| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import multiprocessing |
| 4 | import os |
| 5 | import pickle |
| 6 | import queue |
| 7 | import signal |
| 8 | import threading |
| 9 | import time |
| 10 | import traceback |
| 11 | import weakref |
| 12 | from collections import deque |
| 13 | from collections.abc import Callable, Sequence |
| 14 | from concurrent.futures import Future, InvalidStateError |
| 15 | from contextlib import suppress |
| 16 | from dataclasses import dataclass |
| 17 | from enum import Enum, auto |
| 18 | from functools import partial |
| 19 | from multiprocessing.connection import Connection |
| 20 | from multiprocessing.process import BaseProcess |
| 21 | from multiprocessing.synchronize import Lock as LockType |
| 22 | from threading import Thread |
| 23 | from typing import Any, cast |
| 24 | |
| 25 | import cloudpickle |
| 26 | import torch |
| 27 | |
| 28 | import vllm.envs as envs |
| 29 | from vllm.config import VllmConfig |
| 30 | from vllm.distributed import destroy_distributed_environment, destroy_model_parallel |
| 31 | from vllm.distributed.device_communicators.shm_broadcast import Handle, MessageQueue |
| 32 | from vllm.distributed.ec_transfer.ec_connector.utils import ECOutputAggregator |
| 33 | from vllm.distributed.kv_transfer.kv_connector.utils import KVOutputAggregator |
| 34 | from vllm.distributed.parallel_state import ( |
| 35 | get_dcp_group, |
| 36 | get_dp_group, |
| 37 | get_ep_group, |
| 38 | get_inner_dp_world_group, |
| 39 | get_pcp_group, |
| 40 | get_pp_group, |
| 41 | get_tp_group, |
| 42 | model_parallel_is_initialized, |
| 43 | ) |
| 44 | from vllm.envs import enable_envs_cache |
| 45 | from vllm.logger import configure_logging, init_logger |
| 46 | from vllm.platforms import current_platform |
| 47 | from vllm.tracing import instrument, maybe_init_worker_tracer |
| 48 | from vllm.utils import numa_utils |
| 49 | from vllm.utils.network_utils import ( |
| 50 | aiter_requires_tcp_store, |
| 51 | get_distributed_init_method, |
| 52 | get_file_store_init_method, |
| 53 | get_ip, |
| 54 | get_loopback_ip, |
| 55 | get_open_port, |
| 56 | ) |
| 57 | from vllm.utils.ompmultiprocessing import OMPProcessManager |
| 58 | from vllm.utils.system_utils import ( |
| 59 | _maybe_force_spawn, |
| 60 | decorate_logs, |
| 61 | get_mp_context, |
| 62 | set_process_title, |
| 63 | ) |
| 64 | from vllm.utils.torch_utils import ( |
| 65 | OMP_NUM_THREADS_SET_BY_VLLM, |
| 66 | set_torch_threads_for_runtime, |
| 67 | startup_omp_num_threads, |
| 68 | ) |
| 69 | from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput |
| 70 | from vllm.v1.executor.abstract import Executor, FailureCallback |
| 71 | from vllm.v1.executor.vllm_net_devices import set_worker_net_device |
| 72 | from vllm.v1.outputs import AsyncModelRunnerOutput, DraftTokenIds, ModelRunnerOutput |
| 73 | from vllm.v1.worker.worker_base import WorkerWrapperBase |
| 74 | |
| 75 | logger = init_logger(__name__) |
| 76 | |
| 77 | |
| 78 | class FutureWrapper(Future): |
| 79 | def __init__( |
| 80 | self, |
| 81 | futures_queue: deque["FutureWrapper"], |
| 82 | get_response: Callable[[], Any], |
| 83 | aggregate: Callable = lambda x: x, |
| 84 | ): |
| 85 | self.futures_queue = futures_queue |
| 86 | self.get_response = get_response |
| 87 | self.aggregate = aggregate |
| 88 | super().__init__() |
| 89 | self.futures_queue.appendleft(self) |
| 90 | |
| 91 | def result(self, timeout=None): |
| 92 | if timeout is not None: |
| 93 | raise RuntimeError("timeout not implemented") |
| 94 | |
| 95 | # Drain any futures ahead of us in the queue. |
| 96 | while not self.done(): |
| 97 | future = self.futures_queue.pop() |
| 98 | future._wait_for_response() |
| 99 | return super().result() |
| 100 | |
| 101 | def _wait_for_response(self): |
| 102 | try: |
| 103 | response = self.aggregate(self.get_response()) |
| 104 | with suppress(InvalidStateError): |
| 105 | self.set_result(response) |
| 106 | except Exception as e: |
| 107 | with suppress(InvalidStateError): |
| 108 | self.set_exception(e) |
| 109 | |
| 110 | |
| 111 | class MultiprocExecutor(Executor): |
| 112 | supports_pp: bool = True |
| 113 | |
| 114 | def __init__(self, vllm_config: VllmConfig, monitor_workers: bool = True): |
| 115 | self.monitor_workers = monitor_workers |
| 116 | super().__init__(vllm_config) |
| 117 | |
| 118 | def _init_executor(self) -> None: |
| 119 | # Call self.shutdown at exit to clean up |
| 120 | # and ensure workers will be terminated. |
| 121 | self._finalizer = weakref.finalize(self, self.shutdown) |
| 122 | self.is_failed = False |
| 123 | self.failure_callback: FailureCallback | None = None |
| 124 | |
| 125 | tp_size, pp_size, pcp_size = self._get_parallel_sizes() |
| 126 | assert self.world_size == tp_size * pp_size * pcp_size, ( |
| 127 | f"world_size ({self.world_size}) must be equal to the " |
| 128 | f"tensor_parallel_size ({tp_size}) x pipeline" |
| 129 | f"_parallel_size ({pp_size}) x prefill_context" |
| 130 | f"_parallel_size ({pcp_size}). " |
| 131 | ) |
| 132 | |
| 133 | num_local_procs = self.local_world_size * max( |
| 134 | 1, self.parallel_config.data_parallel_size_local |
| 135 | ) |
| 136 | set_multiprocessing_worker_envs(num_local_procs) |
| 137 | |
| 138 | if aiter_requires_tcp_store(): |
| 139 | distributed_init_method = get_distributed_init_method( |
| 140 | get_loopback_ip(), get_open_port() |
| 141 | ) |
| 142 | else: |
| 143 | distributed_init_method = get_file_store_init_method() |
| 144 | self.rpc_broadcast_mq: MessageQueue | None = None |
| 145 | scheduler_output_handle: Handle | None = None |
| 146 | # Initialize worker and set up message queues for SchedulerOutputs |
| 147 | # and ModelRunnerOutputs |
| 148 | if self.parallel_config.node_rank_within_dp == 0: |
| 149 | # For leader node within each dp rank, |
| 150 | # each dp will have its own leader multiproc executor. |
| 151 | max_chunk_bytes = envs.VLLM_MQ_MAX_CHUNK_BYTES_MB * 1024 * 1024 |
| 152 | mq_connect_ip = get_ip() |
| 153 | logger.info( |
| 154 | "DP group leader: node_rank=%d, node_rank_within_dp=%d, " |
| 155 | "master_addr=%s, mq_connect_ip=%s (local), " |
| 156 | "world_size=%d, local_world_size=%d", |
| 157 | self.parallel_config.node_rank, |
| 158 | self.parallel_config.node_rank_within_dp, |
| 159 | self.parallel_config.master_addr, |
| 160 | mq_connect_ip, |
| 161 | self.world_size, |
| 162 | self.local_world_size, |
| 163 | ) |
| 164 | self.rpc_broadcast_mq = MessageQueue( |
| 165 | self.world_size, |
| 166 | self.local_world_size, |
| 167 | max_chunk_bytes=max_chunk_bytes, |
| 168 | connect_ip=mq_connect_ip, |
| 169 | ) |
| 170 | scheduler_output_handle = self.rpc_broadcast_mq.export_handle() |
| 171 | # Create workers |
| 172 | context = get_mp_context() |
| 173 | shared_worker_lock = context.Lock() |
| 174 | unready_workers: list[UnreadyWorkerProcHandle] = [] |
| 175 | success = False |
| 176 | try: |
| 177 | global_start_rank = ( |
| 178 | self.local_world_size * self.parallel_config.node_rank_within_dp |
| 179 | ) |
| 180 | # When using fork, keep track of socket file descriptors that are |
| 181 | # inherited by the worker, so that we can close them in subsequent |
| 182 | # workers |
| 183 | inherited_fds: list[int] | None = ( |
| 184 | [] if context.get_start_method() == "fork" else None |
| 185 | ) |
| 186 | |
| 187 | # For CPU backend only, to setup OpenMP threads affinity |
| 188 | cpu_omp_manager = OMPProcessManager(self.vllm_config) |
| 189 | for local_rank in range(self.local_world_size): |
| 190 | global_rank = global_start_rank + local_rank |
| 191 | is_driver_worker = self._is_driver_worker(global_rank) |
| 192 | with cpu_omp_manager.configure_omp_envs( |
| 193 | rank=global_rank, local_rank=local_rank |
| 194 | ): |
| 195 | unready_worker_handle = WorkerProc.make_worker_process( |
| 196 | vllm_config=self.vllm_config, |
| 197 | local_rank=local_rank, |
| 198 | rank=global_rank, |
| 199 | distributed_init_method=distributed_init_method, |
| 200 | input_shm_handle=scheduler_output_handle, |
| 201 | shared_worker_lock=shared_worker_lock, |
| 202 | is_driver_worker=is_driver_worker, |
| 203 | inherited_fds=inherited_fds, |
| 204 | ) |
| 205 | unready_workers.append(unready_worker_handle) |
| 206 | if inherited_fds is not None: |
| 207 | inherited_fds.append(unready_worker_handle.death_writer.fileno()) |
| 208 | inherited_fds.append(unready_worker_handle.ready_pipe.fileno()) |
| 209 | |
| 210 | # Workers must be created before wait_for_ready to avoid |
| 211 | # deadlock, since worker.init_device() does a device sync. |
| 212 | |
| 213 | # Wait for all local workers to be ready. |
| 214 | self.workers = WorkerProc.wait_for_ready(unready_workers) |
| 215 | |
| 216 | # The workers have inherited their thread count (see |
| 217 | # set_multiprocessing_worker_envs); this process only schedules, so |
| 218 | # it gets no benefit from torch intra-op parallelism, just CPU |
| 219 | # contention with them. |
| 220 | set_torch_threads_for_runtime() |
| 221 | |
| 222 | # Start background thread to monitor worker health if not in headless mode. |
| 223 | if self.monitor_workers: |
| 224 | self.start_worker_monitor() |
| 225 | |
| 226 | self.response_mqs = [] |
| 227 | # Only leader node have remote response mqs |
| 228 | if self.parallel_config.node_rank_within_dp == 0: |
| 229 | for rank in range(self.world_size): |
| 230 | if rank < self.local_world_size: |
| 231 | local_message_queue = self.workers[rank].worker_response_mq |
| 232 | assert local_message_queue is not None |
| 233 | self.response_mqs.append(local_message_queue) |
| 234 | else: |
| 235 | remote_message_queue = self.workers[0].peer_worker_response_mqs[ |
| 236 | rank |
| 237 | ] |
| 238 | assert remote_message_queue is not None |
| 239 | self.response_mqs.append(remote_message_queue) |
| 240 | |
| 241 | # Ensure message queues are ready. Will deadlock if re-ordered |
| 242 | # Must be kept consistent with the WorkerProc. |
| 243 | |
| 244 | # Wait for all input mqs to be ready. |
| 245 | if self.rpc_broadcast_mq is not None: |
| 246 | self.rpc_broadcast_mq.wait_until_ready() |
| 247 | # Wait for all remote response mqs to be ready. |
| 248 | for response_mq in self.response_mqs: |
| 249 | response_mq.wait_until_ready() |
| 250 | |
| 251 | self.futures_queue = deque[FutureWrapper]() |
| 252 | |
| 253 | self._post_init_executor() |
| 254 | |
| 255 | success = True |
| 256 | finally: |
| 257 | if not success: |
| 258 | # Clean up the worker procs if there was a failure. |
| 259 | # Close death_writers first to signal workers to exit |
| 260 | for uw in unready_workers: |
| 261 | if uw.death_writer is not None: |
| 262 | uw.death_writer.close() |
| 263 | uw.death_writer = None |
| 264 | self._ensure_worker_termination([uw.proc for uw in unready_workers]) |
| 265 | |
| 266 | self.output_rank = self._get_output_rank() |
| 267 | |
| 268 | def get_response_mqs(self, unique_reply_rank: int = -1) -> list[MessageQueue]: |
| 269 | assert unique_reply_rank >= -1 and unique_reply_rank < self.world_size, ( |
| 270 | f"unique_reply_rank must be -1 or < world_size," |
| 271 | f"unique_reply_rank = {unique_reply_rank}, " |
| 272 | f"world_size={self.world_size}" |
| 273 | ) |
| 274 | ranks = ( |
| 275 | [unique_reply_rank] if unique_reply_rank != -1 else range(self.world_size) |
| 276 | ) |
| 277 | return [self.workers[rank].worker_response_mq for rank in ranks] |
| 278 | |
| 279 | def _get_parallel_sizes(self) -> tuple[int, int, int]: |
| 280 | self.world_size = self.parallel_config.world_size |
| 281 | assert self.world_size % self.parallel_config.nnodes_within_dp == 0, ( |
| 282 | f"global world_size ({self.parallel_config.world_size}) must be " |
| 283 | f"divisible by nnodes_within_dp " |
| 284 | f"({self.parallel_config.nnodes_within_dp}). " |
| 285 | ) |
| 286 | self.local_world_size = self.parallel_config.local_world_size |
| 287 | tp_size = self.parallel_config.tensor_parallel_size |
| 288 | pp_size = self.parallel_config.pipeline_parallel_size |
| 289 | pcp_size = self.parallel_config.prefill_context_parallel_size |
| 290 | return tp_size, pp_size, pcp_size |
| 291 | |
| 292 | def _post_init_executor(self) -> None: |
| 293 | pass |
| 294 | |
| 295 | def _is_driver_worker(self, rank: int) -> bool: |
| 296 | return rank % self.parallel_config.tensor_parallel_size == 0 |
| 297 | |
| 298 | def start_worker_monitor(self, inline=False) -> None: |
| 299 | workers = self.workers |
| 300 | self_ref = weakref.ref(self) |
| 301 | |
| 302 | # Monitors worker process liveness. If any die unexpectedly, |
| 303 | # logs an error, shuts down the executor and invokes the failure |
| 304 | # callback to inform the engine. |
| 305 | def monitor_workers(): |
| 306 | sentinels = [h.proc.sentinel for h in workers] |
| 307 | died = multiprocessing.connection.wait(sentinels) |
| 308 | _self = self_ref() |
| 309 | if not _self or getattr(_self, "shutting_down", False): |
| 310 | logger.debug("MultiprocWorkerMonitor: shutdown already initiated") |
| 311 | return |
| 312 | _self.is_failed = True |
| 313 | proc = next(h.proc for h in workers if h.proc.sentinel == died[0]) |
| 314 | logger.error( |
| 315 | "Worker proc %s died unexpectedly (exit code: %s), " |
| 316 | "shutting down executor.", |
| 317 | proc.name, |
| 318 | proc.exitcode, |
| 319 | ) |
| 320 | _self.shutdown() |
| 321 | callback = _self.failure_callback |
| 322 | if callback is not None: |
| 323 | _self.failure_callback = None |
| 324 | callback() |
| 325 | |
| 326 | if not inline: |
| 327 | Thread( |
| 328 | target=monitor_workers, daemon=True, name="MultiprocWorkerMonitor" |
| 329 | ).start() |
| 330 | return |
| 331 | |
| 332 | monitor_workers() |
| 333 | |
| 334 | def register_failure_callback(self, callback: FailureCallback): |
| 335 | if self.is_failed: |
| 336 | callback() |
| 337 | else: |
| 338 | self.failure_callback = callback |
| 339 | |
| 340 | def execute_model( # type: ignore[override] |
| 341 | self, scheduler_output: SchedulerOutput, non_block: bool = False |
| 342 | ) -> ModelRunnerOutput | None | Future[ModelRunnerOutput | None]: |
| 343 | return self.collective_rpc( |
| 344 | "execute_model", |
| 345 | args=(scheduler_output,), |
| 346 | unique_reply_rank=self.output_rank, |
| 347 | non_block=non_block, |
| 348 | timeout=envs.VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS, |
| 349 | kv_output_aggregator=self.kv_output_aggregator, |
| 350 | ec_output_aggregator=self.ec_output_aggregator, |
| 351 | ) |
| 352 | |
| 353 | def sample_tokens( # type: ignore[override] |
| 354 | self, grammar_output: GrammarOutput | None, non_block: bool = False |
| 355 | ) -> ModelRunnerOutput | Future[ModelRunnerOutput]: |
| 356 | return self.collective_rpc( |
| 357 | "sample_tokens", |
| 358 | args=(grammar_output,), |
| 359 | unique_reply_rank=self.output_rank, |
| 360 | non_block=non_block, |
| 361 | timeout=envs.VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS, |
| 362 | kv_output_aggregator=self.kv_output_aggregator, |
| 363 | ec_output_aggregator=self.ec_output_aggregator, |
| 364 | ) |
| 365 | |
| 366 | def execute_dummy_batch(self) -> None: |
| 367 | self.collective_rpc("execute_dummy_batch", unique_reply_rank=self.output_rank) |
| 368 | |
| 369 | def take_draft_token_ids(self) -> DraftTokenIds | None: |
| 370 | # OPTIMIZATION: Get output only from a single worker (output_rank) |
| 371 | return self.collective_rpc( |
| 372 | "take_draft_token_ids", |
| 373 | unique_reply_rank=self.output_rank, |
| 374 | timeout=envs.VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS, |
| 375 | ) |
| 376 | |
| 377 | def collective_rpc( # type: ignore[override] |
| 378 | self, |
| 379 | method: str | Callable, |
| 380 | timeout: float | None = None, |
| 381 | args: tuple = (), |
| 382 | kwargs: dict | None = None, |
| 383 | non_block: bool = False, |
| 384 | unique_reply_rank: int | None = None, |
| 385 | kv_output_aggregator: KVOutputAggregator | None = None, |
| 386 | ec_output_aggregator: ECOutputAggregator | None = None, |
| 387 | ) -> Any: |
| 388 | """Returns single result if unique_reply_rank and/or an output |
| 389 | aggregator is provided, otherwise list.""" |
| 390 | assert self.rpc_broadcast_mq is not None, ( |
| 391 | "collective_rpc should not be called on follower node" |
| 392 | ) |
| 393 | if self.is_failed: |
| 394 | raise RuntimeError("Executor failed.") |
| 395 | |
| 396 | deadline = None if timeout is None else time.monotonic() + timeout |
| 397 | kwargs = kwargs or {} |
| 398 | |
| 399 | aggregators = [a for a in (kv_output_aggregator, ec_output_aggregator) if a] |
| 400 | aggregate: Callable[[Any], Any] |
| 401 | if aggregators: |
| 402 | output_rank = None |
| 403 | |
| 404 | def _aggregate(outputs: Any) -> Any: |
| 405 | # Each aggregator merges its own connector's output onto |
| 406 | # outputs[output_rank] in place and returns it, so chaining is safe. |
| 407 | rank = unique_reply_rank or 0 |
| 408 | result = outputs[rank] |
| 409 | for a in aggregators: |
| 410 | result = a.aggregate(outputs, output_rank=rank) |
| 411 | return result |
| 412 | |
| 413 | aggregate = _aggregate |
| 414 | else: |
| 415 | output_rank = unique_reply_rank |
| 416 | aggregate = lambda x: x |
| 417 | |
| 418 | if isinstance(method, str): |
| 419 | send_method = method |
| 420 | else: |
| 421 | send_method = cloudpickle.dumps(method, protocol=pickle.HIGHEST_PROTOCOL) |
| 422 | self.rpc_broadcast_mq.enqueue((send_method, args, kwargs, output_rank)) |
| 423 | |
| 424 | response_mqs: Sequence[MessageQueue] = self.response_mqs |
| 425 | if output_rank is not None: |
| 426 | response_mqs = (response_mqs[output_rank],) |
| 427 | |
| 428 | def get_response(): |
| 429 | responses = [] |
| 430 | for mq in response_mqs: |
| 431 | dequeue_timeout = ( |
| 432 | None if deadline is None else max(0.0, deadline - time.monotonic()) |
| 433 | ) |
| 434 | try: |
| 435 | status, result = mq.dequeue(timeout=dequeue_timeout) |
| 436 | except TimeoutError as e: |
| 437 | raise TimeoutError(f"RPC call to {method} timed out.") from e |
| 438 | if status != WorkerProc.ResponseStatus.SUCCESS: |
| 439 | raise RuntimeError( |
| 440 | f"Worker failed with error '{result}', please check the" |
| 441 | " stack trace above for the root cause" |
| 442 | ) |
| 443 | responses.append(result) |
| 444 | return responses[0] if output_rank is not None else responses |
| 445 | |
| 446 | future = FutureWrapper( |
| 447 | self.futures_queue, get_response=get_response, aggregate=aggregate |
| 448 | ) |
| 449 | |
| 450 | return future if non_block else future.result() |
| 451 | |
| 452 | @staticmethod |
| 453 | def _ensure_worker_termination(worker_procs: list[BaseProcess]): |
| 454 | """Ensure that all worker processes are terminated. Assumes workers have |
| 455 | received termination requests. Waits for processing, then sends |
| 456 | termination and kill signals if needed.""" |
| 457 | |
| 458 | def wait_for_termination(procs, timeout): |
| 459 | if not time: |
| 460 | # If we are in late stage shutdown, the interpreter may replace |
| 461 | # `time` with `None`. |
| 462 | return all(not proc.is_alive() for proc in procs) |
| 463 | start_time = time.time() |
| 464 | while time.time() - start_time < timeout: |
| 465 | if all(not proc.is_alive() for proc in procs): |
| 466 | return True |
| 467 | time.sleep(0.1) |
| 468 | return False |
| 469 | |
| 470 | active_procs = lambda: [proc for proc in worker_procs if proc.is_alive()] |
| 471 | initial_count = len(active_procs()) |
| 472 | |
| 473 | # Give processes time to clean themselves up properly first |
| 474 | logger.info( |
| 475 | "[shutdown] Executor: waiting for worker exit count=%d", |
| 476 | initial_count, |
| 477 | ) |
| 478 | if wait_for_termination( |
| 479 | active_procs(), timeout=envs.VLLM_WORKER_SHUTDOWN_TIMEOUT_SECONDS |
| 480 | ): |
| 481 | logger.info_once("[shutdown] Executor: all workers exited gracefully") |
| 482 | return |
| 483 | |
| 484 | # Send SIGTERM if still running |
| 485 | remaining = active_procs() |
| 486 | logger.warning( |
| 487 | "[shutdown] Executor: workers still running after grace period; " |
| 488 | "sending SIGTERM count=%d", |
| 489 | len(remaining), |
| 490 | ) |
| 491 | for p in remaining: |
| 492 | p.terminate() |
| 493 | if not wait_for_termination(active_procs(), 4): |
| 494 | # Send SIGKILL if still running |
| 495 | remaining = active_procs() |
| 496 | logger.warning( |
| 497 | "[shutdown] Executor: workers still running after SIGTERM; " |
| 498 | "sending SIGKILL count=%d", |
| 499 | len(remaining), |
| 500 | ) |
| 501 | for p in remaining: |
| 502 | p.kill() |
| 503 | |
| 504 | def shutdown(self): |
| 505 | """Properly shut down the executor and its workers.""" |
| 506 | if not getattr(self, "shutting_down", False): |
| 507 | worker_count = len(getattr(self, "workers", None) or []) |
| 508 | logger.debug( |
| 509 | "[shutdown] Executor: start worker_count=%d", |
| 510 | worker_count, |
| 511 | ) |
| 512 | self.shutting_down = True |
| 513 | |
| 514 | # Make sure all the worker processes are terminated first. |
| 515 | if workers := getattr(self, "workers", None): |
| 516 | for w in workers: |
| 517 | # Close death_writer to signal child processes to exit |
| 518 | if w.death_writer is not None: |
| 519 | w.death_writer.close() |
| 520 | w.death_writer = None |
| 521 | self._ensure_worker_termination([w.proc for w in workers]) |
| 522 | |
| 523 | for w in workers: |
| 524 | # Shutdown response queues |
| 525 | if w.worker_response_mq is not None: |
| 526 | w.worker_response_mq.shutdown() |
| 527 | w.worker_response_mq = None |
| 528 | |
| 529 | if rpc_broadcast_mq := getattr(self, "rpc_broadcast_mq", None): |
| 530 | rpc_broadcast_mq.shutdown() |
| 531 | self.rpc_broadcast_mq = None |
| 532 | if response_mqs := getattr(self, "response_mqs", None): |
| 533 | for mq in response_mqs: |
| 534 | mq.shutdown() |
| 535 | self.response_mqs = [] |
| 536 | |
| 537 | logger.debug_once("[shutdown] Executor: complete") |
| 538 | |
| 539 | def check_health(self) -> None: |
| 540 | self.collective_rpc("check_health", timeout=10) |
| 541 | return |
| 542 | |
| 543 | def _get_output_rank(self) -> int: |
| 544 | # Only returns ModelRunnerOutput from TP rank=0 and PP rank=-1 |
| 545 | # (the first TP worker of the last PP stage). |
| 546 | # Example: |
| 547 | # Assuming TP=8, PP=4, then the world_size=32 |
| 548 | # 0-7, PP rank 0 |
| 549 | # 8-15, PP rank 1 |
| 550 | # 16-23, PP rank 2 |
| 551 | # 24-31, PP rank 3 |
| 552 | # so world_size - tp_size = 32 - 8 = 24 should be PP rank = -1 (i.e. 3) |
| 553 | return ( |
| 554 | self.world_size |
| 555 | - self.parallel_config.tensor_parallel_size |
| 556 | * self.parallel_config.prefill_context_parallel_size |
| 557 | ) |
| 558 | |
| 559 | @classmethod |
| 560 | def supports_async_scheduling(cls) -> bool: |
| 561 | return True |
| 562 | |
| 563 | |
| 564 | @dataclass |
| 565 | class UnreadyWorkerProcHandle: |
| 566 | """WorkerProcess handle before READY.""" |
| 567 | |
| 568 | proc: BaseProcess |
| 569 | rank: int |
| 570 | ready_pipe: Connection |
| 571 | death_writer: Connection | None = None |
| 572 | |
| 573 | |
| 574 | @dataclass |
| 575 | class WorkerProcHandle: |
| 576 | proc: BaseProcess |
| 577 | rank: int |
| 578 | # The worker process writes to this MQ in single-node mode |
| 579 | worker_response_mq: MessageQueue | None |
| 580 | # This is only non empty on driver node, |
| 581 | # the peer worker process i writes to MQ |
| 582 | # `peer_worker_response_mqs[i]` |
| 583 | peer_worker_response_mqs: list[MessageQueue | None] |
| 584 | death_writer: Connection | None = None |
| 585 | |
| 586 | @classmethod |
| 587 | def from_unready_handle( |
| 588 | cls, |
| 589 | unready_handle: UnreadyWorkerProcHandle, |
| 590 | worker_response_mq: MessageQueue | None, |
| 591 | peer_worker_response_mqs: list[MessageQueue | None], |
| 592 | ) -> "WorkerProcHandle": |
| 593 | return cls( |
| 594 | proc=unready_handle.proc, |
| 595 | rank=unready_handle.rank, |
| 596 | worker_response_mq=worker_response_mq, |
| 597 | peer_worker_response_mqs=peer_worker_response_mqs, |
| 598 | death_writer=unready_handle.death_writer, |
| 599 | ) |
| 600 | |
| 601 | |
| 602 | class WorkerProc: |
| 603 | """Wrapper that runs one Worker in a separate process.""" |
| 604 | |
| 605 | READY_STR = "READY" |
| 606 | rpc_broadcast_mq: MessageQueue | None |
| 607 | worker_response_mq: MessageQueue | None |
| 608 | |
| 609 | def _init_message_queues( |
| 610 | self, input_shm_handle: Handle, vllm_config: VllmConfig |
| 611 | ) -> None: |
| 612 | if vllm_config.parallel_config.nnodes_within_dp == 1: |
| 613 | # Initialize MessageQueue for receiving SchedulerOutput |
| 614 | self.rpc_broadcast_mq = MessageQueue.create_from_handle( |
| 615 | input_shm_handle, self.worker.rank |
| 616 | ) |
| 617 | |
| 618 | # Initializes a message queue for sending the model output |
| 619 | self.worker_response_mq = MessageQueue(1, 1) |
| 620 | self.peer_response_handles = [] |
| 621 | else: |
| 622 | # Initialize remote MessageQueue for receiving SchedulerOutput across nodes |
| 623 | self.rpc_broadcast_mq = get_inner_dp_world_group().create_mq_broadcaster( |
| 624 | external_writer_handle=input_shm_handle, |
| 625 | # Since there is external_writer_handle from executor proc, |
| 626 | # where the ready signal from actual writer is sent out of the |
| 627 | # create_mq_broadcaster method and after this setup, we make it |
| 628 | # non blocking. The handshake will be triggered when |
| 629 | # worker.rpc_broadcast_mq.wait_until_ready() is called |
| 630 | blocking=False, |
| 631 | ) |
| 632 | # Initializes remote message queue for sending the model output to the |
| 633 | # driver worker, exposing peer_response_handles for driver worker |
| 634 | # that include handles for all ranks |
| 635 | self.worker_response_mq, self.peer_response_handles = ( |
| 636 | get_inner_dp_world_group().create_single_reader_mq_broadcasters( |
| 637 | reader_rank_in_group=0 |
| 638 | ) |
| 639 | ) |
| 640 | |
| 641 | @instrument(span_name="Worker init") |
| 642 | def __init__( |
| 643 | self, |
| 644 | vllm_config: VllmConfig, |
| 645 | local_rank: int, |
| 646 | rank: int, |
| 647 | distributed_init_method: str, |
| 648 | input_shm_handle: Handle, |
| 649 | shared_worker_lock: LockType, |
| 650 | is_driver_worker: bool, |
| 651 | ): |
| 652 | self.rank = rank |
| 653 | wrapper = WorkerWrapperBase(rpc_rank=local_rank, global_rank=rank) |
| 654 | # TODO: move `init_worker` to executor level as a collective rpc call |
| 655 | all_kwargs: list[dict] = [ |
| 656 | {} for _ in range(vllm_config.parallel_config.world_size) |
| 657 | ] |
| 658 | all_kwargs[local_rank] = { |
| 659 | "vllm_config": vllm_config, |
| 660 | "local_rank": local_rank, |
| 661 | "rank": rank, |
| 662 | "distributed_init_method": distributed_init_method, |
| 663 | "is_driver_worker": is_driver_worker, |
| 664 | "shared_worker_lock": shared_worker_lock, |
| 665 | } |
| 666 | wrapper.init_worker(all_kwargs) |
| 667 | self.worker = wrapper |
| 668 | |
| 669 | self.setup_proc_title_and_log_prefix( |
| 670 | enable_ep=vllm_config.parallel_config.enable_expert_parallel |
| 671 | ) |
| 672 | |
| 673 | # Load model |
| 674 | self.worker.init_device() |
| 675 | # Update process title now that parallel groups are initialized |
| 676 | self.setup_proc_title_and_log_prefix( |
| 677 | enable_ep=vllm_config.parallel_config.enable_expert_parallel |
| 678 | ) |
| 679 | if envs.VLLM_ELASTIC_EP_SCALE_UP_LAUNCH: |
| 680 | self.worker.elastic_ep_execute("load_model") |
| 681 | else: |
| 682 | self.worker.load_model() |
| 683 | |
| 684 | scheduler_config = vllm_config.scheduler_config |
| 685 | self.use_async_scheduling = scheduler_config.async_scheduling |
| 686 | if self.use_async_scheduling: |
| 687 | self.async_output_queue: queue.Queue = queue.Queue() |
| 688 | self.async_output_copy_thread = Thread( |
| 689 | target=self.async_output_busy_loop, |
| 690 | daemon=True, |
| 691 | name="WorkerAsyncOutputCopy", |
| 692 | ) |
| 693 | self.async_output_copy_thread.start() |
| 694 | |
| 695 | # Set block size based on the attention backends |
| 696 | current_platform.update_block_size_for_backend(vllm_config) |
| 697 | |
| 698 | # Initialize message queues after init_device() since multi-node setups |
| 699 | # (nnodes_within_dp > 1) require distributed groups to be initialized |
| 700 | self._init_message_queues(input_shm_handle, vllm_config) |
| 701 | |
| 702 | # Enable environment variable cache (e.g. assume no more |
| 703 | # environment variable overrides after this point) |
| 704 | enable_envs_cache() |
| 705 | |
| 706 | @staticmethod |
| 707 | def make_worker_process( |
| 708 | vllm_config: VllmConfig, |
| 709 | local_rank: int, |
| 710 | rank: int, |
| 711 | distributed_init_method: str, |
| 712 | input_shm_handle, # Receive SchedulerOutput |
| 713 | shared_worker_lock: LockType, |
| 714 | is_driver_worker: bool, |
| 715 | inherited_fds: list[int] | None = None, |
| 716 | ) -> UnreadyWorkerProcHandle: |
| 717 | context = get_mp_context() |
| 718 | # Ready pipe to communicate readiness from child to parent |
| 719 | ready_reader, ready_writer = context.Pipe(duplex=False) |
| 720 | # Death pipe to let child detect parent process exit |
| 721 | death_reader, death_writer = context.Pipe(duplex=False) |
| 722 | if inherited_fds is not None: |
| 723 | inherited_fds = inherited_fds.copy() |
| 724 | inherited_fds.extend((ready_reader.fileno(), death_writer.fileno())) |
| 725 | process_kwargs = { |
| 726 | "vllm_config": vllm_config, |
| 727 | "local_rank": local_rank, |
| 728 | "rank": rank, |
| 729 | "distributed_init_method": distributed_init_method, |
| 730 | "input_shm_handle": input_shm_handle, |
| 731 | "ready_pipe": ready_writer, |
| 732 | "death_pipe": death_reader, |
| 733 | "shared_worker_lock": shared_worker_lock, |
| 734 | "is_driver_worker": is_driver_worker, |
| 735 | # Have the worker close parent end of this worker's pipes too |
| 736 | "inherited_fds": inherited_fds if inherited_fds is not None else [], |
| 737 | } |
| 738 | # Run EngineCore busy loop in background process. |
| 739 | proc = context.Process( |
| 740 | target=WorkerProc.worker_main, |
| 741 | kwargs=process_kwargs, |
| 742 | name=f"VllmWorker-{rank}", |
| 743 | daemon=True, |
| 744 | ) |
| 745 | |
| 746 | # Apply NUMA binding if configured |
| 747 | with numa_utils.configure_subprocess( |
| 748 | vllm_config, local_rank, process_kind="worker" |
| 749 | ): |
| 750 | proc.start() |
| 751 | |
| 752 | # Close child ends of pipes here in the parent |
| 753 | ready_writer.close() |
| 754 | death_reader.close() |
| 755 | # Keep death_writer open in parent - when parent exits, |
| 756 | # death_reader in child will get EOFError |
| 757 | return UnreadyWorkerProcHandle(proc, rank, ready_reader, death_writer) |
| 758 | |
| 759 | @staticmethod |
| 760 | def wait_for_response_handle_ready( |
| 761 | handles: dict[str, Any], proc_handle: UnreadyWorkerProcHandle |
| 762 | ) -> WorkerProcHandle: |
| 763 | response_handle = handles["handle"] |
| 764 | worker_response_mq: MessageQueue | None = None |
| 765 | if len(response_handle.local_reader_ranks) > 0: |
| 766 | worker_response_mq = MessageQueue.create_from_handle(response_handle, 0) |
| 767 | peer_response_handles = handles["peer_response_handles"] |
| 768 | peer_worker_response_mqs = [ |
| 769 | MessageQueue.create_from_handle(handle, -1) |
| 770 | if handle.remote_subscribe_addr is not None |
| 771 | else None |
| 772 | for handle in peer_response_handles |
| 773 | ] |
| 774 | return WorkerProcHandle.from_unready_handle( |
| 775 | proc_handle, |
| 776 | worker_response_mq, |
| 777 | peer_worker_response_mqs=peer_worker_response_mqs, |
| 778 | ) |
| 779 | |
| 780 | @staticmethod |
| 781 | def wait_for_ready( |
| 782 | unready_proc_handles: list[UnreadyWorkerProcHandle], |
| 783 | ) -> list[WorkerProcHandle]: |
| 784 | e = Exception( |
| 785 | "WorkerProc initialization failed due to an exception in a " |
| 786 | "background process. See stack trace for root cause." |
| 787 | ) |
| 788 | |
| 789 | pipes = {handle.ready_pipe: handle for handle in unready_proc_handles} |
| 790 | ready_proc_handles: list[WorkerProcHandle | None] = [None] * len( |
| 791 | unready_proc_handles |
| 792 | ) |
| 793 | while pipes: |
| 794 | ready = multiprocessing.connection.wait(pipes.keys()) |
| 795 | for pipe in ready: |
| 796 | assert isinstance(pipe, Connection) |
| 797 | try: |
| 798 | # Wait until the WorkerProc is ready. |
| 799 | unready_proc_handle = pipes.pop(pipe) |
| 800 | response: dict[str, Any] = pipe.recv() |
| 801 | if response["status"] != "READY": |
| 802 | raise e |
| 803 | |
| 804 | idx = unready_proc_handle.rank % len(ready_proc_handles) |
| 805 | ready_proc_handles[idx] = WorkerProc.wait_for_response_handle_ready( |
| 806 | response, unready_proc_handle |
| 807 | ) |
| 808 | except EOFError: |
| 809 | e.__suppress_context__ = True |
| 810 | raise e from None |
| 811 | |
| 812 | finally: |
| 813 | # Close connection. |
| 814 | pipe.close() |
| 815 | |
| 816 | return cast(list[WorkerProcHandle], ready_proc_handles) |
| 817 | |
| 818 | def shutdown(self): |
| 819 | if self.rpc_broadcast_mq is not None: |
| 820 | self.rpc_broadcast_mq.shutdown() |
| 821 | if self.worker_response_mq is not None: |
| 822 | self.worker_response_mq.shutdown() |
| 823 | self.worker.shutdown() |
| 824 | self.rpc_broadcast_mq = None |
| 825 | self.worker_response_mq = None |
| 826 | destroy_model_parallel() |
| 827 | destroy_distributed_environment() |
| 828 | |
| 829 | def monitor_death_pipe(self, death_pipe, shutdown_requested: threading.Event): |
| 830 | if death_pipe is None: |
| 831 | return |
| 832 | |
| 833 | def death_pipe_monitor(queues_to_shutdown: list[MessageQueue]): |
| 834 | try: |
| 835 | # This will block until parent process exits (pipe closes) |
| 836 | death_pipe.recv() |
| 837 | except EOFError: |
| 838 | logger.info_once("Parent process exited, terminating worker queues") |
| 839 | shutdown_requested.set() |
| 840 | for mq in queues_to_shutdown: |
| 841 | if mq is not None: |
| 842 | mq.shutdown() |
| 843 | except Exception as e: |
| 844 | logger.warning("Death monitoring error: %s", e) |
| 845 | |
| 846 | # Pass queue references directly to avoid gc issues if passing self |
| 847 | Thread( |
| 848 | target=death_pipe_monitor, |
| 849 | args=([self.rpc_broadcast_mq, self.worker_response_mq],), |
| 850 | daemon=True, |
| 851 | name="DeathPipeMonitor", |
| 852 | ).start() |
| 853 | |
| 854 | @staticmethod |
| 855 | def worker_main(*args, **kwargs): |
| 856 | """Worker initialization and execution loops. |
| 857 | This runs a background process""" |
| 858 | if logging_config := getattr(kwargs["vllm_config"], "logging_config", None): |
| 859 | configure_logging(logging_config) |
| 860 | |
| 861 | # Signal handler used for graceful termination. |
| 862 | # SystemExit exception is only raised once to allow this and worker |
| 863 | # processes to terminate without error |
| 864 | shutdown_requested = threading.Event() |
| 865 | |
| 866 | def signal_handler(signum, frame): |
| 867 | nonlocal shutdown_requested |
| 868 | if not shutdown_requested.is_set(): |
| 869 | shutdown_requested.set() |
| 870 | logger.debug( |
| 871 | "WorkerProc handling signal %d, raising SystemExit", signum |
| 872 | ) |
| 873 | raise SystemExit() |
| 874 | |
| 875 | # Either SIGTERM or SIGINT will terminate the worker |
| 876 | signal.signal(signal.SIGTERM, signal_handler) |
| 877 | signal.signal(signal.SIGINT, signal_handler) |
| 878 | |
| 879 | # Publish the logical-to-physical mapping early so topology helpers |
| 880 | # work before init_device (needed by set_worker_net_device below). |
| 881 | assigned_physical_gpu_ids = kwargs[ |
| 882 | "vllm_config" |
| 883 | ].parallel_config.assigned_physical_gpu_ids |
| 884 | if assigned_physical_gpu_ids is not None: |
| 885 | from vllm.platforms.interface import set_assigned_physical_gpu_ids |
| 886 | |
| 887 | set_assigned_physical_gpu_ids(assigned_physical_gpu_ids) |
| 888 | |
| 889 | # Set net device env vars for the worker if VLLM_GPU_NIC_PCIE_MAPPING is set |
| 890 | set_worker_net_device(kwargs.get("local_rank", 0), kwargs["vllm_config"]) |
| 891 | |
| 892 | worker = None |
| 893 | ready_writer = kwargs.pop("ready_pipe") |
| 894 | death_pipe = kwargs.pop("death_pipe", None) |
| 895 | |
| 896 | # Close inherited pipes from parent (incl. other worker pipes) |
| 897 | # Explicitly passing in existing pipes and closing them makes the pipe |
| 898 | # behave when using fork. Otherwise, a hidden reference to the pipes |
| 899 | # exist in the child process and prevents EOF closure. |
| 900 | for fd in kwargs.pop("inherited_fds", []): |
| 901 | try: |
| 902 | os.close(fd) |
| 903 | except Exception as e: |
| 904 | logger.warning("Error closing inherited connection: %s: %s", type(e), e) |
| 905 | |
| 906 | try: |
| 907 | # Initialize tracer |
| 908 | rank = kwargs.get("rank", 0) |
| 909 | maybe_init_worker_tracer( |
| 910 | instrumenting_module_name="vllm.worker", |
| 911 | process_kind="worker", |
| 912 | process_name=f"Worker_{rank}", |
| 913 | ) |
| 914 | |
| 915 | worker = WorkerProc(*args, **kwargs) |
| 916 | assert worker.worker_response_mq is not None |
| 917 | if kwargs["vllm_config"].parallel_config.numa_bind: |
| 918 | numa_utils.log_current_affinity_state(f"Worker_{worker.rank}") |
| 919 | |
| 920 | worker.monitor_death_pipe(death_pipe, shutdown_requested) |
| 921 | |
| 922 | # Send READY once we know everything is loaded |
| 923 | ready_writer.send( |
| 924 | { |
| 925 | "status": WorkerProc.READY_STR, |
| 926 | "handle": worker.worker_response_mq.export_handle(), |
| 927 | "peer_response_handles": worker.peer_response_handles, |
| 928 | } |
| 929 | ) |
| 930 | |
| 931 | # Ensure message queues are ready. Will deadlock if re-ordered. |
| 932 | # Must be kept consistent with the Executor |
| 933 | if worker.rpc_broadcast_mq is not None: |
| 934 | worker.rpc_broadcast_mq.wait_until_ready() |
| 935 | worker.worker_response_mq.wait_until_ready() |
| 936 | ready_writer.close() |
| 937 | ready_writer = None |
| 938 | |
| 939 | worker.worker_busy_loop() |
| 940 | |
| 941 | except Exception: |
| 942 | # NOTE: if an Exception arises in busy_loop, we send |
| 943 | # a FAILURE message over the MQ RPC to notify the Executor, |
| 944 | # which triggers system shutdown. |
| 945 | # TODO(rob): handle case where the MQ itself breaks. |
| 946 | |
| 947 | if ready_writer is not None: |
| 948 | logger.exception("WorkerProc failed to start.") |
| 949 | elif shutdown_requested.is_set(): |
| 950 | logger.debug_once( |
| 951 | "[shutdown] WorkerProc: exiting after shutdown request" |
| 952 | ) |
| 953 | else: |
| 954 | logger.exception("WorkerProc failed.") |
| 955 | |
| 956 | # The parent sends a SIGTERM to all worker processes if |
| 957 | # any worker dies. Set this value so we don't re-throw |
| 958 | # SystemExit() to avoid zmq exceptions in __del__. |
| 959 | shutdown_requested.set() |
| 960 | |
| 961 | except SystemExit as e: |
| 962 | # SystemExit is raised on SIGTERM or SIGKILL, which usually indicates that |
| 963 | # the graceful shutdown process did not succeed |
| 964 | if shutdown_requested.is_set(): |
| 965 | logger.debug_once( |
| 966 | "[shutdown] WorkerProc: terminated by shutdown signal" |
| 967 | ) |
| 968 | else: |
| 969 | logger.warning("WorkerProc was terminated") |
| 970 | # SystemExit must never be ignored |
| 971 | raise e |
| 972 | |
| 973 | finally: |
| 974 | if ready_writer is not None: |
| 975 | ready_writer.close() |
| 976 | if death_pipe is not None: |
| 977 | death_pipe.close() |
| 978 | # Clean up once worker exits busy loop |
| 979 | if worker is not None: |
| 980 | worker.shutdown() |
| 981 | |
| 982 | class ResponseStatus(Enum): |
| 983 | SUCCESS = auto() |
| 984 | FAILURE = auto() |
| 985 | |
| 986 | def enqueue_output(self, output: Any): |
| 987 | """Prepares output from the worker and enqueues it to the |
| 988 | worker_response_mq. If the output is an Exception, it is |
| 989 | converted to a FAILURE response. |
| 990 | """ |
| 991 | if isinstance(output, AsyncModelRunnerOutput): |
| 992 | try: |
| 993 | output = output.get_output() |
| 994 | except Exception as e: |
| 995 | logger.exception("Error getting async model runner output") |
| 996 | output = e |
| 997 | |
| 998 | if isinstance(output, Exception): |
| 999 | result = (WorkerProc.ResponseStatus.FAILURE, str(output)) |
| 1000 | else: |
| 1001 | result = (WorkerProc.ResponseStatus.SUCCESS, output) |
| 1002 | if (response_mq := self.worker_response_mq) is not None: |
| 1003 | response_mq.enqueue(result) |
| 1004 | |
| 1005 | def handle_output(self, output: Any): |
| 1006 | """Handles output from the worker. If async scheduling is enabled, |
| 1007 | it is passed to the async_output_busy_loop thread. Otherwise, it is |
| 1008 | enqueued directly to the worker_response_mq. |
| 1009 | """ |
| 1010 | if self.use_async_scheduling: |
| 1011 | self.async_output_queue.put(output) |
| 1012 | else: |
| 1013 | self.enqueue_output(output) |
| 1014 | |
| 1015 | def async_output_busy_loop(self): |
| 1016 | """Entrypoint for the thread which handles outputs asynchronously.""" |
| 1017 | # set device to the worker device for the thread. |
| 1018 | # a thread will not inherit the context of the main thread. |
| 1019 | # when calling any cuda runtime functions, it will implicitly |
| 1020 | # create a new cuda context on device 0, consuming extra memory. |
| 1021 | # here we set the device to the worker device for the thread, |
| 1022 | # enforcing the context to be the same as the main thread. |
| 1023 | from vllm.platforms import current_platform |
| 1024 | |
| 1025 | if hasattr(self.worker, "device"): |
| 1026 | current_platform.set_device(self.worker.device) |
| 1027 | |
| 1028 | while True: |
| 1029 | output = self.async_output_queue.get() |
| 1030 | self.enqueue_output(output) |
| 1031 | |
| 1032 | def worker_busy_loop(self): |
| 1033 | """Main busy loop for Multiprocessing Workers.""" |
| 1034 | assert self.rpc_broadcast_mq is not None |
| 1035 | while True: |
| 1036 | self._execute_worker_rpc(self.rpc_broadcast_mq.dequeue(indefinite=True)) |
| 1037 | |
| 1038 | def _execute_worker_rpc( |
| 1039 | self, |
| 1040 | rpc_request: tuple[str | bytes, tuple[Any, ...], dict[str, Any], int | None], |
| 1041 | ) -> None: |
| 1042 | """Execute one RPC in a separate frame from the dequeue loop.""" |
| 1043 | method, args, kwargs, output_rank = rpc_request |
| 1044 | try: |
| 1045 | if isinstance(method, str): |
| 1046 | func = getattr(self.worker, method) |
| 1047 | elif isinstance(method, bytes): |
| 1048 | func = partial(cloudpickle.loads(method), self.worker) |
| 1049 | |
| 1050 | output = func(*args, **kwargs) |
| 1051 | |
| 1052 | if output_rank is None or self.rank == output_rank: |
| 1053 | self.handle_output(output) |
| 1054 | except Exception as e: |
| 1055 | # Notes have been introduced in python 3.11 |
| 1056 | if hasattr(e, "add_note"): |
| 1057 | e.add_note(traceback.format_exc()) |
| 1058 | logger.exception("WorkerProc hit an exception.") |
| 1059 | # enqueue_output converts the exception to a FAILURE response |
| 1060 | # containing its string representation before transport. |
| 1061 | if output_rank is None or self.rank == output_rank: |
| 1062 | self.handle_output(e) |
| 1063 | |
| 1064 | @staticmethod |
| 1065 | def setup_proc_title_and_log_prefix(enable_ep: bool) -> None: |
| 1066 | # Check if parallel groups are initialized first |
| 1067 | if not model_parallel_is_initialized(): |
| 1068 | # Parallel groups not yet initialized, use default process name |
| 1069 | set_process_title(name="Worker") |
| 1070 | decorate_logs("Worker") |
| 1071 | return |
| 1072 | |
| 1073 | dp_size = get_dp_group().world_size |
| 1074 | dp_rank = get_dp_group().rank_in_group |
| 1075 | pp_size = get_pp_group().world_size |
| 1076 | pp_rank = get_pp_group().rank_in_group |
| 1077 | pcp_size = get_pcp_group().world_size |
| 1078 | pcp_rank = get_pcp_group().rank_in_group |
| 1079 | tp_size = get_tp_group().world_size |
| 1080 | tp_rank = get_tp_group().rank_in_group |
| 1081 | dcp_size = get_dcp_group().world_size |
| 1082 | dcp_rank = get_dcp_group().rank_in_group |
| 1083 | process_name = "Worker" |
| 1084 | if dp_size > 1: |
| 1085 | process_name += f"_DP{dp_rank}" |
| 1086 | if pp_size > 1: |
| 1087 | process_name += f"_PP{pp_rank}" |
| 1088 | if pcp_size > 1: |
| 1089 | process_name += f"_PCP{pcp_rank}" |
| 1090 | if tp_size > 1: |
| 1091 | process_name += f"_TP{tp_rank}" |
| 1092 | if dcp_size > 1: |
| 1093 | process_name += f"_DCP{dcp_rank}" |
| 1094 | if enable_ep: |
| 1095 | ep_rank = get_ep_group().rank_in_group |
| 1096 | process_name += f"_EP{ep_rank}" |
| 1097 | set_process_title(name=process_name) |
| 1098 | decorate_logs(process_name) |
| 1099 | |
| 1100 | |
| 1101 | def set_multiprocessing_worker_envs(local_world_size: int = 1): |
| 1102 | """Set up environment variables that should be used when there are workers |
| 1103 | in a multiprocessing environment. This should be called by the parent |
| 1104 | process before worker processes are created""" |
| 1105 | _maybe_force_spawn() |
| 1106 | |
| 1107 | if current_platform.is_cpu() or "OMP_NUM_THREADS" in os.environ: |
| 1108 | return |
| 1109 | |
| 1110 | # Choose the workers' thread count here, before they start, since a worker |
| 1111 | # must not set its own: `torch.set_num_threads()` spawns the thread pool |
| 1112 | # eagerly, and doing that part way through a worker's startup either races |
| 1113 | # the dlopen of shared objects or, in a forked worker, deadlocks (libgomp |
| 1114 | # is not fork-safe). |
| 1115 | num_threads = startup_omp_num_threads(local_world_size) |
| 1116 | os.environ["OMP_NUM_THREADS"] = str(num_threads) |
| 1117 | os.environ[OMP_NUM_THREADS_SET_BY_VLLM] = "1" |
| 1118 | |
| 1119 | # A spawned worker picks the count up from the environment when it imports |
| 1120 | # torch. A forked worker instead inherits it from this process, so set it |
| 1121 | # here too. This is safe as long as we don't *use* the pool before forking: |
| 1122 | # a forked child whose parent had run a parallel region deadlocks, whereas |
| 1123 | # one whose parent merely sized the pool does not. |
| 1124 | torch.set_num_threads(num_threads) |
| 1125 | logger.debug( |
| 1126 | "Set OMP_NUM_THREADS=%d for %d worker process(es).", |
| 1127 | num_threads, |
| 1128 | local_world_size, |
| 1129 | ) |