utils.py 41.3 KB
Newer Older
1
2
3
4
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

import contextlib
5
import os
6
import weakref
7
from collections.abc import Callable, Iterator
8
9
10
11
from dataclasses import dataclass
from enum import Enum, auto
from multiprocessing import Process, connection
from multiprocessing.process import BaseProcess
12
from typing import TYPE_CHECKING
13
from unittest.mock import patch
14
15
16
17

import msgspec
import zmq

18
from vllm import envs
19
20
from vllm.config import CacheConfig, ParallelConfig, VllmConfig
from vllm.logger import init_logger
21
from vllm.platforms import current_platform
22
from vllm.ray.ray_env import get_env_vars_to_copy
23
from vllm.utils.network_utils import get_open_zmq_ipc_path, zmq_socket_ctx
24
from vllm.utils.system_utils import get_mp_context
25
from vllm.v1.engine.coordinator import DPCoordinator
26
from vllm.v1.executor import Executor
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
from vllm.v1.utils import get_engine_client_zmq_addr, shutdown

if TYPE_CHECKING:
    from ray.util.placement_group import PlacementGroup

logger = init_logger(__name__)

STARTUP_POLL_PERIOD_MS = 10000


class CoreEngineState(Enum):
    NEW = auto()
    CONNECTED = auto()
    READY = auto()


class CoreEngine:
    """One per data parallel rank, used to track state during handshaking."""

    def __init__(self, index: int = 0, local: bool = True):
        self.local = local
        self.identity = index.to_bytes(2, "little")

        self.state = CoreEngineState.NEW


@dataclass
class EngineZmqAddresses:
    # ZMQ input socket addresses for each front-end client (requests)
    inputs: list[str]
    # ZMQ output socket addresses for each front-end client (responses)
    outputs: list[str]
    # ZMQ input socket address of DP coordinator if applicable
60
    coordinator_input: str | None = None
61
    # ZMQ output socket address of DP coordinator if applicable
62
    coordinator_output: str | None = None
63
64
65
    # ZMQ socket for front-end to connect to DP coordinator.
    # Not used by engine, just relayed to front-end in handshake response.
    # Only required for external DP LB case.
66
    frontend_stats_publish_address: str | None = None
67
68
69
70
71
72
73
74


@dataclass
class EngineHandshakeMetadata:
    """Metadata sent to each engine process during startup handshake,
    including addresses of the front-end ZMQ queues that they should
    connect to.
    """
75

76
    addresses: EngineZmqAddresses
77
    parallel_config: dict[str, int | str | list[int]]
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96


class CoreEngineProcManager:
    """
    Utility class to handle creation, readiness, and shutdown
    of background processes used by the AsyncLLM and LLMEngine.
    """

    def __init__(
        self,
        target_fn: Callable,
        local_engine_count: int,
        start_index: int,
        local_start_index: int,
        vllm_config: VllmConfig,
        local_client: bool,
        handshake_address: str,
        executor_class: type[Executor],
        log_stats: bool,
97
        client_handshake_address: str | None = None,
98
99
100
101
102
103
104
105
106
107
108
    ):
        context = get_mp_context()
        common_kwargs = {
            "vllm_config": vllm_config,
            "local_client": local_client,
            "handshake_address": handshake_address,
            "executor_class": executor_class,
            "log_stats": log_stats,
        }

        if client_handshake_address:
109
            common_kwargs["client_handshake_address"] = client_handshake_address
110
111

        self.processes: list[BaseProcess] = []
112
        local_dp_ranks = []
113
114
115
        for index in range(local_engine_count):
            local_index = local_start_index + index
            global_index = start_index + index
116

117
            # Start EngineCore in background process.
118
            local_dp_ranks.append(local_index)
119
            self.processes.append(
120
121
122
123
124
125
126
127
128
129
                context.Process(
                    target=target_fn,
                    name=f"EngineCore_DP{global_index}",
                    kwargs=common_kwargs
                    | {
                        "dp_rank": global_index,
                        "local_dp_rank": local_index,
                    },
                )
            )
130
131

        self._finalizer = weakref.finalize(self, shutdown, self.processes)
132
133

        data_parallel = vllm_config.parallel_config.data_parallel_size > 1
134
        try:
135
            for proc, local_dp_rank in zip(self.processes, local_dp_ranks):
136
137
138
                # Adjust device control in DP for non-CUDA platforms
                # as well as external and ray launchers
                # For CUDA platforms, we use torch.cuda.set_device()
139
140
                with (
                    set_device_control_env_var(vllm_config, local_dp_rank)
141
142
143
144
145
146
147
                    if (
                        data_parallel
                        and (
                            not current_platform.is_cuda_alike()
                            or vllm_config.parallel_config.use_ray
                        )
                    )
148
149
                    else contextlib.nullcontext()
                ):
150
                    proc.start()
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
        finally:
            # Kill other procs if not all are running.
            if self.finished_procs():
                self.close()

    def close(self):
        """Shutdown all procs."""
        self._finalizer()

    def join_first(self):
        """Wait for any process to exit."""
        connection.wait(proc.sentinel for proc in self.processes)

    def sentinels(self) -> list:
        return [proc.sentinel for proc in self.processes]

    def finished_procs(self) -> dict[str, int]:
        """Returns dict of proc name -> exit code for any finished procs."""
        return {
            proc.name: proc.exitcode
171
172
            for proc in self.processes
            if proc.exitcode is not None
173
174
175
        }


176
@contextlib.contextmanager
177
178
179
def set_device_control_env_var(
    vllm_config: VllmConfig, local_dp_rank: int
) -> Iterator[None]:
180
181
182
183
184
    """
    Temporarily set CUDA_VISIBLE_DEVICES or equivalent
    for engine subprocess.
    """
    world_size = vllm_config.parallel_config.world_size
185
    local_world_size = vllm_config.parallel_config.local_world_size
186
    evar = current_platform.device_control_env_var
187

188
    value = get_device_indices(evar, local_dp_rank, world_size, local_world_size)
189
    with patch.dict(os.environ, values=((evar, value),)):
190
191
192
        yield


193
def get_device_indices(
194
195
196
197
    device_control_env_var: str,
    local_dp_rank: int,
    world_size: int,
    local_world_size: int | None = None,
198
):
199
200
201
202
203
204
205
    """
    Returns a comma-separated string of device indices for the specified
    data parallel rank.

    For example, if world_size=2 and local_dp_rank=1, and there are 4 devices,
    this will select devices 2 and 3 for local_dp_rank=1.
    """
206
207
    if local_world_size is None:
        local_world_size = world_size
208
209
210
    try:
        value = ",".join(
            str(current_platform.device_id_to_physical_device_id(i))
211
212
213
214
            for i in range(
                local_dp_rank * world_size,
                local_dp_rank * world_size + local_world_size,
            )
215
        )
216
    except IndexError as e:
217
218
219
220
221
222
223
        raise Exception(
            f"Error setting {device_control_env_var}: "
            f"local range: [{local_dp_rank * world_size}, "
            f"{(local_dp_rank + 1) * world_size}) "
            "base value: "
            f'"{os.getenv(device_control_env_var)}"'
        ) from e
224
    return value
225
226


227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
class CoreEngineActorManager:
    """
    Utility class to handle creation, readiness, and shutdown
    of core engine Ray actors used by the AsyncLLM and LLMEngine.

    Different from CoreEngineProcManager, this class manages
    core engines for both local and remote nodes.
    """

    def __init__(
        self,
        vllm_config: VllmConfig,
        addresses: EngineZmqAddresses,
        executor_class: type[Executor],
        log_stats: bool,
242
243
        placement_groups: list["PlacementGroup"] | None = None,
        local_dp_ranks: list[int] | None = None,
244
245
246
247
    ):
        import copy

        import ray
248
        from ray.runtime_env import RuntimeEnv
249
        from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy
250

251
252
253
254
255
256
257
258
        from vllm.v1.engine.core import DPMoEEngineCoreActor, EngineCoreActor

        dp_size = vllm_config.parallel_config.data_parallel_size
        actor_class = (
            DPMoEEngineCoreActor
            if dp_size > 1 and vllm_config.model_config.is_moe
            else EngineCoreActor
        )
259
260
261

        self.local_engine_actors: list[ray.ActorHandle] = []
        self.remote_engine_actors: list[ray.ActorHandle] = []
262

263
        env_vars_list = get_env_vars_to_copy(destination=actor_class.__name__)
264
        self.env_vars_dict = {
265
            name: os.environ[name] for name in env_vars_list if name in os.environ
266
267
268
269
270
271
        }
        runtime_env = RuntimeEnv(env_vars=self.env_vars_dict)

        self.addresses = addresses
        self.executor_class = executor_class
        self.log_stats = log_stats
272
        local_engine_count = vllm_config.parallel_config.data_parallel_size_local
273
274
275
        world_size = vllm_config.parallel_config.world_size

        if ray.is_initialized():
276
            logger.info("Ray is already initialized. Skipping Ray initialization.")
277
278
279
280
281
        else:
            ray.init()

        if placement_groups is not None:
            assert local_dp_ranks is not None, (
282
283
                "local_dp_ranks must be provided if placement_groups is provided"
            )
284
            assert len(placement_groups) == len(local_dp_ranks), (
285
286
                "placement_groups and local_dp_ranks must have the same length"
            )
287
288
289
290
            logger.info("Using provided placement groups")
            # TODO(rui): validate passed-in placement groups
            self.created_placement_groups = []
        else:
291
            placement_groups, local_dp_ranks = (
292
                CoreEngineActorManager.create_dp_placement_groups(vllm_config)
293
            )
294
295
            self.created_placement_groups = placement_groups
        assert len(placement_groups) == dp_size, (
296
297
            "Number of placement groups must match data parallel size"
        )
298

299
        self.placement_group_is_local = []
300
        refs = []
301
302
303
        for index, local_index, pg in zip(
            range(dp_size), local_dp_ranks, placement_groups
        ):
304
305
306
            dp_vllm_config = copy.deepcopy(vllm_config)
            dp_vllm_config.parallel_config.placement_group = pg
            local_client = index < local_engine_count
307

308
309
310
311
312
313
314
            if dp_size > 1 and dp_vllm_config.kv_transfer_config is not None:
                # modify the engine_id and append the local_dp_rank to it to ensure
                # that the kv_transfer_config is unique for each DP rank.
                dp_vllm_config.kv_transfer_config.engine_id = (
                    f"{dp_vllm_config.kv_transfer_config.engine_id}_dp{local_index}"
                )

315
316
317
318
319
320
            # Ray XPU known issue: dpctl initializes the GPU runtime early, so
            # setting device env vars in Ray actor's initialization method
            # will not affect device selection. See:
            # https://github.com/ray-project/ray/blob/master/python/ray/_private/accelerators/intel_gpu.py#L56 # noqa: E501
            if current_platform.is_xpu():
                device_evar = current_platform.device_control_env_var
321
322
323
                device_indices = get_device_indices(
                    device_evar, local_index, world_size
                )
324
325
326
327
                actor_env_vars = self.env_vars_dict.copy()
                actor_env_vars[device_evar] = device_indices
                runtime_env = RuntimeEnv(env_vars=actor_env_vars)

328
            actor = (
329
                ray.remote(actor_class)
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
                .options(
                    scheduling_strategy=PlacementGroupSchedulingStrategy(
                        placement_group=pg,
                        placement_group_bundle_index=world_size,
                    ),
                    runtime_env=runtime_env,
                )
                .remote(
                    vllm_config=dp_vllm_config,
                    executor_class=executor_class,
                    log_stats=log_stats,
                    local_client=local_client,
                    addresses=addresses,
                    dp_rank=index,
                    local_dp_rank=local_index,
                )
            )
347
348
349
350
            if local_client:
                self.local_engine_actors.append(actor)
            else:
                self.remote_engine_actors.append(actor)
351
            self.placement_group_is_local.append(local_client)
352
353
354
355
356
357
358
359
360
            refs.append(actor.wait_for_init.remote())

        ray.get(refs)
        self.run_refs = []
        for actor in self.local_engine_actors + self.remote_engine_actors:
            self.run_refs.append(actor.run.remote())

    @staticmethod
    def create_dp_placement_groups(
361
        vllm_config: VllmConfig,
362
    ) -> tuple[list["PlacementGroup"], list[int]]:
363
364
365
        """
        Create placement groups for data parallel.
        """
366
367
368
369
370

        import ray
        from ray._private.state import available_resources_per_node

        logger.info("Creating placement groups for data parallel")
371
        dp_master_ip = vllm_config.parallel_config.data_parallel_master_ip
372
373
        dp_size = vllm_config.parallel_config.data_parallel_size
        dp_size_local = vllm_config.parallel_config.data_parallel_size_local
374
375
376
377
378

        available_resources = available_resources_per_node()
        world_size = vllm_config.parallel_config.world_size
        placement_groups: list[PlacementGroup] = []
        local_dp_ranks: list[int] = []
379

380
381
382
383
384
        dp_master_ip_key = f"node:{dp_master_ip}"
        nodes = sorted(
            available_resources.values(), key=lambda x: dp_master_ip_key not in x
        )
        assert len(nodes) > 0, "No nodes with resources found in Ray cluster."
385
        assert dp_master_ip_key in nodes[0], (
386
            f"The DP master node (ip: {dp_master_ip}) is missing or dead"
387
        )
388
        device_str = current_platform.ray_device_key
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
        n_node_devices: list[int] = [
            int(node_resources[device_str])
            for node_resources in nodes
            if device_str in node_resources
        ]
        assert n_node_devices, f"No {device_str} found in Ray cluster."
        max_device_per_node = max(n_node_devices)

        pack_strategy = envs.VLLM_RAY_DP_PACK_STRATEGY
        _supported_pack_strategies = ("strict", "fill", "span")
        if pack_strategy not in _supported_pack_strategies:
            raise ValueError(
                f"{envs.VLLM_RAY_DP_PACK_STRATEGY} is not supported. "
                "Make sure to set `VLLM_RAY_DP_PACK_STRATEGY` "
                f"to one of {_supported_pack_strategies}"
            )
405

406
        all2all_backend = vllm_config.parallel_config.all2all_backend
407
        if pack_strategy == "fill" and (
408
409
            all2all_backend == "deepep_high_throughput"
            or all2all_backend == "deepep_low_latency"
410
411
412
413
414
415
416
417
        ):
            raise ValueError(
                "DeepEP kernels require EP ranks [0,7] (same for [8,15], ...) "
                "to be on the same node, but VLLM_RAY_DP_PACK_STRATEGY=fill "
                "does not guarantee that. "
                "Please use VLLM_RAY_DP_PACK_STRATEGY=strict instead."
            )

418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
        if pack_strategy in ("strict", "fill"):
            placement_strategy = "STRICT_PACK"
        else:
            placement_strategy = "PACK"
            assert world_size > max_device_per_node, (
                f"World size {world_size} is smaller than the "
                "maximum number of devices per node "
                f"{max_device_per_node}. Make sure to set "
                "`VLLM_RAY_DP_PACK_STRATEGY` to `strict` or `fill`"
            )

            # if we need multiple nodes per dp group, we require for now that
            # available nodes are homogenous
            assert set(n_node_devices) == {max_device_per_node}, (
                f"Nodes are not homogenous, {nodes}"
            )
            assert world_size % max_device_per_node == 0, (
                f"For multi-node data parallel groups, world_size ({world_size}) must "
                f"be a multiple of number of devices per node ({max_device_per_node})."
            )
            assert len(n_node_devices) * max_device_per_node >= world_size * dp_size, (
                f"Not enough total available nodes ({len(n_node_devices)}) "
                f"and devices per node ({max_device_per_node}) "
                f"to satisfy required world size {world_size} and data parallel size "
                f"{dp_size}"
            )
            assert dp_size_local == 1, (
                f"data-parallel-size-local {dp_size_local} should be set as the "
                "default (1) for VLLM_RAY_DP_PACK_STRATEGY=span. "
                "The actual data-parallel-size-local will be auto determined."
            )

        # bundles collected for a single DP rank from multiple nodes,
        # for "span" pack strategy
        collected_bundles = []
453
        for node_resources in nodes:
454
455
456
457
458
459
            node_ip_keys = [
                key
                for key in node_resources
                if key != "node:__internal_head__" and key.startswith("node:")
            ]
            assert len(node_ip_keys) == 1, (
460
                f"Zero or multiple node IP keys found in node resources: {node_ip_keys}"
461
462
463
464
            )
            node_ip_key = node_ip_keys[0]
            node_ip = node_ip_key.split(":")[1]

465
466
467
468
469
470
471
472
            n_device_on_node = int(node_resources.get(device_str, 0))
            if pack_strategy == "span" and n_device_on_node != 0:
                # Strictly speaking,
                # dp_size_available = n_device_on_node / world_size
                # and is a fraction, but we use 1 for easier processing
                dp_size_available = 1
            else:
                dp_size_available = n_device_on_node // world_size
473
474
475
476

            if node_ip == dp_master_ip:
                if dp_size_available < dp_size_local:
                    raise ValueError(
477
478
479
                        f"Not enough resources to allocate {dp_size_local} DP ranks "
                        f"on DP master node {dp_master_ip}, possible to fit "
                        f"{dp_size_available} DP ranks."
480
                    )
481
                dp_size_to_allocate = dp_size_local
482
            elif pack_strategy == "strict":
483
484
485
486
487
488
489
                if dp_size_available < dp_size_local:
                    logger.info(
                        "Skipping node %s as %s DP ranks could not fit, "
                        "possible to fit %s DP ranks",
                        node_ip,
                        dp_size_local,
                        dp_size_available,
490
                    )
491
492
493
                    continue
                dp_size_to_allocate = dp_size_local
            else:
494
495
                # for "pack_strategy" in "fill" and "span"
                # we always take everything that's available
496
497
498
                dp_size_to_allocate = dp_size_available

            for i in range(dp_size_to_allocate):
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
                device_bundle = [{device_str: 1.0, "node:" + node_ip: 0.001}]
                if pack_strategy == "span":
                    collected_bundles += device_bundle * n_device_on_node
                    assert len(collected_bundles) <= world_size, (
                        "collected_bundles should be <= world_size, "
                        f"but got {len(collected_bundles)=} and {world_size=}"
                    )

                    # we only create a placement group if we collected enough devices
                    if len(collected_bundles) < world_size:
                        continue

                    bundles = collected_bundles + [{"CPU": 1.0}]
                    collected_bundles = []
                else:
                    bundles = device_bundle * world_size + [{"CPU": 1.0}]

516
517
                pg = ray.util.placement_group(
                    name=f"dp_rank_{len(placement_groups)}",
518
                    strategy=placement_strategy,
519
520
521
522
                    bundles=bundles,
                )
                placement_groups.append(pg)
                local_dp_ranks.append(i)
523
524
                if len(placement_groups) == dp_size:
                    break
525
526

        if len(placement_groups) < dp_size:
527
            raise ValueError(
528
                f"Not enough resources to allocate {dp_size} "
529
530
531
                "placement groups, only created "
                f"{len(placement_groups)} placement groups. "
                "Available resources: "
532
533
                f"{available_resources}"
            )
534
535
536
537
538
539
540
        assert len(placement_groups) == dp_size, (
            f"Created {len(placement_groups)} DP placement groups, expected {dp_size}"
        )
        assert len(local_dp_ranks) == dp_size, (
            f"local_dp_ranks length {len(local_dp_ranks)} does not match "
            f"expected {dp_size}"
        )
541
542
        return placement_groups, local_dp_ranks

543
544
545
546
547
548
549
550
    @staticmethod
    def add_dp_placement_groups(
        old_vllm_config: VllmConfig, new_data_parallel_size: int
    ) -> tuple[list["PlacementGroup"], list[int]]:
        """
        Add placement groups for new data parallel size.
        """
        import ray
551
552
553
554
        from ray._private.state import (
            available_resources_per_node,
            total_resources_per_node,
        )
555
556
557
558
559
560
561
562
563
564
565
566
567
        from ray.util.state import list_nodes

        old_dp_size = old_vllm_config.parallel_config.data_parallel_size
        num_pg_to_create = new_data_parallel_size - old_dp_size

        if num_pg_to_create <= 0:
            return [], []

        dp_master_ip = old_vllm_config.parallel_config.data_parallel_master_ip
        world_size = old_vllm_config.parallel_config.world_size

        nodes = list_nodes()
        nodes = sorted(nodes, key=lambda node: node.node_ip != dp_master_ip)
568
        assert nodes[0].node_ip == dp_master_ip, "The first node must be the head node"
569
        assert len(nodes) == 1 or nodes[1].node_ip != dp_master_ip, (
570
571
            "There can only be one head node"
        )
572
573
574
575
576
577
578
579

        available_resources = available_resources_per_node()
        total_resources = total_resources_per_node()

        placement_groups = []
        local_dp_ranks = []
        num_pg_created = 0

580
        device_str = current_platform.ray_device_key
581
582
583
584
585
586
        for node in nodes:
            if num_pg_created >= num_pg_to_create:
                break

            node_ip = node.node_ip
            node_id = node.node_id
587
            available_gpus = int(available_resources[node_id][device_str])
588
589
590

            # Get total GPUs on this node from the node's resources
            # Ray stores node resources with node ID as key
591
            total_gpus = int(total_resources[node_id][device_str])
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608

            # Calculate used GPUs and used engines on this node
            used_gpus = max(0, total_gpus - available_gpus)
            used_engines_on_node = used_gpus // world_size

            # Calculate how many new engines this node can accommodate
            available_engine_count = available_gpus // world_size

            # Create placement groups for new engines on this node
            for i in range(available_engine_count):
                if num_pg_created >= num_pg_to_create:
                    break

                rank = old_dp_size + num_pg_created

                # Create bundles with node constraint for master node
                if node_ip == dp_master_ip:
609
610
611
                    bundles = [
                        {device_str: 1.0, "node:" + dp_master_ip: 0.001}
                    ] * world_size + [{"CPU": 1.0}]
612
                else:
613
                    bundles = [{device_str: 1.0}] * world_size + [{"CPU": 1.0}]
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629

                pg = ray.util.placement_group(
                    name=f"dp_rank_{rank}",
                    strategy="STRICT_PACK",
                    bundles=bundles,
                )
                placement_groups.append(pg)

                # Local rank starts from the number of engines already used
                # on this node
                local_rank = used_engines_on_node + i
                local_dp_ranks.append(local_rank)
                num_pg_created += 1

        return placement_groups, local_dp_ranks

630
631
632
    def scale_up_elastic_ep(
        self, cur_vllm_config: VllmConfig, new_data_parallel_size: int
    ) -> None:
633
634
635
636
        import copy

        import ray
        from ray.runtime_env import RuntimeEnv
637
        from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy
638

639
640
641
642
643
644
645
        from vllm.v1.engine.core import DPMoEEngineCoreActor, EngineCoreActor

        actor_class = (
            DPMoEEngineCoreActor
            if cur_vllm_config.model_config.is_moe
            else EngineCoreActor
        )
646

647
648
649
        cur_data_parallel_size = len(self.local_engine_actors) + len(
            self.remote_engine_actors
        )
650
651
652
653

        assert new_data_parallel_size > cur_data_parallel_size, (
            f"New data parallel size {new_data_parallel_size} must be greater "
            f"than current data parallel size {cur_data_parallel_size} "
654
655
            "for scale up"
        )
656

657
658
659
        placement_groups, local_dp_ranks = self.add_dp_placement_groups(
            cur_vllm_config, new_data_parallel_size
        )
660
661
662
663
664

        world_size = cur_vllm_config.parallel_config.world_size
        dp_master_ip = cur_vllm_config.parallel_config.data_parallel_master_ip
        new_local_engines = 0

665
666
667
668
        runtime_env = RuntimeEnv(
            env_vars=self.env_vars_dict | {"VLLM_ELASTIC_EP_SCALE_UP_LAUNCH": "1"}
        )
        for i, (pg, local_rank) in enumerate(zip(placement_groups, local_dp_ranks)):
669
670
            rank = cur_data_parallel_size + i
            dp_vllm_config = copy.deepcopy(cur_vllm_config)
671
            dp_vllm_config.parallel_config.data_parallel_size = new_data_parallel_size
672
673
674
675
            dp_vllm_config.parallel_config.placement_group = pg

            # Check if this placement group is on the head node
            local_client = any(
676
677
                bundle.get("node:" + dp_master_ip, 0) > 0 for bundle in pg.bundle_specs
            )
678
679
680
681
682

            if local_client:
                new_local_engines += 1
                # Update data_parallel_size_local
                dp_vllm_config.parallel_config.data_parallel_size_local = (
683
684
685
686
687
                    cur_vllm_config.parallel_config.data_parallel_size_local
                    + new_local_engines
                )

            actor = (
688
                ray.remote(actor_class)
689
690
691
692
693
694
695
696
                .options(
                    scheduling_strategy=PlacementGroupSchedulingStrategy(
                        placement_group=pg,
                        placement_group_bundle_index=world_size,
                    ),
                    runtime_env=runtime_env,
                )
                .remote(
697
698
699
700
701
702
                    vllm_config=dp_vllm_config,
                    executor_class=self.executor_class,
                    log_stats=self.log_stats,
                    local_client=local_client,
                    addresses=self.addresses,
                    dp_rank=rank,
703
704
705
                    local_dp_rank=local_rank,
                )
            )
706
707
708
709
710
711
712
713

            if local_client:
                self.local_engine_actors.append(actor)
            else:
                self.remote_engine_actors.append(actor)
            self.created_placement_groups.append(pg)
            self.placement_group_is_local.append(local_client)

714
715
716
717
718
719
720
721
722
723
724
725
726
        ray.get(
            [
                actor.wait_for_init.remote()
                for actor in (
                    self.local_engine_actors[-new_local_engines:]
                    if new_local_engines > 0
                    else []
                )
                + self.remote_engine_actors[
                    -(len(placement_groups) - new_local_engines) :
                ]
            ]
        )
727

728
729
730
731
732
        actors = (
            self.local_engine_actors[-new_local_engines:]
            if new_local_engines > 0
            else []
        ) + self.remote_engine_actors[-(len(placement_groups) - new_local_engines) :]
733
734
735
736

        for actor in actors:
            self.run_refs.append(actor.run.remote())

737
        cur_vllm_config.parallel_config.data_parallel_size = new_data_parallel_size
738
739
740
        # Update old_vllm_config with new data_parallel_size_local if any new
        # local engines were added
        if new_local_engines > 0:
741
            cur_vllm_config.parallel_config.data_parallel_size_local += (
742
                new_local_engines
743
            )
744

745
746
747
    def scale_down_elastic_ep(
        self, cur_data_parallel_size: int, new_data_parallel_size: int
    ) -> None:
748
        import ray
749

750
751
752
        assert cur_data_parallel_size > new_data_parallel_size, (
            f"cur_data_parallel_size {cur_data_parallel_size} must be greater "
            f"than new_data_parallel_size {new_data_parallel_size} "
753
754
            "for scale down"
        )
755
756
757
758
759
760
761
762
763
        for _ in range(cur_data_parallel_size - new_data_parallel_size):
            pg = self.created_placement_groups.pop()
            is_local = self.placement_group_is_local.pop()
            if is_local:
                self.local_engine_actors.pop()
            else:
                self.remote_engine_actors.pop()
            ray.util.remove_placement_group(pg)

764
765
766
767
768
    def get_run_refs(self):
        return self.run_refs

    def close(self):
        import ray
769

770
771
772
773
774
775
776
777
778
779
780
781
        for actor in self.local_engine_actors + self.remote_engine_actors:
            ray.kill(actor)
        for pg in self.created_placement_groups:
            ray.util.remove_placement_group(pg)


@contextlib.contextmanager
def launch_core_engines(
    vllm_config: VllmConfig,
    executor_class: type[Executor],
    log_stats: bool,
    num_api_servers: int = 1,
782
783
) -> Iterator[
    tuple[
784
785
        CoreEngineProcManager | CoreEngineActorManager | None,
        DPCoordinator | None,
786
        EngineZmqAddresses,
787
788
    ]
]:
789
790
791
792
793
794
795
796
    """Launch engine and DP coordinator processes as needed."""

    parallel_config = vllm_config.parallel_config
    dp_size = parallel_config.data_parallel_size
    local_engine_count = parallel_config.data_parallel_size_local
    local_start_index = parallel_config.data_parallel_rank_local
    dp_rank = parallel_config.data_parallel_rank
    host = parallel_config.data_parallel_master_ip
797
    local_engines_only = parallel_config.local_engines_only
798
799
800
801
802
803
804
805

    # In offline mode there is an LLM instance per DP rank and
    # one core engine per LLM, see
    # examples/offline_inference/data_parallel.py.
    offline_mode = local_start_index is not None

    # client_local_only = True for cases where this front-end
    # sends requests only to colocated engines.
806
807
808
    client_local_only = (
        offline_mode or local_engines_only or (local_engine_count == dp_size)
    )
809
810
811
812
813
814
815
816
817
818
819
820
821

    # Set up input and output addresses.
    addresses = EngineZmqAddresses(
        inputs=[
            get_engine_client_zmq_addr(client_local_only, host)
            for _ in range(num_api_servers)
        ],
        outputs=[
            get_engine_client_zmq_addr(client_local_only, host)
            for _ in range(num_api_servers)
        ],
    )

822
823
824
825
826
827
828
    # Run the DP Coordinator process with rank 0 when in online DP mode.
    # The coordinator is needed for:
    # 1. Internal/hybrid LB: collecting and publishing queue stats for load balancing
    # 2. MoE models: wave coordination in addition to stats
    run_coordinator = (
        vllm_config.needs_dp_coordinator and not offline_mode and dp_rank == 0
    )
829
830

    if run_coordinator:
831
832
833
834
        coordinator = DPCoordinator(
            parallel_config,
            enable_wave_coordination=vllm_config.model_config.is_moe,
        )
835
836

        addresses.coordinator_input, addresses.coordinator_output = (
837
838
            coordinator.get_engine_socket_addresses()
        )
839
        addresses.frontend_stats_publish_address = (
840
841
            coordinator.get_stats_publish_address()
        )
842

843
        logger.info("Started DP Coordinator process (PID: %d)", coordinator.proc.pid)
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
    else:
        coordinator = None

    if parallel_config.data_parallel_backend == "ray":
        logger.info("Starting ray-based data parallel backend")

        engine_actor_manager = CoreEngineActorManager(
            vllm_config=vllm_config,
            addresses=addresses,
            executor_class=executor_class,
            log_stats=log_stats,
        )

        yield engine_actor_manager, coordinator, addresses
        return

860
    if offline_mode:
861
862
        assert local_engine_count == 1
        engines_to_handshake = [CoreEngine(index=dp_rank, local=True)]
863
864
865
866
867
    elif dp_rank == 0:
        # Rank 0 holds Coordinator, so it handshakes with all Cores
        # in both external dplb and internal dplb mode.
        # Note this also covers the case where we have zero local engines
        # and rank 0 is headless.
868
        engines_to_handshake = [
869
            CoreEngine(index=i, local=(i < local_engine_count)) for i in range(dp_size)
870
        ]
871
872
873
874
    else:
        # Rank > 0 handshakes with just the local cores it is managing.
        assert local_engines_only, (
            "Attempting to launch core_engines from dp_rank > 0, but "
875
876
            "found internal DPLB, which is incompatible."
        )
877
878
879
880
        engines_to_handshake = [
            CoreEngine(index=i, local=True)
            for i in range(dp_rank, dp_rank + local_engine_count)
        ]
881
882
883
884
885
886
887
888

    # Whether the started engines will handshake only with co-located
    # front-end processes. In external_dp_lb mode, ranks > 0 handshake with
    # their co-located frontend and also the rank 0 front-end, and hence this
    # will be False.
    handshake_local_only = offline_mode or local_engine_count == dp_size

    handshake_address = get_engine_client_zmq_addr(
889
890
        handshake_local_only, host, parallel_config.data_parallel_rpc_port
    )
891

892
    if local_engines_only and dp_rank > 0:
893
894
895
896
897
898
899
        assert not handshake_local_only
        local_handshake_address = get_open_zmq_ipc_path()
        client_handshake_address = local_handshake_address
    else:
        local_handshake_address = handshake_address
        client_handshake_address = None

900
901
902
    with zmq_socket_ctx(
        local_handshake_address, zmq.ROUTER, bind=True
    ) as handshake_socket:
903
904
905
906
907
908
909
910
911
912
913
914
915
916
        from vllm.v1.engine.core import EngineCoreProc

        # Start local engines.
        if local_engine_count:
            local_engine_manager = CoreEngineProcManager(
                EngineCoreProc.run_engine_core,
                vllm_config=vllm_config,
                executor_class=executor_class,
                log_stats=log_stats,
                handshake_address=handshake_address,
                client_handshake_address=client_handshake_address,
                local_client=True,
                local_engine_count=local_engine_count,
                start_index=dp_rank,
917
918
                local_start_index=local_start_index or 0,
            )
919
920
921
922
923
924
925
926
927
928
929
        else:
            local_engine_manager = None

        yield local_engine_manager, coordinator, addresses

        # Now wait for engines to start.
        wait_for_engine_startup(
            handshake_socket,
            addresses,
            engines_to_handshake,
            parallel_config,
930
            dp_size > 1 and vllm_config.model_config.is_moe,
931
932
933
934
935
936
937
938
939
940
941
            vllm_config.cache_config,
            local_engine_manager,
            coordinator.proc if coordinator else None,
        )


def wait_for_engine_startup(
    handshake_socket: zmq.Socket,
    addresses: EngineZmqAddresses,
    core_engines: list[CoreEngine],
    parallel_config: ParallelConfig,
942
    coordinated_dp: bool,
943
    cache_config: CacheConfig,
944
945
    proc_manager: CoreEngineProcManager | None,
    coord_process: Process | None,
946
947
948
949
950
951
952
953
954
):
    # Wait for engine core process(es) to send ready messages.
    local_count = parallel_config.data_parallel_size_local
    remote_count = len(core_engines) - local_count
    # [local, remote] counts
    conn_pending, start_pending = [local_count, remote_count], [0, 0]
    poller = zmq.Poller()
    poller.register(handshake_socket, zmq.POLLIN)

955
956
    remote_should_be_headless = (
        not parallel_config.data_parallel_hybrid_lb
957
        and not parallel_config.data_parallel_external_lb
958
    )
959

960
961
962
963
964
965
966
967
968
969
    if proc_manager is not None:
        for sentinel in proc_manager.sentinels():
            poller.register(sentinel, zmq.POLLIN)
    if coord_process is not None:
        poller.register(coord_process.sentinel, zmq.POLLIN)
    while any(conn_pending) or any(start_pending):
        events = poller.poll(STARTUP_POLL_PERIOD_MS)
        if not events:
            if any(conn_pending):
                logger.debug(
970
971
972
                    "Waiting for %d local, %d remote core engine proc(s) to connect.",
                    *conn_pending,
                )
973
974
            if any(start_pending):
                logger.debug(
975
976
977
                    "Waiting for %d local, %d remote core engine proc(s) to start.",
                    *start_pending,
                )
978
979
980
981
982
983
            continue
        if len(events) > 1 or events[0][0] != handshake_socket:
            # One of the local core processes exited.
            finished = proc_manager.finished_procs() if proc_manager else {}
            if coord_process is not None and coord_process.exitcode is not None:
                finished[coord_process.name] = coord_process.exitcode
984
985
986
987
988
            raise RuntimeError(
                "Engine core initialization failed. "
                "See root cause above. "
                f"Failed core proc(s): {finished}"
            )
989
990
991
992

        # Receive HELLO and READY messages from the input socket.
        eng_identity, ready_msg_bytes = handshake_socket.recv_multipart()
        eng_index = int.from_bytes(eng_identity, "little")
993
        engine = next((e for e in core_engines if e.identity == eng_identity), None)
994
        if engine is None:
995
996
997
            raise RuntimeError(
                f"Message from engine with unexpected data parallel rank: {eng_index}"
            )
998
        msg = msgspec.msgpack.decode(ready_msg_bytes)
999
        status, local, headless = msg["status"], msg["local"], msg["headless"]
1000
        if local != engine.local:
1001
1002
1003
1004
1005
1006
            raise RuntimeError(
                f"{status} message from "
                f"{'local' if local else 'remote'} "
                f"engine {eng_index}, expected it to be "
                f"{'local' if engine.local else 'remote'}"
            )
1007

1008
1009
1010
        # Remote engines must be headless iff we aren't in hybrid dp lb mode.
        if not local and headless != remote_should_be_headless:
            if headless:
1011
1012
1013
1014
1015
                raise RuntimeError(
                    f"Remote engine {eng_index} must not use "
                    f"--headless in external or hybrid dp lb "
                    f"mode"
                )
1016
            else:
1017
1018
1019
1020
1021
                raise RuntimeError(
                    f"Remote engine {eng_index} must use "
                    f"--headless unless in external or hybrid "
                    f"dp lb mode"
                )
1022

1023
        if status == "HELLO" and engine.state == CoreEngineState.NEW:
1024
            # Send init message with DP config info.
1025
1026
1027
1028
            init_message = msgspec.msgpack.encode(
                EngineHandshakeMetadata(
                    addresses=addresses,
                    parallel_config={
1029
1030
1031
1032
1033
1034
1035
                        k: getattr(parallel_config, k)
                        for k in (
                            "data_parallel_master_ip",
                            "data_parallel_master_port",
                            "_data_parallel_master_port_list",
                            "data_parallel_size",
                        )
1036
1037
1038
                    }
                    if coordinated_dp
                    else {},
1039
1040
1041
                )
            )
            handshake_socket.send_multipart((eng_identity, init_message), copy=False)
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
            conn_pending[0 if local else 1] -= 1
            start_pending[0 if local else 1] += 1
            engine.state = CoreEngineState.CONNECTED
        elif status == "READY" and engine.state == CoreEngineState.CONNECTED:
            # Setup KV cache config with initialization state from
            # engine core process. Sum values from all engines in DP case.
            num_gpu_blocks = cache_config.num_gpu_blocks or 0
            num_gpu_blocks += msg["num_gpu_blocks"]
            cache_config.num_gpu_blocks = num_gpu_blocks

            # In external DP LB mode, the coordinator address that the
            # front-end procs connect to is obtained from rank 0 via
            # one of the engine handshakes, and passed to the local
            # front-end process in the response from the other.
            if addresses.frontend_stats_publish_address is None:
1057
                addresses.frontend_stats_publish_address = msg.get("dp_stats_address")
1058

1059
1060
            # Validate config hash consistency across DP workers for MoE models.
            if coordinated_dp:
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
                worker_config_hash = msg.get("parallel_config_hash")
                expected_hash = parallel_config.compute_hash()
                if worker_config_hash != expected_hash:
                    raise RuntimeError(
                        f"Configuration mismatch detected for engine "
                        f"{eng_index}. All DP workers must have identical "
                        f"configurations for parameters that affect collective "
                        f"communication (e.g., enable_eplb, "
                        f"eplb_config.log_balancedness). "
                        f"Worker hash: {worker_config_hash}, "
                        f"Expected hash: {expected_hash}. "
                        f"Please ensure all workers are started with the same "
                        f"command-line arguments."
                    )

1076
1077
1078
            start_pending[0 if local else 1] -= 1
            engine.state = CoreEngineState.READY
        else:
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
            raise RuntimeError(
                f"Unexpected {status} message for "
                f"{'local' if local else 'remote'} engine "
                f"{eng_index} in {engine.state} state."
            )

        logger.debug(
            "%s from %s core engine process %s.",
            status,
            "local" if local else "remote",
            eng_index,
        )