core.py 56.2 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
import gc
4
import os
5
import queue
6
import signal
7
8
import threading
import time
9
from collections import deque
10
from collections.abc import Callable, Generator
11
from concurrent.futures import Future
Rui Qiao's avatar
Rui Qiao committed
12
from contextlib import ExitStack, contextmanager
13
from inspect import isclass, signature
14
from logging import DEBUG
15
from typing import Any, TypeVar, cast
16

17
import msgspec
18
19
import zmq

20
21
from vllm.config import ParallelConfig, VllmConfig
from vllm.distributed import stateless_destroy_torch_distributed_process_group
22
from vllm.envs import enable_envs_cache
23
from vllm.logger import init_logger
24
from vllm.logging_utils.dump_input import dump_engine_exception
25
from vllm.lora.request import LoRARequest
26
from vllm.multimodal import MULTIMODAL_REGISTRY
27
from vllm.multimodal.cache import engine_receiver_cache_from_config
28
from vllm.tasks import POOLING_TASKS, SupportedTask
29
from vllm.transformers_utils.config import maybe_register_config_serialize_by_value
30
from vllm.utils.gc_utils import maybe_attach_gc_debug_callback
31
from vllm.utils.hashing import get_hash_fn_by_name
32
from vllm.utils.network_utils import make_zmq_socket
33
from vllm.utils.system_utils import decorate_logs, set_process_title
34
35
36
37
38
39
40
from vllm.v1.core.kv_cache_utils import (
    BlockHash,
    generate_scheduler_kv_cache_config,
    get_kv_cache_configs,
    get_request_block_hasher,
    init_none_hash,
)
41
from vllm.v1.core.sched.interface import SchedulerInterface
42
from vllm.v1.core.sched.output import SchedulerOutput
43
44
45
46
47
48
49
50
51
52
53
54
55
56
from vllm.v1.engine import (
    EngineCoreOutputs,
    EngineCoreRequest,
    EngineCoreRequestType,
    ReconfigureDistributedRequest,
    ReconfigureRankType,
    UtilityOutput,
    UtilityResult,
)
from vllm.v1.engine.utils import (
    EngineHandshakeMetadata,
    EngineZmqAddresses,
    get_device_indices,
)
57
from vllm.v1.executor import Executor
58
from vllm.v1.kv_cache_interface import KVCacheConfig
59
from vllm.v1.metrics.stats import SchedulerStats
60
from vllm.v1.outputs import ModelRunnerOutput
61
from vllm.v1.request import Request, RequestStatus
62
from vllm.v1.serial_utils import MsgpackDecoder, MsgpackEncoder
63
from vllm.v1.structured_output import StructuredOutputManager
64
from vllm.v1.utils import record_function_or_nullcontext
65
66
67
68
from vllm.version import __version__ as VLLM_VERSION

logger = init_logger(__name__)

69
POLLING_TIMEOUT_S = 2.5
70
HANDSHAKE_TIMEOUT_MINS = 5
71

72
_R = TypeVar("_R")  # Return type for collective_rpc
73

74
75
76
77

class EngineCore:
    """Inner loop of vLLM's Engine."""

78
79
80
81
82
    def __init__(
        self,
        vllm_config: VllmConfig,
        executor_class: type[Executor],
        log_stats: bool,
83
        executor_fail_callback: Callable | None = None,
84
    ):
85
86
        # plugins need to be loaded at the engine/scheduler level too
        from vllm.plugins import load_general_plugins
87

88
89
        load_general_plugins()

90
        self.vllm_config = vllm_config
91
        if vllm_config.parallel_config.data_parallel_rank == 0:
92
93
94
95
96
            logger.info(
                "Initializing a V1 LLM engine (v%s) with config: %s",
                VLLM_VERSION,
                vllm_config,
            )
97

98
99
        self.log_stats = log_stats

100
101
        # Setup Model.
        self.model_executor = executor_class(vllm_config)
102
        if executor_fail_callback is not None:
103
            self.model_executor.register_failure_callback(executor_fail_callback)
104

105
106
        self.available_gpu_memory_for_kv_cache = -1

107
        # Setup KV Caches and update CacheConfig after profiling.
108
109
110
        num_gpu_blocks, num_cpu_blocks, kv_cache_config = self._initialize_kv_caches(
            vllm_config
        )
111

112
113
        vllm_config.cache_config.num_gpu_blocks = num_gpu_blocks
        vllm_config.cache_config.num_cpu_blocks = num_cpu_blocks
114
        self.collective_rpc("initialize_cache", args=(num_gpu_blocks, num_cpu_blocks))
115

116
117
        self.structured_output_manager = StructuredOutputManager(vllm_config)

118
        # Setup scheduler.
119
        Scheduler = vllm_config.scheduler_config.get_scheduler_cls()
120

121
122
123
124
125
126
        if len(kv_cache_config.kv_cache_groups) == 0:
            # Encoder models without KV cache don't support
            # chunked prefill. But do SSM models?
            logger.info("Disabling chunked prefill for model without KVCache")
            vllm_config.scheduler_config.chunked_prefill_enabled = False

127
128
129
130
131
        scheduler_block_size = (
            vllm_config.cache_config.block_size
            * vllm_config.parallel_config.decode_context_parallel_size
        )

132
        self.scheduler: SchedulerInterface = Scheduler(
133
            vllm_config=vllm_config,
134
135
            kv_cache_config=kv_cache_config,
            structured_output_manager=self.structured_output_manager,
136
            include_finished_set=vllm_config.parallel_config.data_parallel_size > 1,
137
            log_stats=self.log_stats,
138
            block_size=scheduler_block_size,
139
        )
140
        self.use_spec_decode = vllm_config.speculative_config is not None
141
        if self.scheduler.connector is not None:  # type: ignore
142
            self.model_executor.init_kv_output_aggregator(self.scheduler.connector)  # type: ignore
143

144
        self.mm_registry = mm_registry = MULTIMODAL_REGISTRY
145
        self.mm_receiver_cache = engine_receiver_cache_from_config(
146
147
            vllm_config, mm_registry
        )
148

149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
        # If a KV connector is initialized for scheduler, we want to collect
        # handshake metadata from all workers so the connector in the scheduler
        # will have the full context
        kv_connector = self.scheduler.get_kv_connector()
        if kv_connector is not None:
            # Collect and store KV connector xfer metadata from workers
            # (after KV cache registration)
            xfer_handshake_metadata = (
                self.model_executor.get_kv_connector_handshake_metadata()
            )

            if xfer_handshake_metadata:
                # xfer_handshake_metadata is list of dicts from workers
                # Each dict already has structure {tp_rank: metadata}
                # Merge all worker dicts into a single dict
                content: dict[int, Any] = {}
                for worker_dict in xfer_handshake_metadata:
                    if worker_dict is not None:
                        content.update(worker_dict)
                kv_connector.set_xfer_handshake_metadata(content)

170
171
172
173
174
        # Setup batch queue for pipeline parallelism.
        # Batch queue for scheduled batches. This enables us to asynchronously
        # schedule and execute batches, and is required by pipeline parallelism
        # to eliminate pipeline bubbles.
        self.batch_queue_size = self.model_executor.max_concurrent_batches
175
176
177
        self.batch_queue: (
            deque[tuple[Future[ModelRunnerOutput], SchedulerOutput]] | None
        ) = None
178
        if self.batch_queue_size > 1:
179
            logger.info("Batch queue is enabled with size %d", self.batch_queue_size)
180
            self.batch_queue = deque(maxlen=self.batch_queue_size)
181

182
        self.request_block_hasher: Callable[[Request], list[BlockHash]] | None = None
183
184
        if (
            self.vllm_config.cache_config.enable_prefix_caching
185
            or kv_connector is not None
186
        ):
187
            caching_hash_fn = get_hash_fn_by_name(
188
189
                vllm_config.cache_config.prefix_caching_hash_algo
            )
190
191
192
            init_none_hash(caching_hash_fn)

            self.request_block_hasher = get_request_block_hasher(
193
                scheduler_block_size, caching_hash_fn
194
            )
195

196
197
198
        self.step_fn = (
            self.step if self.batch_queue is None else self.step_with_batch_queue
        )
199

200
    def _initialize_kv_caches(
201
202
        self, vllm_config: VllmConfig
    ) -> tuple[int, int, KVCacheConfig]:
203
        start = time.time()
204

205
        # Get all kv cache needed by the model
206
        kv_cache_specs = self.model_executor.get_kv_cache_specs()
207

208
209
        has_kv_cache = any(kv_cache_spec for kv_cache_spec in kv_cache_specs)
        if has_kv_cache:
210
211
212
            if os.environ.get("VLLM_ELASTIC_EP_SCALE_UP_LAUNCH") == "1":
                dp_group = getattr(self, "dp_group", None)
                assert dp_group is not None
213
                self.available_gpu_memory_for_kv_cache = (
214
                    ParallelConfig.sync_kv_cache_memory_size(dp_group, -1)
215
216
217
218
                )
                available_gpu_memory = [self.available_gpu_memory_for_kv_cache] * len(
                    kv_cache_specs
                )
219
220
221
            else:
                # Profiles the peak memory usage of the model to determine how
                # much memory can be allocated for kv cache.
222
223
                available_gpu_memory = self.model_executor.determine_available_memory()
                self.available_gpu_memory_for_kv_cache = available_gpu_memory[0]
224
225
226
        else:
            # Attention free models don't need memory for kv cache
            available_gpu_memory = [0] * len(kv_cache_specs)
227

228
        assert len(kv_cache_specs) == len(available_gpu_memory)
229

230
231
232
233
        kv_cache_configs = get_kv_cache_configs(
            vllm_config, kv_cache_specs, available_gpu_memory
        )
        scheduler_kv_cache_config = generate_scheduler_kv_cache_config(kv_cache_configs)
234
        num_gpu_blocks = scheduler_kv_cache_config.num_blocks
235
        num_cpu_blocks = 0
236
237

        # Initialize kv cache and warmup the execution
238
        self.model_executor.initialize_from_config(kv_cache_configs)
239

240
        elapsed = time.time() - start
241
        logger.info_once(
242
243
            ("init engine (profile, create kv cache, warmup model) took %.2f seconds"),
            elapsed,
244
            scope="local",
245
        )
246
        return num_gpu_blocks, num_cpu_blocks, scheduler_kv_cache_config
247

248
249
250
    def get_supported_tasks(self) -> tuple[SupportedTask, ...]:
        return self.model_executor.supported_tasks

251
252
    def add_request(self, request: Request, request_wave: int = 0):
        """Add request to the scheduler.
253

254
255
256
        `request_wave`: indicate which wave of requests this is expected to
        belong to in DP case
        """
257
258
259
        # Validate the request_id type.
        if not isinstance(request.request_id, str):
            raise TypeError(
260
261
                f"request_id must be a string, got {type(request.request_id)}"
            )
262

263
        if pooling_params := request.pooling_params:
264
            supported_pooling_tasks = [
265
                task for task in self.get_supported_tasks() if task in POOLING_TASKS
266
267
            ]

268
            if pooling_params.task not in supported_pooling_tasks:
269
270
271
272
                raise ValueError(
                    f"Unsupported task: {pooling_params.task!r} "
                    f"Supported tasks: {supported_pooling_tasks}"
                )
273

274
        if request.kv_transfer_params is not None and (
275
276
277
278
279
280
            not self.scheduler.get_kv_connector()
        ):
            logger.warning(
                "Got kv_transfer_params, but no KVConnector found. "
                "Disabling KVTransfer for this request."
            )
Robert Shaw's avatar
Robert Shaw committed
281

282
        self.scheduler.add_request(request)
283

284
    def abort_requests(self, request_ids: list[str]):
285
286
287
288
289
        """Abort requests from the scheduler."""

        # TODO: The scheduler doesn't really need to know the
        # specific finish reason, TBD whether we propagate that
        # (i.e. client-aborted vs stop criteria met).
290
        self.scheduler.finish_requests(request_ids, RequestStatus.FINISHED_ABORTED)
291

292
293
    @contextmanager
    def log_error_detail(self, scheduler_output: SchedulerOutput):
294
        """Execute the model and log detailed info on failure."""
295
        try:
296
            yield
297
298
299
300
301
        except Exception as err:
            # We do not want to catch BaseException here since we're only
            # interested in dumping info when the exception is due to an
            # error from execute_model itself.

302
            # NOTE: This method is exception-free
303
304
305
            dump_engine_exception(
                self.vllm_config, scheduler_output, self.scheduler.make_stats()
            )
306
307
            raise err

308
    def step(self) -> tuple[dict[int, EngineCoreOutputs], bool]:
309
310
311
312
313
        """Schedule, execute, and make output.

        Returns tuple of outputs and a flag indicating whether the model
        was executed.
        """
314

315
316
317
        # Check for any requests remaining in the scheduler - unfinished,
        # or finished and not yet removed from the batch.
        if not self.scheduler.has_requests():
318
            return {}, False
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
        with record_function_or_nullcontext("core step: schedule"):
            scheduler_output = self.scheduler.schedule()

        with record_function_or_nullcontext("core step: execute_model"):
            future = self.model_executor.execute_model(scheduler_output, non_block=True)
            grammar_output = self.scheduler.get_grammar_bitmask(scheduler_output)
            with self.log_error_detail(scheduler_output):
                model_output = future.result()
                if model_output is None:
                    model_output = self.model_executor.sample_tokens(grammar_output)

        with record_function_or_nullcontext("core step: update_from_output"):
            engine_core_outputs = self.scheduler.update_from_output(
                scheduler_output, model_output
            )
334

335
        return engine_core_outputs, scheduler_output.total_num_scheduled_tokens > 0
336

337
338
339
340
341
342
343
    def post_step(self, model_executed: bool) -> None:
        if self.use_spec_decode and model_executed:
            # Take the draft token ids.
            draft_token_ids = self.model_executor.take_draft_token_ids()
            if draft_token_ids is not None:
                self.scheduler.update_draft_token_ids(draft_token_ids)

344
    def step_with_batch_queue(
345
        self,
346
    ) -> tuple[dict[int, EngineCoreOutputs] | None, bool]:
347
348
349
350
        """Schedule and execute batches with the batch queue.
        Note that if nothing to output in this step, None is returned.

        The execution flow is as follows:
351
352
353
354
        1. Try to schedule a new batch if the batch queue is not full.
        If a new batch is scheduled, directly return an empty engine core
        output. In other words, fulfilling the batch queue has a higher priority
        than getting model outputs.
355
356
357
358
359
        2. If there is no new scheduled batch, meaning that the batch queue
        is full or no other requests can be scheduled, we block until the first
        batch in the job queue is finished.
        3. Update the scheduler from the output.
        """
360
361
        batch_queue = self.batch_queue
        assert batch_queue is not None
362

363
364
365
        # Try to schedule a new batch if the batch queue is not full, but
        # the scheduler may return an empty batch if all requests are scheduled.
        # Note that this is not blocking.
366
        assert len(batch_queue) < self.batch_queue_size
367

368
        model_executed = False
369
        deferred_scheduler_output = None
370
        if self.scheduler.has_requests():
371
372
373
374
375
376
377
378
            with record_function_or_nullcontext("core step_with_batch_queue: schedule"):
                scheduler_output = self.scheduler.schedule()
            with record_function_or_nullcontext(
                "core step_with_batch_queue: execute_model"
            ):
                exec_future = self.model_executor.execute_model(
                    scheduler_output, non_block=True
                )
379
            model_executed = scheduler_output.total_num_scheduled_tokens > 0
380
381

            if scheduler_output.pending_structured_output_tokens:
382
383
384
385
386
387
388
389
390
391
392
                with record_function_or_nullcontext(
                    "core step_with_batch_queue: pending_structured_output_tokens"
                ):
                    # We need to defer sampling until we have processed the model output
                    # from the prior step.
                    deferred_scheduler_output = scheduler_output
                    # Block-wait for execute to return
                    # (continues running async on the GPU).
                    with self.log_error_detail(scheduler_output):
                        exec_result = exec_future.result()
                        assert exec_result is None
393
            else:
394
395
396
397
398
399
400
401
                with record_function_or_nullcontext(
                    "core step_with_batch_queue: get_grammar_bitmask"
                ):
                    # We aren't waiting for any tokens, get any grammar
                    # output immediately.
                    grammar_output = self.scheduler.get_grammar_bitmask(
                        scheduler_output
                    )
402
403
404
405
406
                # Block-wait for execute to return (continues running async on the GPU).
                with self.log_error_detail(scheduler_output):
                    exec_result = exec_future.result()

                if exec_result is None:
407
408
409
410
411
412
413
                    with record_function_or_nullcontext(
                        "core step_with_batch_queue: sample_tokens"
                    ):
                        # Call sample tokens.
                        future = self.model_executor.sample_tokens(
                            grammar_output, non_block=True
                        )
414
415
416
417
418
419
420
421
422
423
424
425
426
                else:
                    # No sampling required (e.g. all requests finished).
                    future = cast(Future[ModelRunnerOutput], exec_future)
                # Add this step's future to the queue.
                batch_queue.appendleft((future, scheduler_output))
                if (
                    model_executed
                    and len(batch_queue) < self.batch_queue_size
                    and not batch_queue[-1][0].done()
                ):
                    # Don't block on next worker response unless the queue is full
                    # or there are no more requests to schedule.
                    return None, True
427
428
429
430
431
432

        elif not batch_queue:
            # Queue is empty. We should not reach here since this method should
            # only be called when the scheduler contains requests or the queue
            # is non-empty.
            return None, False
433
434
435
436
437
438
439
440
441
442
443
        with record_function_or_nullcontext("core step_with_batch_queue: model_output"):
            # Block until the next result is available.
            future, scheduler_output = batch_queue.pop()
            with self.log_error_detail(scheduler_output):
                model_output = future.result()
        with record_function_or_nullcontext(
            "core step_with_batch_queue: update_from_output"
        ):
            engine_core_outputs = self.scheduler.update_from_output(
                scheduler_output, model_output
            )
444
445
446
447
448

        # NOTE(nick): We can either handle the deferred tasks here or save
        # in a field and do it immediately once step_with_batch_queue is
        # re-called. The latter slightly favors TTFT over TPOT/throughput.
        if deferred_scheduler_output:
449
450
451
452
453
454
455
456
457
458
459
460
            with record_function_or_nullcontext(
                "core step_with_batch_queue: deferred_scheduler_output"
            ):
                # We now have the tokens needed to compute the bitmask for the
                # deferred request. Get the bitmask and call sample tokens.
                grammar_output = self.scheduler.get_grammar_bitmask(
                    deferred_scheduler_output
                )
                future = self.model_executor.sample_tokens(
                    grammar_output, non_block=True
                )
                batch_queue.appendleft((future, deferred_scheduler_output))
461

462
        return engine_core_outputs, model_executed
463

464
    def shutdown(self):
465
        self.structured_output_manager.clear_backend()
466
467
        if self.model_executor:
            self.model_executor.shutdown()
468
469
        if self.scheduler:
            self.scheduler.shutdown()
470

471
    def profile(self, is_start: bool = True):
472
        self.model_executor.profile(is_start)
473

474
475
    def reset_mm_cache(self):
        # NOTE: Since this is mainly for debugging, we don't attempt to
476
        # re-sync the internal caches (P0 sender, P1 receiver)
477
        if self.scheduler.has_unfinished_requests():
478
479
480
481
            logger.warning(
                "Resetting the multi-modal cache when requests are "
                "in progress may lead to desynced internal caches."
            )
482

483
        # The cache either exists in EngineCore or WorkerWrapperBase
484
485
        if self.mm_receiver_cache is not None:
            self.mm_receiver_cache.clear_cache()
486

487
488
        self.model_executor.reset_mm_cache()

489
490
491
    def reset_prefix_cache(self):
        self.scheduler.reset_prefix_cache()

492
493
494
    def sleep(self, level: int = 1):
        self.model_executor.sleep(level)

495
    def wake_up(self, tags: list[str] | None = None):
496
        self.model_executor.wake_up(tags)
497

498
499
500
    def is_sleeping(self) -> bool:
        return self.model_executor.is_sleeping

501
    def execute_dummy_batch(self):
502
        self.model_executor.execute_dummy_batch()
503

504
505
506
507
508
509
    def add_lora(self, lora_request: LoRARequest) -> bool:
        return self.model_executor.add_lora(lora_request)

    def remove_lora(self, lora_id: int) -> bool:
        return self.model_executor.remove_lora(lora_id)

510
    def list_loras(self) -> set[int]:
511
512
513
514
        return self.model_executor.list_loras()

    def pin_lora(self, lora_id: int) -> bool:
        return self.model_executor.pin_lora(lora_id)
515

516
517
518
    def save_sharded_state(
        self,
        path: str,
519
520
        pattern: str | None = None,
        max_size: int | None = None,
521
    ) -> None:
522
523
524
525
526
527
        self.model_executor.save_sharded_state(
            path=path, pattern=pattern, max_size=max_size
        )

    def collective_rpc(
        self,
528
529
        method: str | Callable[..., _R],
        timeout: float | None = None,
530
        args: tuple = (),
531
        kwargs: dict[str, Any] | None = None,
532
533
    ) -> list[_R]:
        return self.model_executor.collective_rpc(method, timeout, args, kwargs)
534

535
    def preprocess_add_request(self, request: EngineCoreRequest) -> tuple[Request, int]:
536
        """Preprocess the request.
537

538
539
540
        This function could be directly used in input processing thread to allow
        request initialization running in parallel with Model forward
        """
541
542
        # Note on thread safety: no race condition.
        # `mm_receiver_cache` is reset at the end of LLMEngine init,
543
        # and will only be accessed in the input processing thread afterwards.
544
        if self.mm_receiver_cache is not None and request.mm_features:
545
546
547
            request.mm_features = self.mm_receiver_cache.get_and_update_features(
                request.mm_features
            )
548

549
        req = Request.from_engine_core_request(request, self.request_block_hasher)
550
551
552
553
554
555
556
557
558
        if req.use_structured_output:
            # Note on thread safety: no race condition.
            # `grammar_init` is only invoked in input processing thread. For
            # `structured_output_manager`, each request is independent and
            # grammar compilation is async. Scheduler always checks grammar
            # compilation status before scheduling request.
            self.structured_output_manager.grammar_init(req)
        return req, request.current_wave

559
560
561
562

class EngineCoreProc(EngineCore):
    """ZMQ-wrapper for running EngineCore in background process."""

563
    ENGINE_CORE_DEAD = b"ENGINE_CORE_DEAD"
564

565
566
    def __init__(
        self,
567
        vllm_config: VllmConfig,
568
        local_client: bool,
569
        handshake_address: str,
570
        executor_class: type[Executor],
571
        log_stats: bool,
572
        client_handshake_address: str | None = None,
573
        engine_index: int = 0,
574
    ):
Rui Qiao's avatar
Rui Qiao committed
575
        self.input_queue = queue.Queue[tuple[EngineCoreRequestType, Any]]()
576
        self.output_queue = queue.Queue[tuple[int, EngineCoreOutputs] | bytes]()
Rui Qiao's avatar
Rui Qiao committed
577
        executor_fail_callback = lambda: self.input_queue.put_nowait(
578
579
            (EngineCoreRequestType.EXECUTOR_FAILED, b"")
        )
580

Rui Qiao's avatar
Rui Qiao committed
581
582
583
        self.engine_index = engine_index
        identity = self.engine_index.to_bytes(length=2, byteorder="little")
        self.engines_running = False
584

585
586
587
588
589
590
591
        with self._perform_handshakes(
            handshake_address,
            identity,
            local_client,
            vllm_config,
            client_handshake_address,
        ) as addresses:
592
            self.client_count = len(addresses.outputs)
593
594

            # Set up data parallel environment.
595
            self.has_coordinator = addresses.coordinator_output is not None
596
            self.frontend_stats_publish_address = (
597
598
599
600
601
602
603
                addresses.frontend_stats_publish_address
            )
            logger.debug(
                "Has DP Coordinator: %s, stats publish address: %s",
                self.has_coordinator,
                self.frontend_stats_publish_address,
            )
604
            # Only publish request queue stats to coordinator for "internal"
605
            # and "hybrid" LB modes .
606
607
            self.publish_dp_lb_stats = (
                self.has_coordinator
608
609
                and not vllm_config.parallel_config.data_parallel_external_lb
            )
610

611
612
            self._init_data_parallel(vllm_config)

613
614
615
            super().__init__(
                vllm_config, executor_class, log_stats, executor_fail_callback
            )
616

617
618
619
620
621
622
            # Background Threads and Queues for IO. These enable us to
            # overlap ZMQ socket IO with GPU since they release the GIL,
            # and to overlap some serialization/deserialization with the
            # model forward pass.
            # Threads handle Socket <-> Queues and core_busy_loop uses Queue.
            ready_event = threading.Event()
623
624
625
626
627
628
629
630
631
632
            input_thread = threading.Thread(
                target=self.process_input_sockets,
                args=(
                    addresses.inputs,
                    addresses.coordinator_input,
                    identity,
                    ready_event,
                ),
                daemon=True,
            )
633
634
635
636
            input_thread.start()

            self.output_thread = threading.Thread(
                target=self.process_output_sockets,
637
638
639
640
641
642
643
                args=(
                    addresses.outputs,
                    addresses.coordinator_output,
                    self.engine_index,
                ),
                daemon=True,
            )
644
645
646
647
648
649
            self.output_thread.start()

            # Don't complete handshake until DP coordinator ready message is
            # received.
            while not ready_event.wait(timeout=10):
                if not input_thread.is_alive():
650
                    raise RuntimeError("Input socket thread died during startup")
651
652
653
                assert addresses.coordinator_input is not None
                logger.info("Waiting for READY message from DP Coordinator...")

654
655
656
657
658
        # Mark the startup heap as static so that it's ignored by GC.
        # Reduces pause times of oldest generation collections.
        gc.collect()
        gc.freeze()

659
660
661
        # If enable, attach GC debugger after static variable freeze.
        maybe_attach_gc_debug_callback()

662
663
664
665
        # Enable environment variable cache (e.g. assume no more
        # environment variable overrides after this point)
        enable_envs_cache()

Rui Qiao's avatar
Rui Qiao committed
666
    @contextmanager
667
668
669
670
671
672
    def _perform_handshakes(
        self,
        handshake_address: str,
        identity: bytes,
        local_client: bool,
        vllm_config: VllmConfig,
673
        client_handshake_address: str | None,
Rui Qiao's avatar
Rui Qiao committed
674
    ) -> Generator[EngineZmqAddresses, None, None]:
675
676
677
678
679
        """
        Perform startup handshakes.

        For DP=1 or offline mode, this is with the colocated front-end process.

680
        For DP>1 with internal load-balancing this is with the shared front-end
681
682
        process which may reside on a different node.

683
        For DP>1 with external or hybrid load-balancing, two handshakes are
684
        performed:
685
686
687
688
            - With the rank 0 front-end process which retrieves the
              DP Coordinator ZMQ addresses and DP process group address.
            - With the colocated front-end process which retrieves the
              client input/output socket addresses.
689
690
        with the exception of the rank 0 and colocated engines themselves which
        don't require the second handshake.
691
692
693
694
695
696

        Here, "front-end" process can mean the process containing the engine
        core client (which is the API server process in the case the API
        server is not scaled out), OR the launcher process running the
        run_multi_api_server() function in serve.py.
        """
Rui Qiao's avatar
Rui Qiao committed
697
        input_ctx = zmq.Context()
698
        is_local = local_client and client_handshake_address is None
699
        headless = not local_client
700
701
702
703
704
705
706
707
708
        handshake = self._perform_handshake(
            input_ctx,
            handshake_address,
            identity,
            is_local,
            headless,
            vllm_config,
            vllm_config.parallel_config,
        )
709
710
711
712
        if client_handshake_address is None:
            with handshake as addresses:
                yield addresses
        else:
713
            assert local_client
714
            local_handshake = self._perform_handshake(
715
716
                input_ctx, client_handshake_address, identity, True, False, vllm_config
            )
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
            with handshake as addresses, local_handshake as client_addresses:
                addresses.inputs = client_addresses.inputs
                addresses.outputs = client_addresses.outputs
                yield addresses

        # Update config which may have changed from the handshake
        vllm_config.__post_init__()

    @contextmanager
    def _perform_handshake(
        self,
        ctx: zmq.Context,
        handshake_address: str,
        identity: bytes,
        local_client: bool,
732
        headless: bool,
733
        vllm_config: VllmConfig,
734
        parallel_config_to_update: ParallelConfig | None = None,
735
    ) -> Generator[EngineZmqAddresses, None, None]:
736
737
738
739
740
741
742
743
        with make_zmq_socket(
            ctx,
            handshake_address,
            zmq.DEALER,
            identity=identity,
            linger=5000,
            bind=False,
        ) as handshake_socket:
Rui Qiao's avatar
Rui Qiao committed
744
            # Register engine with front-end.
745
746
747
            addresses = self.startup_handshake(
                handshake_socket, local_client, headless, parallel_config_to_update
            )
Rui Qiao's avatar
Rui Qiao committed
748
749
750
751
            yield addresses

            # Send ready message.
            num_gpu_blocks = vllm_config.cache_config.num_gpu_blocks
752
753
754
755
            # We pass back the coordinator stats update address here for the
            # external LB case for our colocated front-end to use (coordinator
            # only runs with rank 0).
            dp_stats_address = self.frontend_stats_publish_address
756
757
758
759
760
761
762
763
764
765
766
767

            # Include config hash for DP configuration validation
            ready_msg = {
                "status": "READY",
                "local": local_client,
                "headless": headless,
                "num_gpu_blocks": num_gpu_blocks,
                "dp_stats_address": dp_stats_address,
            }
            if vllm_config.parallel_config.data_parallel_size > 1:
                ready_msg["parallel_config_hash"] = (
                    vllm_config.parallel_config.compute_hash()
768
                )
769
770

            handshake_socket.send(msgspec.msgpack.encode(ready_msg))
Rui Qiao's avatar
Rui Qiao committed
771

772
    @staticmethod
773
    def startup_handshake(
774
775
        handshake_socket: zmq.Socket,
        local_client: bool,
776
        headless: bool,
777
        parallel_config: ParallelConfig | None = None,
778
    ) -> EngineZmqAddresses:
779
        # Send registration message.
780
        handshake_socket.send(
781
782
783
784
785
786
787
788
            msgspec.msgpack.encode(
                {
                    "status": "HELLO",
                    "local": local_client,
                    "headless": headless,
                }
            )
        )
789
790

        # Receive initialization message.
791
        logger.debug("Waiting for init message from front-end.")
792
        if not handshake_socket.poll(timeout=HANDSHAKE_TIMEOUT_MINS * 60_000):
793
794
795
796
797
            raise RuntimeError(
                "Did not receive response from front-end "
                f"process within {HANDSHAKE_TIMEOUT_MINS} "
                f"minutes"
            )
798
799
        init_bytes = handshake_socket.recv()
        init_message: EngineHandshakeMetadata = msgspec.msgpack.decode(
800
801
            init_bytes, type=EngineHandshakeMetadata
        )
802
803
        logger.debug("Received init message: %s", init_message)

804
805
806
        if parallel_config is not None:
            for key, value in init_message.parallel_config.items():
                setattr(parallel_config, key, value)
807

808
        return init_message.addresses
809
810

    @staticmethod
811
    def run_engine_core(*args, dp_rank: int = 0, local_dp_rank: int = 0, **kwargs):
812
813
        """Launch EngineCore busy loop in background process."""

814
815
816
817
818
        # Signal handler used for graceful termination.
        # SystemExit exception is only raised once to allow this and worker
        # processes to terminate without error
        shutdown_requested = False

819
820
821
        # Ensure we can serialize transformer config after spawning
        maybe_register_config_serialize_by_value()

822
823
824
825
826
827
828
829
830
831
        def signal_handler(signum, frame):
            nonlocal shutdown_requested
            if not shutdown_requested:
                shutdown_requested = True
                raise SystemExit()

        # Either SIGTERM or SIGINT will terminate the engine_core
        signal.signal(signal.SIGTERM, signal_handler)
        signal.signal(signal.SIGINT, signal_handler)

832
        engine_core: EngineCoreProc | None = None
833
        try:
834
            parallel_config: ParallelConfig = kwargs["vllm_config"].parallel_config
835
            if parallel_config.data_parallel_size > 1 or dp_rank > 0:
836
                set_process_title("EngineCore", f"DP{dp_rank}")
837
                decorate_logs()
838
839
840
841
842
                # Set data parallel rank for this engine process.
                parallel_config.data_parallel_rank = dp_rank
                parallel_config.data_parallel_rank_local = local_dp_rank
                engine_core = DPEngineCoreProc(*args, **kwargs)
            else:
843
                set_process_title("EngineCore")
844
                decorate_logs()
845
846
                engine_core = EngineCoreProc(*args, **kwargs)

847
848
            engine_core.run_busy_loop()

849
        except SystemExit:
850
            logger.debug("EngineCore exiting.")
851
            raise
852
853
854
855
856
857
858
        except Exception as e:
            if engine_core is None:
                logger.exception("EngineCore failed to start.")
            else:
                logger.exception("EngineCore encountered a fatal error.")
                engine_core._send_engine_dead()
            raise e
859
860
861
862
        finally:
            if engine_core is not None:
                engine_core.shutdown()

863
864
865
    def _init_data_parallel(self, vllm_config: VllmConfig):
        pass

866
867
868
    def run_busy_loop(self):
        """Core busy loop of the EngineCore."""

869
870
        # Loop until process is sent a SIGINT or SIGTERM
        while True:
871
            # 1) Poll the input queue until there is work to do.
872
873
874
875
876
877
878
879
            self._process_input_queue()
            # 2) Step the engine core and return the outputs.
            self._process_engine_step()

    def _process_input_queue(self):
        """Exits when an engine step needs to be performed."""

        waited = False
880
881
882
883
884
        while (
            not self.engines_running
            and not self.scheduler.has_requests()
            and not self.batch_queue
        ):
885
886
887
888
889
890
891
            if logger.isEnabledFor(DEBUG) and self.input_queue.empty():
                logger.debug("EngineCore waiting for work.")
                waited = True
            req = self.input_queue.get()
            self._handle_client_request(*req)

        if waited:
892
            logger.debug("EngineCore loop active.")
893
894
895
896
897
898

        # Handle any more client requests.
        while not self.input_queue.empty():
            req = self.input_queue.get_nowait()
            self._handle_client_request(*req)

899
    def _process_engine_step(self) -> bool:
900
901
902
        """Called only when there are unfinished local requests."""

        # Step the engine core.
903
        outputs, model_executed = self.step_fn()
904
        # Put EngineCoreOutputs into the output queue.
905
        for output in outputs.items() if outputs else ():
906
            self.output_queue.put_nowait(output)
907
908
        # Post-step hook.
        self.post_step(model_executed)
909

910
911
        return model_executed

912
913
914
    def _handle_client_request(
        self, request_type: EngineCoreRequestType, request: Any
    ) -> None:
915
        """Dispatch request from client."""
916

917
        if request_type == EngineCoreRequestType.ADD:
918
919
            req, request_wave = request
            self.add_request(req, request_wave)
920
        elif request_type == EngineCoreRequestType.ABORT:
921
            self.abort_requests(request)
922
        elif request_type == EngineCoreRequestType.UTILITY:
923
            client_idx, call_id, method_name, args = request
924
925
926
            output = UtilityOutput(call_id)
            try:
                method = getattr(self, method_name)
927
928
                result = method(*self._convert_msgspec_args(method, args))
                output.result = UtilityResult(result)
929
930
            except BaseException as e:
                logger.exception("Invocation of %s method failed", method_name)
931
932
933
                output.failure_message = (
                    f"Call to {method_name} method failed: {str(e)}"
                )
934
            self.output_queue.put_nowait(
935
936
                (client_idx, EngineCoreOutputs(utility_output=output))
            )
937
938
939
        elif request_type == EngineCoreRequestType.EXECUTOR_FAILED:
            raise RuntimeError("Executor failed.")
        else:
940
941
942
            logger.error(
                "Unrecognized input request type encountered: %s", request_type
            )
943
944
945
946

    @staticmethod
    def _convert_msgspec_args(method, args):
        """If a provided arg type doesn't match corresponding target method
947
        arg type, try converting to msgspec object."""
948
949
950
951
952
        if not args:
            return args
        arg_types = signature(method).parameters.values()
        assert len(args) <= len(arg_types)
        return tuple(
953
954
            msgspec.convert(v, type=p.annotation)
            if isclass(p.annotation)
955
            and issubclass(p.annotation, msgspec.Struct)
956
957
958
959
            and not isinstance(v, p.annotation)
            else v
            for v, p in zip(args, arg_types)
        )
960

961
962
963
964
965
966
967
968
969
    def _send_engine_dead(self):
        """Send EngineDead status to the EngineCoreClient."""

        # Put ENGINE_CORE_DEAD in the queue.
        self.output_queue.put_nowait(EngineCoreProc.ENGINE_CORE_DEAD)

        # Wait until msg sent by the daemon before shutdown.
        self.output_thread.join(timeout=5.0)
        if self.output_thread.is_alive():
970
971
972
973
            logger.fatal(
                "vLLM shutdown signal from EngineCore failed "
                "to send. Please report this issue."
            )
974

975
976
977
    def process_input_sockets(
        self,
        input_addresses: list[str],
978
        coord_input_address: str | None,
979
980
981
        identity: bytes,
        ready_event: threading.Event,
    ):
982
983
984
        """Input socket IO thread."""

        # Msgpack serialization decoding.
985
986
        add_request_decoder = MsgpackDecoder(EngineCoreRequest)
        generic_decoder = MsgpackDecoder()
987

988
989
990
        with ExitStack() as stack, zmq.Context() as ctx:
            input_sockets = [
                stack.enter_context(
991
992
993
994
                    make_zmq_socket(
                        ctx, input_address, zmq.DEALER, identity=identity, bind=False
                    )
                )
995
996
997
998
999
1000
                for input_address in input_addresses
            ]
            if coord_input_address is None:
                coord_socket = None
            else:
                coord_socket = stack.enter_context(
1001
1002
1003
1004
1005
1006
1007
1008
                    make_zmq_socket(
                        ctx,
                        coord_input_address,
                        zmq.XSUB,
                        identity=identity,
                        bind=False,
                    )
                )
1009
                # Send subscription message to coordinator.
1010
                coord_socket.send(b"\x01")
1011
1012
1013
1014
1015
1016
1017

            # Register sockets with poller.
            poller = zmq.Poller()
            for input_socket in input_sockets:
                # Send initial message to each input socket - this is required
                # before the front-end ROUTER socket can send input messages
                # back to us.
1018
                input_socket.send(b"")
1019
                poller.register(input_socket, zmq.POLLIN)
1020

1021
            if coord_socket is not None:
1022
1023
                # Wait for ready message from coordinator.
                assert coord_socket.recv() == b"READY"
1024
                poller.register(coord_socket, zmq.POLLIN)
1025

1026
1027
            ready_event.set()
            del ready_event
1028
1029
1030
            while True:
                for input_socket, _ in poller.poll():
                    # (RequestType, RequestData)
1031
1032
                    type_frame, *data_frames = input_socket.recv_multipart(copy=False)
                    request_type = EngineCoreRequestType(bytes(type_frame.buffer))
1033
1034

                    # Deserialize the request data.
1035
1036
1037
1038
1039
                    if request_type == EngineCoreRequestType.ADD:
                        request = add_request_decoder.decode(data_frames)
                        request = self.preprocess_add_request(request)
                    else:
                        request = generic_decoder.decode(data_frames)
1040
1041
1042
1043

                    # Push to input queue for core busy loop.
                    self.input_queue.put_nowait((request_type, request))

1044
1045
1046
    def process_output_sockets(
        self,
        output_paths: list[str],
1047
        coord_output_path: str | None,
1048
1049
        engine_index: int,
    ):
1050
1051
1052
        """Output socket IO thread."""

        # Msgpack serialization encoding.
1053
        encoder = MsgpackEncoder()
1054
1055
1056
1057
1058
1059
        # Send buffers to reuse.
        reuse_buffers: list[bytearray] = []
        # Keep references to outputs and buffers until zmq is finished
        # with them (outputs may contain tensors/np arrays whose
        # backing buffers were extracted for zero-copy send).
        pending = deque[tuple[zmq.MessageTracker, Any, bytearray]]()
1060

1061
1062
        # We must set linger to ensure the ENGINE_CORE_DEAD
        # message is sent prior to closing the socket.
1063
1064
1065
        with ExitStack() as stack, zmq.Context() as ctx:
            sockets = [
                stack.enter_context(
1066
1067
                    make_zmq_socket(ctx, output_path, zmq.PUSH, linger=4000)
                )
1068
1069
                for output_path in output_paths
            ]
1070
1071
1072
1073
1074
1075
1076
1077
1078
            coord_socket = (
                stack.enter_context(
                    make_zmq_socket(
                        ctx, coord_output_path, zmq.PUSH, bind=False, linger=4000
                    )
                )
                if coord_output_path is not None
                else None
            )
1079
1080
            max_reuse_bufs = len(sockets) + 1

1081
            while True:
1082
1083
1084
1085
                output = self.output_queue.get()
                if output == EngineCoreProc.ENGINE_CORE_DEAD:
                    for socket in sockets:
                        socket.send(output)
1086
                    break
1087
1088
                assert not isinstance(output, bytes)
                client_index, outputs = output
1089
                outputs.engine_index = engine_index
1090

1091
1092
1093
1094
1095
1096
1097
                if client_index == -1:
                    # Don't reuse buffer for coordinator message
                    # which will be very small.
                    assert coord_socket is not None
                    coord_socket.send_multipart(encoder.encode(outputs))
                    continue

1098
1099
1100
1101
1102
                # Reclaim buffers that zmq is finished with.
                while pending and pending[-1][0].done:
                    reuse_buffers.append(pending.pop()[2])

                buffer = reuse_buffers.pop() if reuse_buffers else bytearray()
1103
                buffers = encoder.encode_into(outputs, buffer)
1104
1105
1106
                tracker = sockets[client_index].send_multipart(
                    buffers, copy=False, track=True
                )
1107
1108
1109
                if not tracker.done:
                    ref = outputs if len(buffers) > 1 else None
                    pending.appendleft((tracker, ref, buffer))
1110
1111
                elif len(reuse_buffers) < max_reuse_bufs:
                    # Limit the number of buffers to reuse.
1112
                    reuse_buffers.append(buffer)
1113
1114
1115
1116
1117
1118
1119
1120
1121


class DPEngineCoreProc(EngineCoreProc):
    """ZMQ-wrapper for running EngineCore in background process
    in a data parallel context."""

    def __init__(
        self,
        vllm_config: VllmConfig,
1122
        local_client: bool,
1123
        handshake_address: str,
1124
1125
        executor_class: type[Executor],
        log_stats: bool,
1126
        client_handshake_address: str | None = None,
1127
    ):
1128
1129
        # Counts forward-passes of the model so that we can synchronize
        # finished with DP peers every N steps.
1130
        self.step_counter = 0
1131
        self.current_wave = 0
Rui Qiao's avatar
Rui Qiao committed
1132
        self.last_counts = (0, 0)
1133
1134
1135

        # Initialize the engine.
        dp_rank = vllm_config.parallel_config.data_parallel_rank
1136
1137
1138
1139
1140
1141
1142
1143
1144
        super().__init__(
            vllm_config,
            local_client,
            handshake_address,
            executor_class,
            log_stats,
            client_handshake_address,
            dp_rank,
        )
1145
1146
1147

    def _init_data_parallel(self, vllm_config: VllmConfig):
        # Configure GPUs and stateless process group for data parallel.
1148
        dp_rank = vllm_config.parallel_config.data_parallel_rank
1149
        dp_size = vllm_config.parallel_config.data_parallel_size
1150
1151
1152
        local_dp_rank = vllm_config.parallel_config.data_parallel_rank_local

        assert dp_size > 1
1153
        assert local_dp_rank is not None
1154
1155
        assert 0 <= local_dp_rank <= dp_rank < dp_size

1156
1157
1158
1159
1160
1161
        if 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.
            vllm_config.kv_transfer_config.engine_id = (
                f"{vllm_config.kv_transfer_config.engine_id}_dp{local_dp_rank}"
            )
1162
1163
1164
1165
            logger.debug(
                "Setting kv_transfer_config.engine_id to %s",
                vllm_config.kv_transfer_config.engine_id,
            )
1166

1167
        self.dp_rank = dp_rank
1168
1169
1170
1171
1172
1173
1174
        self.dp_group = vllm_config.parallel_config.stateless_init_dp_group()

    def shutdown(self):
        super().shutdown()
        if dp_group := getattr(self, "dp_group", None):
            stateless_destroy_torch_distributed_process_group(dp_group)

1175
1176
1177
1178
    def add_request(self, request: Request, request_wave: int = 0):
        if self.has_coordinator and request_wave != self.current_wave:
            if request_wave > self.current_wave:
                self.current_wave = request_wave
1179
1180
1181
1182
            elif not self.engines_running:
                # Request received for an already-completed wave, notify
                # front-end that we need to start the next one.
                self.output_queue.put_nowait(
1183
1184
                    (-1, EngineCoreOutputs(start_wave=self.current_wave))
                )
1185

1186
        super().add_request(request, request_wave)
1187

1188
1189
1190
    def _handle_client_request(
        self, request_type: EngineCoreRequestType, request: Any
    ) -> None:
1191
        if request_type == EngineCoreRequestType.START_DP_WAVE:
1192
1193
            new_wave, exclude_eng_index = request
            if exclude_eng_index != self.engine_index and (
1194
1195
                new_wave >= self.current_wave
            ):
1196
1197
                self.current_wave = new_wave
                if not self.engines_running:
1198
                    logger.debug("EngineCore starting idle loop for wave %d.", new_wave)
1199
1200
1201
1202
                    self.engines_running = True
        else:
            super()._handle_client_request(request_type, request)

1203
    def _maybe_publish_request_counts(self):
1204
        if not self.publish_dp_lb_stats:
1205
1206
1207
1208
1209
1210
            return

        # Publish our request counts (if they've changed).
        counts = self.scheduler.get_request_counts()
        if counts != self.last_counts:
            self.last_counts = counts
1211
1212
1213
1214
            stats = SchedulerStats(
                *counts, step_counter=self.step_counter, current_wave=self.current_wave
            )
            self.output_queue.put_nowait((-1, EngineCoreOutputs(scheduler_stats=stats)))
1215

1216
1217
1218
1219
1220
1221
1222
1223
    def run_busy_loop(self):
        """Core busy loop of the EngineCore for data parallel case."""

        # Loop until process is sent a SIGINT or SIGTERM
        while True:
            # 1) Poll the input queue until there is work to do.
            self._process_input_queue()

1224
1225
            # 2) Step the engine core.
            executed = self._process_engine_step()
1226
1227
            self._maybe_publish_request_counts()

1228
            local_unfinished_reqs = self.scheduler.has_unfinished_requests()
1229
1230
            if not executed:
                if not local_unfinished_reqs and not self.engines_running:
1231
1232
1233
                    # All engines are idle.
                    continue

1234
1235
                # We are in a running state and so must execute a dummy pass
                # if the model didn't execute any ready requests.
1236
1237
1238
                self.execute_dummy_batch()

            # 3) All-reduce operation to determine global unfinished reqs.
1239
            self.engines_running = self._has_global_unfinished_reqs(
1240
1241
                local_unfinished_reqs
            )
1242

1243
            if not self.engines_running:
1244
                if self.dp_rank == 0 or not self.has_coordinator:
1245
                    # Notify client that we are pausing the loop.
1246
1247
1248
                    logger.debug(
                        "Wave %d finished, pausing engine loop.", self.current_wave
                    )
1249
1250
1251
1252
                    # In the coordinator case, dp rank 0 sends updates to the
                    # coordinator. Otherwise (offline spmd case), each rank
                    # sends the update to its colocated front-end process.
                    client_index = -1 if self.has_coordinator else 0
1253
                    self.output_queue.put_nowait(
1254
1255
1256
1257
1258
                        (
                            client_index,
                            EngineCoreOutputs(wave_complete=self.current_wave),
                        )
                    )
1259
                # Increment wave count and reset step counter.
1260
                self.current_wave += 1
1261
                self.step_counter = 0
1262
1263

    def _has_global_unfinished_reqs(self, local_unfinished: bool) -> bool:
1264
        # Optimization - only perform finish-sync all-reduce every 32 steps.
1265
1266
        self.step_counter += 1
        if self.step_counter % 32 != 0:
1267
1268
            return True

1269
        return ParallelConfig.has_unfinished_dp(self.dp_group, local_unfinished)
Rui Qiao's avatar
Rui Qiao committed
1270

1271
    def reinitialize_distributed(
1272
1273
        self, reconfig_request: ReconfigureDistributedRequest
    ) -> None:
1274
1275
1276
1277
1278
        stateless_destroy_torch_distributed_process_group(self.dp_group)
        self.shutdown()

        parallel_config = self.vllm_config.parallel_config
        old_dp_size = parallel_config.data_parallel_size
1279
        parallel_config.data_parallel_size = reconfig_request.new_data_parallel_size
1280
        if reconfig_request.new_data_parallel_rank != -1:
1281
            parallel_config.data_parallel_rank = reconfig_request.new_data_parallel_rank
1282
        # local rank specifies device visibility, it should not be changed
1283
1284
1285
1286
1287
        assert (
            reconfig_request.new_data_parallel_rank_local
            == ReconfigureRankType.KEEP_CURRENT_RANK
        )
        parallel_config.data_parallel_master_ip = (
1288
            reconfig_request.new_data_parallel_master_ip
1289
1290
        )
        parallel_config.data_parallel_master_port = (
1291
            reconfig_request.new_data_parallel_master_port
1292
        )
1293
1294
1295
        if reconfig_request.new_data_parallel_rank != -2:
            self.dp_rank = parallel_config.data_parallel_rank
            self.dp_group = parallel_config.stateless_init_dp_group()
1296
        reconfig_request.new_data_parallel_master_port = (
1297
            parallel_config.data_parallel_master_port
1298
        )
1299
1300
1301
1302
1303
1304
1305
1306

        self.model_executor.reinitialize_distributed(reconfig_request)
        if reconfig_request.new_data_parallel_size > old_dp_size:
            assert self.available_gpu_memory_for_kv_cache > 0
            # pass available_gpu_memory_for_kv_cache from existing
            # engine-cores to new engine-cores so they can directly
            # use it in _initialize_kv_caches() rather than profiling.
            ParallelConfig.sync_kv_cache_memory_size(
1307
1308
                self.dp_group, self.available_gpu_memory_for_kv_cache
            )
1309
1310
1311
            # NOTE(yongji): newly joined workers require dummy_run even
            # CUDA graph is not used
            self.model_executor.collective_rpc("compile_or_warm_up_model")
1312
1313
1314
1315
        if (
            reconfig_request.new_data_parallel_rank
            == ReconfigureRankType.SHUTDOWN_CURRENT_RANK
        ):
1316
1317
1318
            self.shutdown()
            logger.info("DPEngineCoreProc %s shutdown", self.dp_rank)
        else:
1319
1320
1321
            logger.info(
                "Distributed environment reinitialized for DP rank %s", self.dp_rank
            )
1322

Rui Qiao's avatar
Rui Qiao committed
1323
1324
1325
1326
1327
1328
1329
1330
1331

class DPEngineCoreActor(DPEngineCoreProc):
    """
    Ray actor for running EngineCore in a data parallel context
    """

    def __init__(
        self,
        vllm_config: VllmConfig,
1332
        local_client: bool,
Rui Qiao's avatar
Rui Qiao committed
1333
1334
1335
1336
1337
1338
1339
1340
        addresses: EngineZmqAddresses,
        executor_class: type[Executor],
        log_stats: bool,
        dp_rank: int = 0,
        local_dp_rank: int = 0,
    ):
        self.addresses = addresses
        vllm_config.parallel_config.data_parallel_rank = dp_rank
1341
        vllm_config.parallel_config.data_parallel_rank_local = local_dp_rank
Rui Qiao's avatar
Rui Qiao committed
1342

1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
        # Set CUDA_VISIBLE_DEVICES as early as possible in actor life cycle
        # NOTE: in MP we set CUDA_VISIBLE_DEVICES at process creation time,
        # and this cannot be done in the same way for Ray because:
        # 1) Ray manages life cycle of all ray workers (including
        # DPEngineCoreActor)
        # 2) Ray sets CUDA_VISIBLE_DEVICES based on num_gpus configuration
        # To bypass 2, we need to also set
        # RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES, but vLLM workers created
        # thereafter would have CUDA_VISIBLE_DEVICES set, which is sticky:
        # https://github.com/ray-project/ray/blob/e752fc319ddedd9779a0989b6d3613909bad75c9/python/ray/_private/worker.py#L456 # noqa: E501
1353
1354
1355
1356
1357
1358
1359
        # This is problematic because when the vLLM worker (a Ray actor)
        # executes a task, it indexes into the sticky CUDA_VISIBLE_DEVICES
        # rather than directly using the GPU ID, potentially resulting in
        # index out of bounds error. See:
        # https://github.com/ray-project/ray/pull/40461/files#diff-31e8159767361e4bc259b6d9883d9c0d5e5db780fcea4a52ead4ee3ee4a59a78R1860 # noqa: E501
        # and get_accelerator_ids_for_accelerator_resource() in worker.py
        # of ray.
1360
        self._set_visible_devices(vllm_config, local_dp_rank)
Rui Qiao's avatar
Rui Qiao committed
1361

1362
        super().__init__(vllm_config, local_client, "", executor_class, log_stats)
Rui Qiao's avatar
Rui Qiao committed
1363

1364
    def _set_visible_devices(self, vllm_config: VllmConfig, local_dp_rank: int):
1365
        from vllm.platforms import current_platform
1366

1367
1368
1369
1370
        if current_platform.is_xpu():
            pass
        else:
            device_control_env_var = current_platform.device_control_env_var
1371
1372
1373
            self._set_cuda_visible_devices(
                vllm_config, local_dp_rank, device_control_env_var
            )
1374

1375
1376
1377
    def _set_cuda_visible_devices(
        self, vllm_config: VllmConfig, local_dp_rank: int, device_control_env_var: str
    ):
1378
1379
1380
        world_size = vllm_config.parallel_config.world_size
        # Set CUDA_VISIBLE_DEVICES or equivalent.
        try:
1381
1382
1383
            value = get_device_indices(
                device_control_env_var, local_dp_rank, world_size
            )
1384
            os.environ[device_control_env_var] = value
1385
1386
1387
1388
1389
        except IndexError as e:
            raise Exception(
                f"Error setting {device_control_env_var}: "
                f"local range: [{local_dp_rank * world_size}, "
                f"{(local_dp_rank + 1) * world_size}) "
1390
1391
                f'base value: "{os.getenv(device_control_env_var)}"'
            ) from e
1392

Rui Qiao's avatar
Rui Qiao committed
1393
    @contextmanager
1394
1395
1396
1397
1398
1399
    def _perform_handshakes(
        self,
        handshake_address: str,
        identity: bytes,
        local_client: bool,
        vllm_config: VllmConfig,
1400
        client_handshake_address: str | None,
1401
    ):
Rui Qiao's avatar
Rui Qiao committed
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
        """
        For Ray, we don't need to actually perform handshake.
        All addresses information is known before the actor creation.
        Therefore, we simply yield these addresses.
        """
        yield self.addresses

    def wait_for_init(self):
        """
        Wait until the engine core is initialized.

        This is just an empty method. When ray.get() on this method
        (or any other method of the actor) returns, it is guaranteed
        that actor creation (i.e., __init__) is complete.
        """
        pass

    def run(self):
        """
        Run the engine core busy loop.
        """
        try:
            self.run_busy_loop()
        except SystemExit:
            logger.debug("EngineCore exiting.")
            raise
        except Exception:
            logger.exception("EngineCore encountered a fatal error.")
            raise
        finally:
            self.shutdown()