world_size, |
| 581 | ), |
| 582 | runtime_env=runtime_env, |
| 583 | ) |
| 584 | .remote( |
| 585 | vllm_config=dp_vllm_config, |
| 586 | executor_class=executor_class, |
| 587 | log_stats=log_stats, |
| 588 | local_client=local_client, |
| 589 | addresses=addresses, |
| 590 | dp_rank=index, |
| 591 | local_dp_rank=local_index, |
| 592 | ) |
| 593 | ) |
| 594 | if local_client: |
| 595 | self.local_engine_actors.append(actor) |
| 596 | else: |
| 597 | self.remote_engine_actors.append(actor) |
| 598 | self.placement_group_is_local.append(local_client) |
| 599 | refs.append(actor.wait_for_init.remote()) |
| 600 | |
| 601 | ray.get(refs) |
| 602 | self.run_refs = [] |
| 603 | self.actor_run_ref_dict = dict() |
| 604 | for actor in self.local_engine_actors + self.remote_engine_actors: |
| 605 | ref = actor.run.remote() |
| 606 | self.run_refs.append(ref) |
| 607 | self.actor_run_ref_dict[actor] = ref |
| 608 | |
| 609 | @staticmethod |
| 610 | def create_dp_placement_groups( |
| 611 | vllm_config: VllmConfig, |
| 612 | ) -> tuple[list["PlacementGroup"], list[int]]: |
| 613 | """Create placement groups for data parallel.""" |
| 614 | import ray |
| 615 | from ray._private.state import available_resources_per_node |
| 616 | |
| 617 | logger.info("Creating placement groups for data parallel") |
| 618 | dp_master_ip = vllm_config.parallel_config.data_parallel_master_ip |
| 619 | dp_size = vllm_config.parallel_config.data_parallel_size |
| 620 | dp_size_local = vllm_config.parallel_config.data_parallel_size_local |
| 621 | |
| 622 | available_resources = available_resources_per_node() |
| 623 | world_size = vllm_config.parallel_config.world_size |
| 624 | placement_groups: list[PlacementGroup] = [] |
| 625 | local_dp_ranks: list[int] = [] |
| 626 | |
| 627 | nodes = [ |
| 628 | node_resources |
| 629 | for _, node_resources in _dp_nodes_master_first( |
| 630 | available_resources, dp_master_ip |
| 631 | ) |
| 632 | ] |
| 633 | |
| 634 | # optionally restrict DP placement to a caller-provided node set. |
| 635 | requested_node_ips = { |
| 636 | ip.strip() |
| 637 | for ip in envs.VLLM_RAY_DP_PLACEMENT_NODE_IPS.split(",") |
| 638 | if ip.strip() |
| 639 | } |
| 640 | if requested_node_ips: |
| 641 | allowed_node_ips = set(requested_node_ips) |
| 642 | # The master node must host the local ranks, so it has to be allowed. |
| 643 | if dp_master_ip not in allowed_node_ips: |
| 644 | allowed_node_ips.add(dp_master_ip) |
| 645 | filtered_nodes = [ |
| 646 | node_resources |
| 647 | for node_resources in nodes |
| 648 | if _node_ip_from_resources(node_resources) in allowed_node_ips |
| 649 | ] |
| 650 | logger.info( |
| 651 | "VLLM_RAY_DP_PLACEMENT_NODE_IPS set; restricting DP placement " |
| 652 | "from %d to %d node(s): %s", |
| 653 | len(nodes), |
| 654 | len(filtered_nodes), |
| 655 | sorted(allowed_node_ips), |
| 656 | ) |
| 657 | nodes = filtered_nodes |
| 658 | |
| 659 | device_str = current_platform.ray_device_key |
| 660 | n_node_devices: list[int] = [ |
| 661 | int(node_resources[device_str]) |
| 662 | for node_resources in nodes |
| 663 | if device_str in node_resources |
| 664 | ] |
| 665 | assert n_node_devices, f"No {device_str} found in Ray cluster." |
| 666 | max_device_per_node = max(n_node_devices) |
| 667 | |
| 668 | pack_strategy = envs.VLLM_RAY_DP_PACK_STRATEGY |
| 669 | _supported_pack_strategies = ("strict", "fill", "span") |
| 670 | if pack_strategy not in _supported_pack_strategies: |
| 671 | raise ValueError( |
| 672 | f"{envs.VLLM_RAY_DP_PACK_STRATEGY} is not supported. " |
| 673 | "Make sure to set `VLLM_RAY_DP_PACK_STRATEGY` " |
| 674 | f"to one of {_supported_pack_strategies}" |
| 675 | ) |
| 676 | |
| 677 | all2all_backend = vllm_config.parallel_config.all2all_backend |
| 678 | if pack_strategy == "fill" and ( |
| 679 | all2all_backend == "deepep_high_throughput" |
| 680 | or all2all_backend == "deepep_low_latency" |
| 681 | ): |
| 682 | raise ValueError( |
| 683 | "DeepEP kernels require EP ranks [0,7] (same for [8,15], ...) " |
| 684 | "to be on the same node, but VLLM_RAY_DP_PACK_STRATEGY=fill " |
| 685 | "does not guarantee that. " |
| 686 | "Please use VLLM_RAY_DP_PACK_STRATEGY=strict instead." |
| 687 | ) |
| 688 | |
| 689 | if pack_strategy in ("strict", "fill"): |
| 690 | placement_strategy = "STRICT_PACK" |
| 691 | else: |
| 692 | placement_strategy = "PACK" |
| 693 | assert world_size > max_device_per_node, ( |
| 694 | f"World size {world_size} is smaller than the " |
| 695 | "maximum number of devices per node " |
| 696 | f"{max_device_per_node}. Make sure to set " |
| 697 | "`VLLM_RAY_DP_PACK_STRATEGY` to `strict` or `fill`" |
| 698 | ) |
| 699 | |
| 700 | # if we need multiple nodes per dp group, we require for now that |
| 701 | # available nodes are homogeneous |
| 702 | assert set(n_node_devices) == {max_device_per_node}, ( |
| 703 | f"Nodes are not homogeneous, {nodes}" |
| 704 | ) |
| 705 | assert world_size % max_device_per_node == 0, ( |
| 706 | f"For multi-node data parallel groups, world_size ({world_size}) must " |
| 707 | f"be a multiple of number of devices per node ({max_device_per_node})." |
| 708 | ) |
| 709 | assert len(n_node_devices) * max_device_per_node >= world_size * dp_size, ( |
| 710 | f"Not enough total available nodes ({len(n_node_devices)}) " |
| 711 | f"and devices per node ({max_device_per_node}) " |
| 712 | f"to satisfy required world size {world_size} and data parallel size " |
| 713 | f"{dp_size}" |
| 714 | ) |
| 715 | assert dp_size_local == 1, ( |
| 716 | f"data-parallel-size-local {dp_size_local} should be set as the " |
| 717 | "default (1) for VLLM_RAY_DP_PACK_STRATEGY=span. " |
| 718 | "The actual data-parallel-size-local will be auto determined." |
| 719 | ) |
| 720 | |
| 721 | # bundles collected for a single DP rank from multiple nodes, |
| 722 | # for "span" pack strategy |
| 723 | collected_bundles = [] |
| 724 | for node_resources in nodes: |
| 725 | node_ip = _node_ip_from_resources(node_resources) |
| 726 | assert node_ip is not None, ( |
| 727 | f"No node IP key found in node resources: {node_resources}" |
| 728 | ) |
| 729 | |
| 730 | n_device_on_node = int(node_resources.get(device_str, 0)) |
| 731 | if pack_strategy == "span" and n_device_on_node != 0: |
| 732 | # Strictly speaking, |
| 733 | # dp_size_available = n_device_on_node / world_size |
| 734 | # and is a fraction, but we use 1 for easier processing |
| 735 | dp_size_available = 1 |
| 736 | else: |
| 737 | dp_size_available = n_device_on_node // world_size |
| 738 | |
| 739 | if node_ip == dp_master_ip: |
| 740 | if dp_size_available < dp_size_local: |
| 741 | raise ValueError( |
| 742 | f"Not enough resources to allocate {dp_size_local} DP ranks " |
| 743 | f"on DP master node {dp_master_ip}, possible to fit " |
| 744 | f"{dp_size_available} DP ranks." |
| 745 | ) |
| 746 | dp_size_to_allocate = dp_size_local |
| 747 | elif pack_strategy == "strict": |
| 748 | if dp_size_available < dp_size_local: |
| 749 | logger.info( |
| 750 | "Skipping node %s as %s DP ranks could not fit, " |
| 751 | "possible to fit %s DP ranks", |
| 752 | node_ip, |
| 753 | dp_size_local, |
| 754 | dp_size_available, |
| 755 | ) |
| 756 | continue |
| 757 | dp_size_to_allocate = dp_size_local |
| 758 | else: |
| 759 | # for "pack_strategy" in "fill" and "span" |
| 760 | # we always take everything that's available |
| 761 | dp_size_to_allocate = dp_size_available |
| 762 | |
| 763 | for i in range(dp_size_to_allocate): |
| 764 | device_bundle = [{device_str: 1.0, "node:" + node_ip: 0.001}] |
| 765 | if pack_strategy == "span": |
| 766 | collected_bundles += device_bundle * n_device_on_node |
| 767 | assert len(collected_bundles) <= world_size, ( |
| 768 | "collected_bundles should be <= world_size, " |
| 769 | f"but got {len(collected_bundles)=} and {world_size=}" |
| 770 | ) |
| 771 | |
| 772 | # we only create a placement group if we collected enough devices |
| 773 | if len(collected_bundles) < world_size: |
| 774 | continue |
| 775 | |
| 776 | control_node_ip = _get_bundle_node_ip(collected_bundles[0]) |
| 777 | bundles = collected_bundles + [ |
| 778 | _make_control_bundle(control_node_ip) |
| 779 | ] |
| 780 | collected_bundles = [] |
| 781 | else: |
| 782 | # STRICT_PACK already keeps every bundle in the placement |
| 783 | # group on one node, so the explicit node affinity on the |
| 784 | # control bundle is redundant for correctness here. Keep it |
| 785 | # anyway for consistency with the span path and to preserve |
| 786 | # intent if this scheduling strategy changes later. |
| 787 | bundles = device_bundle * world_size + [ |
| 788 | _make_control_bundle(node_ip) |
| 789 | ] |
| 790 | |
| 791 | pg = ray.util.placement_group( |
| 792 | name=f"dp_rank_{len(placement_groups)}", |
| 793 | strategy=placement_strategy, |
| 794 | bundles=bundles, |
| 795 | ) |
| 796 | placement_groups.append(pg) |
| 797 | local_dp_ranks.append(i) |
| 798 | if len(placement_groups) == dp_size: |
| 799 | break |
| 800 | |
| 801 | if len(placement_groups) == dp_size: |
| 802 | break |
| 803 | |
| 804 | if len(placement_groups) < dp_size: |
| 805 | raise ValueError( |
| 806 | f"Not enough resources to allocate {dp_size} " |
| 807 | "placement groups, only created " |
| 808 | f"{len(placement_groups)} placement groups. " |
| 809 | "Available resources: " |
| 810 | f"{available_resources}" |
| 811 | ) |
| 812 | assert len(placement_groups) == dp_size, ( |
| 813 | f"Created {len(placement_groups)} DP placement groups, expected {dp_size}" |
| 814 | ) |
| 815 | assert len(local_dp_ranks) == dp_size, ( |
| 816 | f"local_dp_ranks length {len(local_dp_ranks)} does not match " |
| 817 | f"expected {dp_size}" |
| 818 | ) |
| 819 | return placement_groups, local_dp_ranks |
| 820 | |
| 821 | @staticmethod |
| 822 | def add_dp_placement_groups( |
| 823 | old_vllm_config: VllmConfig, new_data_parallel_size: int |
| 824 | ) -> tuple[list["PlacementGroup"], list[int]]: |
| 825 | """Add placement groups for new data parallel size.""" |
| 826 | import ray |
| 827 | from ray._private.state import ( |
| 828 | available_resources_per_node, |
| 829 | total_resources_per_node, |
| 830 | ) |
| 831 | |
| 832 | old_dp_size = old_vllm_config.parallel_config.data_parallel_size |
| 833 | num_pg_to_create = new_data_parallel_size - old_dp_size |
| 834 | |
| 835 | if num_pg_to_create <= 0: |
| 836 | return [], [] |
| 837 | |
| 838 | dp_master_ip = old_vllm_config.parallel_config.data_parallel_master_ip |
| 839 | world_size = old_vllm_config.parallel_config.world_size |
| 840 | |
| 841 | # Both maps are keyed by node id and only contain live nodes. |
| 842 | available_resources = available_resources_per_node() |
| 843 | total_resources = total_resources_per_node() |
| 844 | nodes = _dp_nodes_master_first(available_resources, dp_master_ip) |
| 845 | |
| 846 | placement_groups = [] |
| 847 | local_dp_ranks = [] |
| 848 | num_pg_created = 0 |
| 849 | |
| 850 | device_str = current_platform.ray_device_key |
| 851 | for node_id, node_resources in nodes: |
| 852 | if num_pg_created >= num_pg_to_create: |
| 853 | break |
| 854 | |
| 855 | node_ip = _node_ip_from_resources(node_resources) |
| 856 | assert node_ip is not None, ( |
| 857 | f"No node IP key found in node resources: {node_resources}" |
| 858 | ) |
| 859 | if device_str not in node_resources: |
| 860 | continue |
| 861 | available_gpus = int(node_resources[device_str]) |
| 862 | |
| 863 | # Get total GPUs on this node from the node's resources |
| 864 | # Ray stores node resources with node ID as key |
| 865 | total_gpus = int(total_resources[node_id][device_str]) |
| 866 | |
| 867 | # Calculate used GPUs and used engines on this node |
| 868 | used_gpus = max(0, total_gpus - available_gpus) |
| 869 | used_engines_on_node = used_gpus // world_size |
| 870 | |
| 871 | # Calculate how many new engines this node can accommodate |
| 872 | available_engine_count = available_gpus // world_size |
| 873 | |
| 874 | # Create placement groups for new engines on this node |
| 875 | for i in range(available_engine_count): |
| 876 | if num_pg_created >= num_pg_to_create: |
| 877 | break |
| 878 | |
| 879 | rank = old_dp_size + num_pg_created |
| 880 | |
| 881 | # Create bundles with node constraint for master node |
| 882 | if node_ip == dp_master_ip: |
| 883 | bundles = [ |
| 884 | {device_str: 1.0, "node:" + dp_master_ip: 0.001} |
| 885 | ] * world_size + [{"CPU": 1.0}] |
| 886 | else: |
| 887 | bundles = [{device_str: 1.0}] * world_size + [{"CPU": 1.0}] |
| 888 | |
| 889 | pg = ray.util.placement_group( |
| 890 | name=f"dp_rank_{rank}", |
| 891 | strategy="STRICT_PACK", |
| 892 | bundles=bundles, |
| 893 | ) |
| 894 | placement_groups.append(pg) |
| 895 | |
| 896 | # Local rank starts from the number of engines already used |
| 897 | # on this node |
| 898 | local_rank = used_engines_on_node + i |
| 899 | local_dp_ranks.append(local_rank) |
| 900 | num_pg_created += 1 |
| 901 | |
| 902 | return placement_groups, local_dp_ranks |
| 903 | |
| 904 | def scale_up_elastic_ep( |
| 905 | self, |
| 906 | cur_vllm_config: VllmConfig, |
| 907 | new_data_parallel_size: int, |
| 908 | num_redundant_experts: int, |
| 909 | ) -> None: |
| 910 | import copy |
| 911 | |
| 912 | import ray |
| 913 | from ray.runtime_env import RuntimeEnv |
| 914 | from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy |
| 915 | |
| 916 | from vllm.v1.engine.core import DPMoEEngineCoreActor, EngineCoreActor |
| 917 | |
| 918 | actor_class = ( |
| 919 | DPMoEEngineCoreActor |
| 920 | if cur_vllm_config.model_config.is_moe |
| 921 | else EngineCoreActor |
| 922 | ) |
| 923 | |
| 924 | cur_data_parallel_size = len(self.local_engine_actors) + len( |
| 925 | self.remote_engine_actors |
| 926 | ) |
| 927 | |
| 928 | assert new_data_parallel_size > cur_data_parallel_size, ( |
| 929 | f"New data parallel size {new_data_parallel_size} must be greater " |
| 930 | f"than current data parallel size {cur_data_parallel_size} " |
| 931 | "for scale up" |
| 932 | ) |
| 933 | |
| 934 | placement_groups, local_dp_ranks = self.add_dp_placement_groups( |
| 935 | cur_vllm_config, new_data_parallel_size |
| 936 | ) |
| 937 | |
| 938 | world_size = cur_vllm_config.parallel_config.world_size |
| 939 | dp_master_ip = cur_vllm_config.parallel_config.data_parallel_master_ip |
| 940 | new_local_engines = 0 |
| 941 | |
| 942 | runtime_env = RuntimeEnv( |
| 943 | env_vars=self.env_vars_dict | {"VLLM_ELASTIC_EP_SCALE_UP_LAUNCH": "1"} |
| 944 | ) |
| 945 | for i, (pg, local_rank) in enumerate(zip(placement_groups, local_dp_ranks)): |
| 946 | rank = cur_data_parallel_size + i |
| 947 | dp_vllm_config = copy.deepcopy(cur_vllm_config) |
| 948 | if new_data_parallel_size > 1: |
| 949 | _apply_dp_identity_suffix(dp_vllm_config, rank) |
| 950 | dp_vllm_config.parallel_config.data_parallel_size = new_data_parallel_size |
| 951 | dp_vllm_config.parallel_config.eplb_config.num_redundant_experts = ( |
| 952 | num_redundant_experts |
| 953 | ) |
| 954 | dp_vllm_config.parallel_config.placement_group = pg |
| 955 | |
| 956 | # Check if this placement group is on the head node |
| 957 | local_client = any( |
| 958 | bundle.get("node:" + dp_master_ip, 0) > 0 for bundle in pg.bundle_specs |
| 959 | ) |
| 960 | |
| 961 | if local_client: |
| 962 | new_local_engines += 1 |
| 963 | # Update data_parallel_size_local |
| 964 | dp_vllm_config.parallel_config.data_parallel_size_local = ( |
| 965 | cur_vllm_config.parallel_config.data_parallel_size_local |
| 966 | + new_local_engines |
| 967 | ) |
| 968 | |
| 969 | actor = ( |
| 970 | ray.remote(actor_class) |
| 971 | .options( |
| 972 | scheduling_strategy=PlacementGroupSchedulingStrategy( |
| 973 | placement_group=pg, |
| 974 | placement_group_bundle_index=world_size, |
| 975 | ), |
| 976 | runtime_env=runtime_env, |
| 977 | ) |
| 978 | .remote( |
| 979 | vllm_config=dp_vllm_config, |
| 980 | executor_class=self.executor_class, |
| 981 | log_stats=self.log_stats, |
| 982 | local_client=local_client, |
| 983 | addresses=self.addresses, |
| 984 | dp_rank=rank, |
| 985 | local_dp_rank=local_rank, |
| 986 | ) |
| 987 | ) |
| 988 | |
| 989 | if local_client: |
| 990 | self.local_engine_actors.append(actor) |
| 991 | else: |
| 992 | self.remote_engine_actors.append(actor) |
| 993 | self.created_placement_groups.append(pg) |
| 994 | self.placement_group_is_local.append(local_client) |
| 995 | |
| 996 | actors = ( |
| 997 | self.local_engine_actors[-new_local_engines:] |
| 998 | if new_local_engines > 0 |
| 999 | else [] |
| 1000 | ) + self.remote_engine_actors[-(len(placement_groups) - new_local_engines) :] |
| 1001 | |
| 1002 | ray.get([actor.wait_for_init.remote() for actor in actors]) |
| 1003 | for actor in actors: |
| 1004 | ref = actor.run.remote() |
| 1005 | self.run_refs.append(ref) |
| 1006 | self.actor_run_ref_dict[actor] = ref |
| 1007 | |
| 1008 | def scale_down_elastic_ep( |
| 1009 | self, cur_data_parallel_size: int, new_data_parallel_size: int |
| 1010 | ) -> None: |
| 1011 | import ray |
| 1012 | |
| 1013 | assert cur_data_parallel_size > new_data_parallel_size, ( |
| 1014 | f"cur_data_parallel_size {cur_data_parallel_size} must be greater " |
| 1015 | f"than new_data_parallel_size {new_data_parallel_size} " |
| 1016 | "for scale down" |
| 1017 | ) |
| 1018 | for _ in range(cur_data_parallel_size - new_data_parallel_size): |
| 1019 | pg = self.created_placement_groups.pop() |
| 1020 | is_local = self.placement_group_is_local.pop() |
| 1021 | if is_local: |
| 1022 | self.local_engine_actors.pop() |
| 1023 | else: |
| 1024 | self.remote_engine_actors.pop() |
| 1025 | ray.util.remove_placement_group(pg) |
| 1026 | |
| 1027 | def remove_run_refs_for_scale_down(self, removed_dp_size: int) -> None: |
| 1028 | if removed_dp_size <= 0: |
| 1029 | return |
| 1030 | flags = self.placement_group_is_local[-removed_dp_size:] |
| 1031 | li = len(self.local_engine_actors) - 1 |
| 1032 | ri = len(self.remote_engine_actors) - 1 |
| 1033 | for is_local in reversed(flags): |
| 1034 | if is_local: |
| 1035 | actor = self.local_engine_actors[li] |
| 1036 | li -= 1 |
| 1037 | else: |
| 1038 | actor = self.remote_engine_actors[ri] |
| 1039 | ri -= 1 |
| 1040 | ref = self.actor_run_ref_dict.pop(actor) |
| 1041 | self.run_refs.remove(ref) |
| 1042 | |
| 1043 | def get_run_refs(self): |
| 1044 | return self.run_refs |
| 1045 | |
| 1046 | def monitor_engine_liveness(self) -> None: |
| 1047 | import ray |
| 1048 | |
| 1049 | while not self.manager_stopped.is_set(): |
| 1050 | actor_run_refs = list(self.get_run_refs()) |
| 1051 | if not actor_run_refs: |
| 1052 | logger.info( |
| 1053 | "There are no actors to monitor currently. " |
| 1054 | "The monitoring function is about to terminate." |
| 1055 | ) |
| 1056 | break |
| 1057 | actor_done_refs, _ = ray.wait(actor_run_refs, timeout=5) |
| 1058 | unexpected_failure = False |
| 1059 | for actor_ref in actor_done_refs: |
| 1060 | if self.manager_stopped.is_set(): |
| 1061 | break |
| 1062 | if actor_ref not in self.get_run_refs(): |
| 1063 | # The run refs may have been updated by elastic scale-down. |
| 1064 | continue |
| 1065 | try: |
| 1066 | ray.get(actor_ref) |
| 1067 | except ray.exceptions.RayActorError: |
| 1068 | self.failed_proc_name = f"Actor {actor_ref}" |
| 1069 | unexpected_failure = True |
| 1070 | |
| 1071 | if unexpected_failure: |
| 1072 | break |
| 1073 | |
| 1074 | self.shutdown() |
| 1075 | |
| 1076 | def shutdown(self, timeout: float | None = None) -> None: |
| 1077 | import ray |
| 1078 | |
| 1079 | self.manager_stopped.set() |
| 1080 | for actor in self.local_engine_actors + self.remote_engine_actors: |
| 1081 | ray.kill(actor) |
| 1082 | for pg in self.created_placement_groups: |
| 1083 | ray.util.remove_placement_group(pg) |
| 1084 | |
| 1085 | |
| 1086 | def get_engine_zmq_addresses( |
| 1087 | vllm_config: VllmConfig, |
| 1088 | num_api_servers: int = 1, |
| 1089 | *, |
| 1090 | defer_api_server_ports: bool = True, |
| 1091 | ) -> EngineZmqAddresses: |
| 1092 | """Allocate ZMQ addresses for engine-client communication. |
| 1093 | |
| 1094 | By default each TCP address is a ``tcp://host:0`` placeholder; the |
| 1095 | consumer (API-server child or single-process ``MPClient``) binds, then |
| 1096 | recovers the kernel-assigned port via ``getsockopt(zmq.LAST_ENDPOINT)`` |
| 1097 | and writes it back into ``addresses`` before the engine handshake. |
| 1098 | |
| 1099 | Set ``defer_api_server_ports=False`` only when the consumer cannot |
| 1100 | report a bound port back (e.g. the Rust front-end). IPC paths are |
| 1101 | unaffected.""" |
| 1102 | parallel_config = vllm_config.parallel_config |
| 1103 | local_engine_count = parallel_config.data_parallel_size_local |
| 1104 | local_start_index = parallel_config.data_parallel_rank_local |
| 1105 | dp_size = parallel_config.data_parallel_size |
| 1106 | host = parallel_config.data_parallel_master_ip |
| 1107 | local_engines_only = parallel_config.local_engines_only |
| 1108 | |
| 1109 | # In offline mode there is an LLM instance per DP rank and |
| 1110 | # one core engine per LLM, see |
| 1111 | # examples/features/data_parallel/data_parallel_offline.py. |
| 1112 | offline_mode = local_start_index is not None |
| 1113 | |
| 1114 | # client_local_only = True for cases where this front-end |
| 1115 | # sends requests only to colocated engines. |
| 1116 | client_local_only = ( |
| 1117 | offline_mode or local_engines_only or (local_engine_count == dp_size) |
| 1118 | ) |
| 1119 | # NOTE(yongji): handling scaling from intra-node to inter-node |
| 1120 | if parallel_config.enable_elastic_ep: |
| 1121 | client_local_only = False |
| 1122 | |
| 1123 | def _addr() -> str: |
| 1124 | if client_local_only: |
| 1125 | return get_open_zmq_ipc_path() |
| 1126 | return get_tcp_uri(host, 0 if defer_api_server_ports else get_open_port()) |
| 1127 | |
| 1128 | return EngineZmqAddresses( |
| 1129 | inputs=[_addr() for _ in range(num_api_servers)], |
| 1130 | outputs=[_addr() for _ in range(num_api_servers)], |
| 1131 | ) |
| 1132 | |
| 1133 | |
| 1134 | FrontendProcess = BaseProcess | _SubprocessWrapper |
| 1135 | |
| 1136 | |
| 1137 | @dataclass |
| 1138 | class CoreEngineLaunch: |
| 1139 | """Resources and startup barrier for launched engine processes.""" |
| 1140 | |
| 1141 | engine_manager: CoreEngineProcManager | CoreEngineActorManager | None |
| 1142 | coordinator: DPCoordinator | None |
| 1143 | addresses: EngineZmqAddresses |
| 1144 | tensor_queue: Queue | None |
| 1145 | # Frontend processes to watch during engine startup; may be assigned by |
| 1146 | # the caller before the startup barrier runs on context manager exit. |
| 1147 | watched_frontend_processes: Sequence[FrontendProcess] = () |
| 1148 | |
| 1149 | |
| 1150 | @contextlib.contextmanager |
| 1151 | def launch_core_engines( |
| 1152 | vllm_config: VllmConfig, |
| 1153 | executor_class: type[Executor], |
| 1154 | log_stats: bool, |
| 1155 | addresses: EngineZmqAddresses, |
| 1156 | ) -> Iterator[CoreEngineLaunch]: |
| 1157 | """Launch engine and DP coordinator processes as needed.""" |
| 1158 | parallel_config = vllm_config.parallel_config |
| 1159 | dp_size = parallel_config.data_parallel_size |
| 1160 | local_engine_count = parallel_config.data_parallel_size_local |
| 1161 | local_start_index = parallel_config.data_parallel_rank_local |
| 1162 | dp_rank = parallel_config.data_parallel_rank |
| 1163 | host = parallel_config.data_parallel_master_ip |
| 1164 | local_engines_only = parallel_config.local_engines_only |
| 1165 | |
| 1166 | offline_mode = local_start_index is not None |
| 1167 | |
| 1168 | # Create a single tensor IPC queue for sharing multimodal tensors between |
| 1169 | # API servers and engine core. Returns a single queue since we only support |
| 1170 | # DP=1 for this data flow. |
| 1171 | tensor_queue: Queue | None = None |
| 1172 | multimodal_config = vllm_config.model_config.multimodal_config |
| 1173 | if multimodal_config is not None and multimodal_config.mm_tensor_ipc == "torch_shm": |
| 1174 | tensor_queue = get_mp_context().Queue() |
| 1175 | |
| 1176 | # Run the DP Coordinator process with rank 0 when in online DP mode. |
| 1177 | # The coordinator is needed for: |
| 1178 | # 1. Internal/hybrid LB: collecting and publishing queue stats for load balancing |
| 1179 | # 2. MoE models: wave coordination in addition to stats |
| 1180 | run_coordinator = ( |
| 1181 | vllm_config.needs_dp_coordinator and not offline_mode and dp_rank == 0 |
| 1182 | ) |
| 1183 | |
| 1184 | if run_coordinator: |
| 1185 | coordinator = DPCoordinator( |
| 1186 | parallel_config, |
| 1187 | enable_wave_coordination=vllm_config.model_config.is_moe, |
| 1188 | logging_config=vllm_config.logging_config, |
| 1189 | ) |
| 1190 | |
| 1191 | addresses.coordinator_input, addresses.coordinator_output = ( |
| 1192 | coordinator.get_engine_socket_addresses() |
| 1193 | ) |
| 1194 | addresses.frontend_stats_publish_address = ( |
| 1195 | coordinator.get_stats_publish_address() |
| 1196 | ) |
| 1197 | |
| 1198 | logger.info("Started DP Coordinator process (PID: %d)", coordinator.proc.pid) |
| 1199 | else: |
| 1200 | coordinator = None |
| 1201 | |
| 1202 | if parallel_config.data_parallel_backend == "ray": |
| 1203 | logger.info("Starting ray-based data parallel backend") |
| 1204 | |
| 1205 | engine_actor_manager = CoreEngineActorManager( |
| 1206 | vllm_config=vllm_config, |
| 1207 | addresses=addresses, |
| 1208 | executor_class=executor_class, |
| 1209 | log_stats=log_stats, |
| 1210 | ) |
| 1211 | |
| 1212 | yield CoreEngineLaunch( |
| 1213 | engine_actor_manager, coordinator, addresses, tensor_queue |
| 1214 | ) |
| 1215 | return |
| 1216 | |
| 1217 | if offline_mode: |
| 1218 | assert local_engine_count == 1 |
| 1219 | engines_to_handshake = [CoreEngine(index=dp_rank, local=True)] |
| 1220 | elif dp_rank == 0: |
| 1221 | # Rank 0 holds Coordinator, so it handshakes with all Cores |
| 1222 | # in both external dplb and internal dplb mode. |
| 1223 | # Note this also covers the case where we have zero local engines |
| 1224 | # and rank 0 is headless. |
| 1225 | engines_to_handshake = [ |
| 1226 | CoreEngine(index=i, local=(i < local_engine_count)) for i in range(dp_size) |
| 1227 | ] |
| 1228 | else: |
| 1229 | # Rank > 0 handshakes with just the local cores it is managing. |
| 1230 | assert local_engines_only, ( |
| 1231 | "Attempting to launch core_engines from dp_rank > 0, but " |
| 1232 | "found internal DPLB, which is incompatible." |
| 1233 | ) |
| 1234 | engines_to_handshake = [ |
| 1235 | CoreEngine(index=i, local=True) |
| 1236 | for i in range(dp_rank, dp_rank + local_engine_count) |
| 1237 | ] |
| 1238 | |
| 1239 | # Whether the started engines will handshake only with co-located |
| 1240 | # front-end processes. In external_dp_lb mode, ranks > 0 handshake with |
| 1241 | # their co-located frontend and also the rank 0 front-end, and hence this |
| 1242 | # will be False. |
| 1243 | handshake_local_only = offline_mode or local_engine_count == dp_size |
| 1244 | |
| 1245 | # NOTE(yongji): handling scaling from intra-node to inter-node |
| 1246 | if parallel_config.enable_elastic_ep: |
| 1247 | handshake_local_only = False |
| 1248 | |
| 1249 | handshake_address = get_engine_client_zmq_addr( |
| 1250 | handshake_local_only, |
| 1251 | host, |
| 1252 | parallel_config.data_parallel_rpc_port, |
| 1253 | ) |
| 1254 | |
| 1255 | if local_engines_only and dp_rank > 0: |
| 1256 | assert not handshake_local_only |
| 1257 | local_handshake_address = get_open_zmq_ipc_path() |
| 1258 | client_handshake_address = local_handshake_address |
| 1259 | else: |
| 1260 | local_handshake_address = handshake_address |
| 1261 | client_handshake_address = None |
| 1262 | |
| 1263 | with zmq_socket_ctx( |
| 1264 | local_handshake_address, zmq.ROUTER, bind=True |
| 1265 | ) as handshake_socket: |
| 1266 | # Start local engines. |
| 1267 | if local_engine_count: |
| 1268 | local_engine_manager = CoreEngineProcManager( |
| 1269 | vllm_config=vllm_config, |
| 1270 | executor_class=executor_class, |
| 1271 | log_stats=log_stats, |
| 1272 | handshake_address=handshake_address, |
| 1273 | client_handshake_address=client_handshake_address, |
| 1274 | local_client=True, |
| 1275 | local_engine_count=local_engine_count, |
| 1276 | start_index=dp_rank, |
| 1277 | local_start_index=local_start_index or 0, |
| 1278 | tensor_queue=tensor_queue, |
| 1279 | ) |
| 1280 | else: |
| 1281 | local_engine_manager = None |
| 1282 | |
| 1283 | launch = CoreEngineLaunch( |
| 1284 | local_engine_manager, coordinator, addresses, tensor_queue |
| 1285 | ) |
| 1286 | yield launch |
| 1287 | wait_for_engine_startup( |
| 1288 | handshake_socket, |
| 1289 | engines_to_handshake, |
| 1290 | parallel_config, |
| 1291 | dp_size > 1 and vllm_config.model_config.is_moe, |
| 1292 | vllm_config.cache_config, |
| 1293 | launch, |
| 1294 | ) |
| 1295 | |
| 1296 | |
| 1297 | def wait_for_engine_startup( |
| 1298 | handshake_socket: zmq.Socket, |
| 1299 | core_engines: list[CoreEngine], |
| 1300 | parallel_config: ParallelConfig, |
| 1301 | coordinated_dp: bool, |
| 1302 | cache_config: CacheConfig, |
| 1303 | launch: CoreEngineLaunch, |
| 1304 | ): |
| 1305 | # Wait for engine core process(es) to send ready messages. |
| 1306 | local_count = parallel_config.data_parallel_size_local |
| 1307 | remote_count = len(core_engines) - local_count |
| 1308 | # [local, remote] counts |
| 1309 | conn_pending, start_pending = [local_count, remote_count], [0, 0] |
| 1310 | poller = zmq.Poller() |
| 1311 | poller.register(handshake_socket, zmq.POLLIN) |
| 1312 | |
| 1313 | remote_should_be_headless = ( |
| 1314 | not parallel_config.data_parallel_hybrid_lb |
| 1315 | and not parallel_config.data_parallel_external_lb |
| 1316 | ) |
| 1317 | |
| 1318 | # 1. Engine processes |
| 1319 | if isinstance(launch.engine_manager, CoreEngineProcManager): |
| 1320 | for sentinel in launch.engine_manager.sentinels(): |
| 1321 | poller.register(sentinel, zmq.POLLIN) |
| 1322 | # 2. DP Coordinator process, if present |
| 1323 | coord_process = launch.coordinator.proc if launch.coordinator else None |
| 1324 | if coord_process is not None: |
| 1325 | poller.register(coord_process.sentinel, zmq.POLLIN) |
| 1326 | # 3. Watched frontend processes, if any |
| 1327 | frontend_process_by_fd: dict[int, FrontendProcess] = {} |
| 1328 | for proc in launch.watched_frontend_processes: |
| 1329 | fd = proc.sentinel if isinstance(proc.sentinel, int) else proc.sentinel.fileno() |
| 1330 | frontend_process_by_fd[fd] = proc |
| 1331 | poller.register(fd, zmq.POLLIN) |
| 1332 | |
| 1333 | while any(conn_pending) or any(start_pending): |
| 1334 | events = poller.poll(STARTUP_POLL_PERIOD_MS) |
| 1335 | if not events: |
| 1336 | if any(conn_pending): |
| 1337 | logger.debug( |
| 1338 | "Waiting for %d local, %d remote core engine proc(s) to connect.", |
| 1339 | *conn_pending, |
| 1340 | ) |
| 1341 | if any(start_pending): |
| 1342 | logger.debug( |
| 1343 | "Waiting for %d local, %d remote core engine proc(s) to start.", |
| 1344 | *start_pending, |
| 1345 | ) |
| 1346 | continue |
| 1347 | if len(events) > 1 or events[0][0] != handshake_socket: |
| 1348 | # One of the local core, coordinator, or watched frontend processes exited. |
| 1349 | if isinstance(launch.engine_manager, CoreEngineProcManager): |
| 1350 | finished = launch.engine_manager.finished_procs() |
| 1351 | else: |
| 1352 | finished = {} |
| 1353 | if coord_process is not None and coord_process.exitcode is not None: |
| 1354 | finished[coord_process.name] = coord_process.exitcode |
| 1355 | failed_frontend_procs = { |
| 1356 | proc.name: proc.exitcode |
| 1357 | for fd, proc in frontend_process_by_fd.items() |
| 1358 | if proc.exitcode is not None |
| 1359 | or any(event_fd == fd for event_fd, _ in events) |
| 1360 | } |
| 1361 | if failed_frontend_procs and not finished: |
| 1362 | raise RuntimeError( |
| 1363 | "Frontend process failed during engine core initialization. " |
| 1364 | "See root cause above. " |
| 1365 | f"Failed frontend proc(s): {failed_frontend_procs}" |
| 1366 | ) |
| 1367 | raise RuntimeError( |
| 1368 | "Engine core initialization failed. " |
| 1369 | "See root cause above. " |
| 1370 | f"Failed core proc(s): {finished}" |
| 1371 | + ( |
| 1372 | f", failed frontend proc(s): {failed_frontend_procs}" |
| 1373 | if failed_frontend_procs |
| 1374 | else "" |
| 1375 | ) |
| 1376 | ) |
| 1377 | |
| 1378 | # Receive HELLO and READY messages from the input socket. |
| 1379 | eng_identity, ready_msg_bytes = handshake_socket.recv_multipart() |
| 1380 | eng_index = int.from_bytes(eng_identity, "little") |
| 1381 | engine = next((e for e in core_engines if e.identity == eng_identity), None) |
| 1382 | if engine is None: |
| 1383 | raise RuntimeError( |
| 1384 | f"Message from engine with unexpected data parallel rank: {eng_index}" |
| 1385 | ) |
| 1386 | msg = msgspec.msgpack.decode(ready_msg_bytes) |
| 1387 | status, local, headless = msg["status"], msg["local"], msg["headless"] |
| 1388 | if local != engine.local: |
| 1389 | raise RuntimeError( |
| 1390 | f"{status} message from " |
| 1391 | f"{'local' if local else 'remote'} " |
| 1392 | f"engine {eng_index}, expected it to be " |
| 1393 | f"{'local' if engine.local else 'remote'}" |
| 1394 | ) |
| 1395 | |
| 1396 | # Remote engines must be headless iff we aren't in hybrid dp lb mode. |
| 1397 | if not local and headless != remote_should_be_headless: |
| 1398 | if headless: |
| 1399 | raise RuntimeError( |
| 1400 | f"Remote engine {eng_index} must not use " |
| 1401 | f"--headless in external or hybrid dp lb " |
| 1402 | f"mode" |
| 1403 | ) |
| 1404 | else: |
| 1405 | raise RuntimeError( |
| 1406 | f"Remote engine {eng_index} must use " |
| 1407 | f"--headless unless in external or hybrid " |
| 1408 | f"dp lb mode" |
| 1409 | ) |
| 1410 | |
| 1411 | if status == "HELLO" and engine.state == CoreEngineState.NEW: |
| 1412 | # Send init message with DP config info. |
| 1413 | init_message = msgspec.msgpack.encode( |
| 1414 | EngineHandshakeMetadata( |
| 1415 | addresses=launch.addresses, |
| 1416 | parallel_config={ |
| 1417 | k: getattr(parallel_config, k) |
| 1418 | for k in ( |
| 1419 | "data_parallel_master_ip", |
| 1420 | "data_parallel_master_port", |
| 1421 | "_data_parallel_master_port_list", |
| 1422 | "data_parallel_size", |
| 1423 | ) |
| 1424 | } |
| 1425 | if coordinated_dp |
| 1426 | else {}, |
| 1427 | ) |
| 1428 | ) |
| 1429 | handshake_socket.send_multipart((eng_identity, init_message), copy=False) |
| 1430 | conn_pending[0 if local else 1] -= 1 |
| 1431 | start_pending[0 if local else 1] += 1 |
| 1432 | engine.state = CoreEngineState.CONNECTED |
| 1433 | elif status == "READY" and engine.state == CoreEngineState.CONNECTED: |
| 1434 | # Validate config hash consistency across DP workers for MoE models. |
| 1435 | if coordinated_dp: |
| 1436 | worker_config_hash = msg.get("parallel_config_hash") |
| 1437 | expected_hash = parallel_config.compute_hash() |
| 1438 | if worker_config_hash != expected_hash: |
| 1439 | raise RuntimeError( |
| 1440 | f"Configuration mismatch detected for engine " |
| 1441 | f"{eng_index}. All DP workers must have identical " |
| 1442 | f"configurations for parameters that affect collective " |
| 1443 | f"communication (e.g., enable_eplb, " |
| 1444 | f"eplb_config.log_balancedness). " |
| 1445 | f"Worker hash: {worker_config_hash}, " |
| 1446 | f"Expected hash: {expected_hash}. " |
| 1447 | f"Please ensure all workers are started with the same " |
| 1448 | f"command-line arguments." |
| 1449 | ) |
| 1450 | |
| 1451 | start_pending[0 if local else 1] -= 1 |
| 1452 | engine.state = CoreEngineState.READY |
| 1453 | else: |
| 1454 | raise RuntimeError( |
| 1455 | f"Unexpected {status} message for " |
| 1456 | f"{'local' if local else 'remote'} engine " |
| 1457 | f"{eng_index} in {engine.state} state." |
| 1458 | ) |
| 1459 | |
| 1460 | logger.debug( |
| 1461 | "%s from %s core engine process %s.", |
| 1462 | status, |
| 1463 | "local" if local else "remote", |
| 1464 | eng_index, |
| 1465 | ) |