core.py 81.4 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
import os
4
import queue
5
import signal
6
7
import threading
import time
8
from collections import defaultdict, deque
9
from collections.abc import Callable, Generator
10
from concurrent.futures import Future
Rui Qiao's avatar
Rui Qiao committed
11
from contextlib import ExitStack, contextmanager
12
from enum import IntEnum
13
from functools import partial
14
from inspect import isclass, signature
15
from logging import DEBUG
16
from multiprocessing.queues import Queue
17
from typing import Any, TypeVar, cast
18

19
import msgspec
20
21
import zmq

22
import vllm.envs as envs
23
24
from vllm.config import ParallelConfig, VllmConfig
from vllm.distributed import stateless_destroy_torch_distributed_process_group
25
from vllm.envs import enable_envs_cache
26
from vllm.logger import init_logger
27
from vllm.logging_utils.dump_input import dump_engine_exception
28
from vllm.lora.request import LoRARequest
29
from vllm.multimodal import MULTIMODAL_REGISTRY
30
from vllm.tasks import POOLING_TASKS, SupportedTask
31
from vllm.tracing import instrument, maybe_init_worker_tracer
32
from vllm.transformers_utils.config import maybe_register_config_serialize_by_value
33
from vllm.utils import numa_utils
34
35
36
37
from vllm.utils.gc_utils import (
    freeze_gc_heap,
    maybe_attach_gc_debug_callback,
)
38
from vllm.utils.hashing import get_hash_fn_by_name
39
from vllm.utils.network_utils import make_zmq_socket
40
from vllm.utils.system_utils import decorate_logs, set_process_title
41
42
43
44
45
46
47
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,
)
48
from vllm.v1.core.sched.interface import PauseState, SchedulerInterface
49
from vllm.v1.core.sched.output import SchedulerOutput
50
from vllm.v1.engine import (
51
52
    EEP_NOTIFICATION_CALL_ID,
    EEPNotificationType,
53
    EngineCoreOutput,
54
    EngineCoreOutputs,
55
    EngineCoreReadyResponse,
56
57
    EngineCoreRequest,
    EngineCoreRequestType,
58
    FinishReason,
59
    PauseMode,
60
61
62
63
64
    ReconfigureDistributedRequest,
    ReconfigureRankType,
    UtilityOutput,
    UtilityResult,
)
65
from vllm.v1.engine.tensor_ipc import TensorIpcReceiver
66
67
68
from vllm.v1.engine.utils import (
    EngineHandshakeMetadata,
    EngineZmqAddresses,
69
    SignalCallback,
70
71
    get_device_indices,
)
72
from vllm.v1.executor import Executor
73
from vllm.v1.kv_cache_interface import KVCacheConfig
74
from vllm.v1.metrics.stats import SchedulerStats
75
from vllm.v1.outputs import ModelRunnerOutput
76
from vllm.v1.request import Request, RequestStatus
77
from vllm.v1.serial_utils import MsgpackDecoder, MsgpackEncoder
78
from vllm.v1.structured_output import StructuredOutputManager
79
from vllm.v1.utils import compute_iteration_details
80
81
82
83
from vllm.version import __version__ as VLLM_VERSION

logger = init_logger(__name__)

84
HANDSHAKE_TIMEOUT_MINS = 5
85

86
_R = TypeVar("_R")  # Return type for collective_rpc
87

88
89
90
91

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

92
93
94
95
96
    def __init__(
        self,
        vllm_config: VllmConfig,
        executor_class: type[Executor],
        log_stats: bool,
97
        executor_fail_callback: Callable | None = None,
98
        include_finished_set: bool = False,
99
    ):
100
101
        # plugins need to be loaded at the engine/scheduler level too
        from vllm.plugins import load_general_plugins
102

103
104
        load_general_plugins()

105
        self.vllm_config = vllm_config
106
        if not vllm_config.parallel_config.data_parallel_rank_local:
107
108
109
110
111
            logger.info(
                "Initializing a V1 LLM engine (v%s) with config: %s",
                VLLM_VERSION,
                vllm_config,
            )
112

113
114
        self.log_stats = log_stats

115
116
        # Setup Model.
        self.model_executor = executor_class(vllm_config)
117
        if executor_fail_callback is not None:
118
            self.model_executor.register_failure_callback(executor_fail_callback)
119

120
121
        self.available_gpu_memory_for_kv_cache = -1

122
123
124
        if envs.VLLM_ELASTIC_EP_SCALE_UP_LAUNCH:
            self._eep_scale_up_before_kv_init()

125
        # Setup KV Caches and update CacheConfig after profiling.
126
        kv_cache_config = self._initialize_kv_caches(vllm_config)
127
128
        self.structured_output_manager = StructuredOutputManager(vllm_config)

129
        # Setup scheduler.
130
        Scheduler = vllm_config.scheduler_config.get_scheduler_cls()
131

132
        if len(kv_cache_config.kv_cache_groups) == 0:  # noqa: SIM102
133
134
            # Encoder models without KV cache don't support
            # chunked prefill. But do SSM models?
135
136
137
            if vllm_config.scheduler_config.enable_chunked_prefill:
                logger.warning("Disabling chunked prefill for model without KVCache")
                vllm_config.scheduler_config.enable_chunked_prefill = False
138

139
140
141
        scheduler_block_size = (
            vllm_config.cache_config.block_size
            * vllm_config.parallel_config.decode_context_parallel_size
142
            * vllm_config.parallel_config.prefill_context_parallel_size
143
144
        )

145
        self.scheduler: SchedulerInterface = Scheduler(
146
            vllm_config=vllm_config,
147
148
            kv_cache_config=kv_cache_config,
            structured_output_manager=self.structured_output_manager,
149
            include_finished_set=include_finished_set,
150
            log_stats=self.log_stats,
151
            block_size=scheduler_block_size,
152
        )
153
        self.use_spec_decode = vllm_config.speculative_config is not None
154
        if self.scheduler.connector is not None:  # type: ignore
155
            self.model_executor.init_kv_output_aggregator(self.scheduler.connector)  # type: ignore
156

157
        mm_registry = MULTIMODAL_REGISTRY
158
159
        self.mm_receiver_cache = mm_registry.engine_receiver_cache_from_config(
            vllm_config
160
        )
161

162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
        # 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)

183
184
185
186
187
        # 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
188
        self.batch_queue: (
189
            deque[tuple[Future[ModelRunnerOutput], SchedulerOutput, Future[Any]]] | None
190
        ) = None
191
        if self.batch_queue_size > 1:
192
            logger.debug("Batch queue is enabled with size %d", self.batch_queue_size)
193
            self.batch_queue = deque(maxlen=self.batch_queue_size)
194

195
196
197
        self.is_ec_consumer = (
            vllm_config.ec_transfer_config is None
            or vllm_config.ec_transfer_config.is_ec_consumer
198
        )
199
        self.is_pooling_model = vllm_config.model_config.runner_type == "pooling"
200

201
        self.request_block_hasher: Callable[[Request], list[BlockHash]] | None = None
202
        if vllm_config.cache_config.enable_prefix_caching or kv_connector is not None:
203
            caching_hash_fn = get_hash_fn_by_name(
204
205
                vllm_config.cache_config.prefix_caching_hash_algo
            )
206
207
208
            init_none_hash(caching_hash_fn)

            self.request_block_hasher = get_request_block_hasher(
209
                scheduler_block_size, caching_hash_fn
210
            )
211

212
213
214
        self.step_fn = (
            self.step if self.batch_queue is None else self.step_with_batch_queue
        )
215
        self.async_scheduling = vllm_config.scheduler_config.async_scheduling
216

217
        self.aborts_queue = queue.Queue[list[str]]()
218

219
        self._idle_state_callbacks: list[Callable] = []
220

221
222
223
        # Mark the startup heap as static so that it's ignored by GC.
        # Reduces pause times of oldest generation collections.
        freeze_gc_heap()
224
225
        # If enable, attach GC debugger after static variable freeze.
        maybe_attach_gc_debug_callback()
226
227
228
        # Enable environment variable cache (e.g. assume no more
        # environment variable overrides after this point)
        enable_envs_cache()
229

230
    @instrument(span_name="Prepare model")
231
    def _initialize_kv_caches(self, vllm_config: VllmConfig) -> KVCacheConfig:
232
        start = time.time()
233

234
        # Get all kv cache needed by the model
235
        kv_cache_specs = self.model_executor.get_kv_cache_specs()
236

237
238
        has_kv_cache = any(kv_cache_spec for kv_cache_spec in kv_cache_specs)
        if has_kv_cache:
239
240
241
242
            if envs.VLLM_ELASTIC_EP_SCALE_UP_LAUNCH:
                # NOTE(yongji): should already be set
                # during _eep_scale_up_before_kv_init
                assert self.available_gpu_memory_for_kv_cache > 0
243
244
245
                available_gpu_memory = [self.available_gpu_memory_for_kv_cache] * len(
                    kv_cache_specs
                )
246
247
248
            else:
                # Profiles the peak memory usage of the model to determine how
                # much memory can be allocated for kv cache.
249
250
                available_gpu_memory = self.model_executor.determine_available_memory()
                self.available_gpu_memory_for_kv_cache = available_gpu_memory[0]
251
252
253
        else:
            # Attention free models don't need memory for kv cache
            available_gpu_memory = [0] * len(kv_cache_specs)
254

255
        assert len(kv_cache_specs) == len(available_gpu_memory)
256

257
258
259
        # Track max_model_len before KV cache config to detect auto-fit changes
        max_model_len_before = vllm_config.model_config.max_model_len

260
261
262
        kv_cache_configs = get_kv_cache_configs(
            vllm_config, kv_cache_specs, available_gpu_memory
        )
263
264
265
266
267
268
269
270

        # If auto-fit reduced max_model_len, sync the new value to workers.
        # This is needed because workers were spawned before memory profiling
        # and have the original (larger) max_model_len cached.
        max_model_len_after = vllm_config.model_config.max_model_len
        if max_model_len_after != max_model_len_before:
            self.collective_rpc("update_max_model_len", args=(max_model_len_after,))

271
        scheduler_kv_cache_config = generate_scheduler_kv_cache_config(kv_cache_configs)
272
273
274
275
276
277
278
279
        vllm_config.cache_config.num_gpu_blocks = scheduler_kv_cache_config.num_blocks
        kv_cache_groups = scheduler_kv_cache_config.kv_cache_groups
        if kv_cache_groups:
            vllm_config.cache_config.block_size = min(
                g.kv_cache_spec.block_size for g in kv_cache_groups
            )

        vllm_config.validate_block_size()
280
281

        # Initialize kv cache and warmup the execution
282
        self.model_executor.initialize_from_config(kv_cache_configs)
283

284
        elapsed = time.time() - start
285
        logger.info_once(
286
            "init engine (profile, create kv cache, warmup model) took %.2f seconds",
287
            elapsed,
288
            scope="local",
289
        )
290
        return scheduler_kv_cache_config
291

292
293
294
    def get_supported_tasks(self) -> tuple[SupportedTask, ...]:
        return self.model_executor.supported_tasks

295
296
    def add_request(self, request: Request, request_wave: int = 0):
        """Add request to the scheduler.
297

298
299
300
        `request_wave`: indicate which wave of requests this is expected to
        belong to in DP case
        """
301
302
303
        # Validate the request_id type.
        if not isinstance(request.request_id, str):
            raise TypeError(
304
305
                f"request_id must be a string, got {type(request.request_id)}"
            )
306

307
        if pooling_params := request.pooling_params:
308
            supported_pooling_tasks = [
309
                task for task in self.get_supported_tasks() if task in POOLING_TASKS
310
311
            ]

312
            if pooling_params.task not in supported_pooling_tasks:
313
314
315
316
                raise ValueError(
                    f"Unsupported task: {pooling_params.task!r} "
                    f"Supported tasks: {supported_pooling_tasks}"
                )
317

318
        if request.kv_transfer_params is not None and (
319
320
321
322
323
324
            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
325

326
        self.scheduler.add_request(request)
327

328
    def abort_requests(self, request_ids: list[str]):
329
330
331
332
333
        """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).
334
        self.scheduler.finish_requests(request_ids, RequestStatus.FINISHED_ABORTED)
335

336
337
    @contextmanager
    def log_error_detail(self, scheduler_output: SchedulerOutput):
338
        """Execute the model and log detailed info on failure."""
339
        try:
340
            yield
341
342
343
344
345
        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.

346
            # NOTE: This method is exception-free
347
348
349
            dump_engine_exception(
                self.vllm_config, scheduler_output, self.scheduler.make_stats()
            )
350
351
            raise err

352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
    @contextmanager
    def log_iteration_details(self, scheduler_output: SchedulerOutput):
        if not self.vllm_config.observability_config.enable_logging_iteration_details:
            yield
            return
        self._iteration_index = getattr(self, "_iteration_index", 0)
        iteration_details = compute_iteration_details(scheduler_output)
        before = time.monotonic()
        yield
        logger.info(
            "".join(
                [
                    "Iteration(",
                    str(self._iteration_index),
                    "): ",
                    str(iteration_details.num_ctx_requests),
                    " context requests, ",
                    str(iteration_details.num_ctx_tokens),
                    " context tokens, ",
                    str(iteration_details.num_generation_requests),
                    " generation requests, ",
                    str(iteration_details.num_generation_tokens),
                    " generation tokens, iteration elapsed time: ",
                    format((time.monotonic() - before) * 1000, ".2f"),
                    " ms",
                ]
            )
        )
        self._iteration_index += 1

382
    def step(self) -> tuple[dict[int, EngineCoreOutputs], bool]:
383
384
385
386
387
        """Schedule, execute, and make output.

        Returns tuple of outputs and a flag indicating whether the model
        was executed.
        """
388

389
390
391
        # Check for any requests remaining in the scheduler - unfinished,
        # or finished and not yet removed from the batch.
        if not self.scheduler.has_requests():
392
            return {}, False
393
394
395
        scheduler_output = self.scheduler.schedule()
        future = self.model_executor.execute_model(scheduler_output, non_block=True)
        grammar_output = self.scheduler.get_grammar_bitmask(scheduler_output)
396
397
398
399
        with (
            self.log_error_detail(scheduler_output),
            self.log_iteration_details(scheduler_output),
        ):
400
401
402
403
            model_output = future.result()
            if model_output is None:
                model_output = self.model_executor.sample_tokens(grammar_output)

404
405
406
        # Before processing the model output, process any aborts that happened
        # during the model execution.
        self._process_aborts_queue()
407
408
409
        engine_core_outputs = self.scheduler.update_from_output(
            scheduler_output, model_output
        )
410

411
        return engine_core_outputs, scheduler_output.total_num_scheduled_tokens > 0
412

413
    def post_step(self, model_executed: bool) -> None:
414
415
416
417
        # When using async scheduling we can't get draft token ids in advance,
        # so we update draft token ids in the worker process and don't
        # need to update draft token ids here.
        if not self.async_scheduling and self.use_spec_decode and model_executed:
418
419
420
421
422
            # 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)

423
    def step_with_batch_queue(
424
        self,
425
    ) -> tuple[dict[int, EngineCoreOutputs] | None, bool]:
426
427
428
429
        """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:
430
431
432
433
        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.
434
435
436
437
438
        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.
        """
439

440
441
        batch_queue = self.batch_queue
        assert batch_queue is not None
442

443
444
445
        # 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.
446
        assert len(batch_queue) < self.batch_queue_size
447

448
        model_executed = False
449
        deferred_scheduler_output = None
450
        if self.scheduler.has_requests():
451
            scheduler_output = self.scheduler.schedule()
452
453
454
455
            with self.log_error_detail(scheduler_output):
                exec_future = self.model_executor.execute_model(
                    scheduler_output, non_block=True
                )
456
            if self.is_ec_consumer:
457
                model_executed = scheduler_output.total_num_scheduled_tokens > 0
458

459
            if self.is_pooling_model or not model_executed:
460
461
                # No sampling required (no requests scheduled).
                future = cast(Future[ModelRunnerOutput], exec_future)
462
            else:
463
464
465
                if not scheduler_output.pending_structured_output_tokens:
                    # We aren't waiting for any tokens, get any grammar output
                    # and sample immediately.
466
467
468
                    grammar_output = self.scheduler.get_grammar_bitmask(
                        scheduler_output
                    )
469
470
471
                    future = self.model_executor.sample_tokens(
                        grammar_output, non_block=True
                    )
472
                else:
473
474
475
476
477
                    # We need to defer sampling until we have processed the model output
                    # from the prior step.
                    deferred_scheduler_output = scheduler_output

            if not deferred_scheduler_output:
478
                # Add this step's future to the queue.
479
                batch_queue.appendleft((future, scheduler_output, exec_future))
480
481
482
483
484
485
486
487
                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
488
489
490
491
492
493

        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
494
495

        # Block until the next result is available.
496
        future, scheduler_output, exec_model_fut = batch_queue.pop()
497
498
499
500
        with (
            self.log_error_detail(scheduler_output),
            self.log_iteration_details(scheduler_output),
        ):
501
            model_output = future.result()
502
503
504
505
506
            if model_output is None:
                # None from sample_tokens() implies that the original execute_model()
                # call failed - raise that exception.
                exec_model_fut.result()
                raise RuntimeError("unexpected error")
507

508
509
510
        # Before processing the model output, process any aborts that happened
        # during the model execution.
        self._process_aborts_queue()
511
512
513
        engine_core_outputs = self.scheduler.update_from_output(
            scheduler_output, model_output
        )
514
515
516
517
518

        # 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:
519
520
521
522
523
524
525
526
527
528
529
530
            # If we are doing speculative decoding with structured output,
            # we need to get the draft token ids from the prior step before
            # we can compute the grammar bitmask for the deferred request.
            if self.use_spec_decode:
                draft_token_ids = self.model_executor.take_draft_token_ids()
                assert draft_token_ids is not None
                # Update the draft token ids in the scheduler output to
                # filter out the invalid spec tokens, which will be padded
                # with -1 and skipped by the grammar bitmask computation.
                self.scheduler.update_draft_token_ids_in_output(
                    draft_token_ids, deferred_scheduler_output
                )
531
532
533
534
535
536
            # 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)
537
            batch_queue.appendleft((future, deferred_scheduler_output, exec_future))
538

539
        return engine_core_outputs, model_executed
540

541
542
543
544
545
    def _process_aborts_queue(self):
        if not self.aborts_queue.empty():
            request_ids = []
            while not self.aborts_queue.empty():
                ids = self.aborts_queue.get_nowait()
546
547
                # Should be a list here, but also handle string just in case.
                request_ids.extend((ids,) if isinstance(ids, str) else ids)
548
549
550
            # More efficient to abort all as a single batch.
            self.abort_requests(request_ids)

551
    def shutdown(self):
552
        self.structured_output_manager.clear_backend()
553
554
        if self.model_executor:
            self.model_executor.shutdown()
555
556
        if self.scheduler:
            self.scheduler.shutdown()
557

558
559
    def profile(self, is_start: bool = True, profile_prefix: str | None = None):
        self.model_executor.profile(is_start, profile_prefix)
560

561
562
    def reset_mm_cache(self):
        # NOTE: Since this is mainly for debugging, we don't attempt to
563
        # re-sync the internal caches (P0 sender, P1 receiver)
564
        if self.scheduler.has_unfinished_requests():
565
566
567
568
            logger.warning(
                "Resetting the multi-modal cache when requests are "
                "in progress may lead to desynced internal caches."
            )
569

570
        # The cache either exists in EngineCore or WorkerWrapperBase
571
572
        if self.mm_receiver_cache is not None:
            self.mm_receiver_cache.clear_cache()
573

574
575
        self.model_executor.reset_mm_cache()

576
577
578
579
580
581
    def reset_prefix_cache(
        self, reset_running_requests: bool = False, reset_connector: bool = False
    ) -> bool:
        return self.scheduler.reset_prefix_cache(
            reset_running_requests, reset_connector
        )
582

583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
    def reset_encoder_cache(self) -> None:
        """Reset the encoder cache to invalidate all cached encoder outputs.

        This should be called when model weights are updated to ensure
        stale vision embeddings computed with old weights are not reused.
        Clears both the scheduler's cache manager and the GPU model runner's cache.
        """
        # NOTE: Since this is mainly for debugging, we don't attempt to
        # re-sync the internal caches (P0 sender, P1 receiver)
        if self.scheduler.has_unfinished_requests():
            logger.warning(
                "Resetting the encoder cache when requests are "
                "in progress may lead to desynced internal caches."
            )

        # Reset the scheduler's encoder cache manager (logical state)
        self.scheduler.reset_encoder_cache()
        # Reset the GPU model runner's encoder cache (physical storage)
        self.model_executor.reset_encoder_cache()

603
604
605
606
607
    def _reset_caches(self, reset_running_requests=True) -> None:
        self.reset_prefix_cache(reset_running_requests=reset_running_requests)
        self.reset_mm_cache()
        self.reset_encoder_cache()

608
609
    def pause_scheduler(
        self, mode: PauseMode = "abort", clear_cache: bool = True
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
    ) -> Future | None:
        """Pause generation; behavior depends on mode.

        All pause modes queue new adds -- "abort" and "keep" skip step();
        "wait" allows step() so in-flight requests can drain.

        - ``abort``: Set PAUSED_NEW, abort all requests, wait for abort
          outputs to be sent (when running with output_queue), optionally
          clear caches, then complete the returned Future.
        - ``wait``: Set PAUSED_NEW (queue adds, keep stepping); when drained,
          optionally clear caches, then complete the returned Future.
        - ``keep``: Set PAUSED_ALL; return a Future that completes when the
          output queue is empty.
        """
        if mode not in ("keep", "abort", "wait"):
            raise ValueError(f"Invalid pause mode: {mode}")
        if mode == "wait":
            raise ValueError("'wait' mode can't be used in inproc-engine mode")

        if mode == "abort":
            self.scheduler.finish_requests(None, RequestStatus.FINISHED_ABORTED)

        pause_state = PauseState.PAUSED_ALL if mode == "keep" else PauseState.PAUSED_NEW
        self.scheduler.set_pause_state(pause_state)
        if clear_cache:
            self._reset_caches()

637
638
639
        return None

    def resume_scheduler(self) -> None:
640
641
        """Resume the scheduler and flush any requests queued while paused."""
        self.scheduler.set_pause_state(PauseState.UNPAUSED)
642
643

    def is_scheduler_paused(self) -> bool:
644
645
        """Return whether the scheduler is in any pause state."""
        return self.scheduler.pause_state != PauseState.UNPAUSED
646

647
    def sleep(self, level: int = 1, mode: PauseMode = "abort") -> None | Future:
648
649
650
651
652
653
654
655
        """Put the engine to sleep at the specified level.

        Args:
            level: Sleep level.
                - Level 0: Pause scheduling only. Requests are still accepted
                           but not processed. No GPU memory changes.
                - Level 1: Offload model weights to CPU, discard KV cache.
                - Level 2: Discard all GPU memory.
656
657
            mode: Pause mode - how to deal with any existing requests, see
                documentation of pause_scheduler method.
658
        """
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683

        # Pause scheduler before sleeping.
        clear_prefix_cache = level >= 1
        pause_future = self.pause_scheduler(mode=mode, clear_cache=clear_prefix_cache)
        if level < 1:
            return pause_future

        # Level 1+: Delegate to executor for GPU memory management
        model_executor = self.model_executor
        if pause_future is None:
            model_executor.sleep(level)
            return None

        future = Future[Any]()

        def pause_complete(f: Future):
            try:
                f.result()  # propagate any exception
                future.set_result(model_executor.sleep(level))
            except Exception as e:
                future.set_exception(e)

        logger.info("Waiting for in-flight requests to complete before sleeping...")
        pause_future.add_done_callback(pause_complete)
        return future
684

685
    def wake_up(self, tags: list[str] | None = None):
686
687
688
689
690
691
        """Wake up the engine from sleep.

        Args:
            tags: Tags to wake up. Use ["scheduling"] for level 0 wake up.
        """
        if tags is not None and "scheduling" in tags:
692
693
694
695
            # Remove "scheduling" from tags if there are other tags to process.
            tags = [t for t in tags if t != "scheduling"]

        if tags is None or tags:
696
            self.model_executor.wake_up(tags)
697

698
699
700
        # Resume scheduling (applies to all levels)
        self.resume_scheduler()

701
    def is_sleeping(self) -> bool:
702
        """Check if engine is sleeping at any level."""
703
        return self.is_scheduler_paused() or self.model_executor.is_sleeping
704

705
    def execute_dummy_batch(self):
706
        self.model_executor.execute_dummy_batch()
707

708
709
710
711
712
713
    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)

714
    def list_loras(self) -> set[int]:
715
716
717
718
        return self.model_executor.list_loras()

    def pin_lora(self, lora_id: int) -> bool:
        return self.model_executor.pin_lora(lora_id)
719

720
721
722
    def save_sharded_state(
        self,
        path: str,
723
724
        pattern: str | None = None,
        max_size: int | None = None,
725
    ) -> None:
726
727
728
729
730
731
        self.model_executor.save_sharded_state(
            path=path, pattern=pattern, max_size=max_size
        )

    def collective_rpc(
        self,
732
733
        method: str | Callable[..., _R],
        timeout: float | None = None,
734
        args: tuple = (),
735
        kwargs: dict[str, Any] | None = None,
736
737
    ) -> list[_R]:
        return self.model_executor.collective_rpc(method, timeout, args, kwargs)
738

739
    def preprocess_add_request(self, request: EngineCoreRequest) -> tuple[Request, int]:
740
        """Preprocess the request.
741

742
743
744
        This function could be directly used in input processing thread to allow
        request initialization running in parallel with Model forward
        """
745
746
        # Note on thread safety: no race condition.
        # `mm_receiver_cache` is reset at the end of LLMEngine init,
747
        # and will only be accessed in the input processing thread afterwards.
748
        if self.mm_receiver_cache is not None and request.mm_features:
749
750
751
            request.mm_features = self.mm_receiver_cache.get_and_update_features(
                request.mm_features
            )
752

753
        req = Request.from_engine_core_request(request, self.request_block_hasher)
754
755
756
757
758
759
760
761
762
        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

763
764
765
766
767
768
769
770
771
772
    def _eep_scale_up_before_kv_init(self):
        raise NotImplementedError

    def _eep_send_engine_core_notification(
        self,
        notification_type: EEPNotificationType,
        vllm_config: VllmConfig | None = None,
    ):
        raise NotImplementedError

773

774
775
776
777
778
779
class EngineShutdownState(IntEnum):
    RUNNING = 0
    REQUESTED = 1
    SHUTTING_DOWN = 2


780
781
782
class EngineCoreProc(EngineCore):
    """ZMQ-wrapper for running EngineCore in background process."""

783
    ENGINE_CORE_DEAD = b"ENGINE_CORE_DEAD"
784
    addresses: EngineZmqAddresses
785

786
    @instrument(span_name="EngineCoreProc init")
787
788
    def __init__(
        self,
789
        vllm_config: VllmConfig,
790
        local_client: bool,
791
        handshake_address: str,
792
        executor_class: type[Executor],
793
        log_stats: bool,
794
        client_handshake_address: str | None = None,
795
        tensor_queue: Queue | None = None,
796
        *,
797
        engine_index: int = 0,
798
    ):
Rui Qiao's avatar
Rui Qiao committed
799
        self.input_queue = queue.Queue[tuple[EngineCoreRequestType, Any]]()
800
        self.output_queue = queue.Queue[tuple[int, EngineCoreOutputs] | bytes]()
Rui Qiao's avatar
Rui Qiao committed
801
        executor_fail_callback = lambda: self.input_queue.put_nowait(
802
803
            (EngineCoreRequestType.EXECUTOR_FAILED, b"")
        )
804

Rui Qiao's avatar
Rui Qiao committed
805
806
807
        self.engine_index = engine_index
        identity = self.engine_index.to_bytes(length=2, byteorder="little")
        self.engines_running = False
808
        self.shutdown_state = EngineShutdownState.RUNNING
809

810
811
812
813
814
815
        # Receiver for tensor IPC
        self.tensor_ipc_receiver: TensorIpcReceiver | None = None
        if tensor_queue is not None:
            self.tensor_ipc_receiver = TensorIpcReceiver(tensor_queue)
            logger.info("Using tensor IPC queue for multimodal tensor sharing")

816
817
818
819
820
821
822
        with self._perform_handshakes(
            handshake_address,
            identity,
            local_client,
            vllm_config,
            client_handshake_address,
        ) as addresses:
823
            # Set up data parallel environment.
824
            self.has_coordinator = addresses.coordinator_output is not None
825
            self.frontend_stats_publish_address = (
826
827
828
829
830
831
832
                addresses.frontend_stats_publish_address
            )
            logger.debug(
                "Has DP Coordinator: %s, stats publish address: %s",
                self.has_coordinator,
                self.frontend_stats_publish_address,
            )
833
            internal_dp_balancing = (
834
                self.has_coordinator
835
836
                and not vllm_config.parallel_config.data_parallel_external_lb
            )
837
838
839
            # Only publish request queue stats to coordinator for "internal"
            # and "hybrid" LB modes.
            self.publish_dp_lb_stats = internal_dp_balancing
840

841
842
843
844
845
846
847
            self.addresses = addresses
            self.process_input_queue_block = True
            if envs.VLLM_ELASTIC_EP_SCALE_UP_LAUNCH:
                self._eep_send_engine_core_notification(
                    EEPNotificationType.NEW_CORE_ENGINES_INIT_READY,
                    vllm_config=vllm_config,
                )
848
849
            self._init_data_parallel(vllm_config)

850
            super().__init__(
851
852
853
854
855
                vllm_config,
                executor_class,
                log_stats,
                executor_fail_callback,
                internal_dp_balancing,
856
            )
857

858
859
860
861
862
863
            # 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()
864
865
866
867
868
869
870
871
872
873
            input_thread = threading.Thread(
                target=self.process_input_sockets,
                args=(
                    addresses.inputs,
                    addresses.coordinator_input,
                    identity,
                    ready_event,
                ),
                daemon=True,
            )
874
875
876
877
            input_thread.start()

            self.output_thread = threading.Thread(
                target=self.process_output_sockets,
878
879
880
881
882
883
884
                args=(
                    addresses.outputs,
                    addresses.coordinator_output,
                    self.engine_index,
                ),
                daemon=True,
            )
885
886
887
888
889
890
            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():
891
                    raise RuntimeError("Input socket thread died during startup")
892
893
894
                assert addresses.coordinator_input is not None
                logger.info("Waiting for READY message from DP Coordinator...")

Rui Qiao's avatar
Rui Qiao committed
895
    @contextmanager
896
897
898
899
900
901
    def _perform_handshakes(
        self,
        handshake_address: str,
        identity: bytes,
        local_client: bool,
        vllm_config: VllmConfig,
902
        client_handshake_address: str | None,
Rui Qiao's avatar
Rui Qiao committed
903
    ) -> Generator[EngineZmqAddresses, None, None]:
904
905
906
907
908
        """
        Perform startup handshakes.

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

909
        For DP>1 with internal load-balancing this is with the shared front-end
910
911
        process which may reside on a different node.

912
        For DP>1 with external or hybrid load-balancing, two handshakes are
913
        performed:
914
915
916
917
            - 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.
918
919
        with the exception of the rank 0 and colocated engines themselves which
        don't require the second handshake.
920
921
922
923
924
925

        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
926
        input_ctx = zmq.Context()
927
        is_local = local_client and client_handshake_address is None
928
        headless = not local_client
929
930
931
932
933
934
935
936
937
        handshake = self._perform_handshake(
            input_ctx,
            handshake_address,
            identity,
            is_local,
            headless,
            vllm_config,
            vllm_config.parallel_config,
        )
938
        if client_handshake_address is None:
939
            # We only need to handshake with one party.
940
941
942
            with handshake as addresses:
                yield addresses
        else:
943
            # We need to handshake with rank 0 front-end and our colocated frontend.
944
            assert local_client
945
            local_handshake = self._perform_handshake(
946
947
                input_ctx, client_handshake_address, identity, True, False, vllm_config
            )
948
            with handshake as addresses, local_handshake as client_addresses:
949
950
951
952
                # 1. Obtain DP Coordinator zmq address and DP process group address
                #    (addresses).
                # 2. Add front-end input/output addresses from colocated front-end
                #    (client_addresses).
953
954
955
956
957
958
959
960
961
962
963
964
965
966
                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,
967
        headless: bool,
968
        vllm_config: VllmConfig,
969
        parallel_config_to_update: ParallelConfig | None = None,
970
    ) -> Generator[EngineZmqAddresses, None, None]:
971
972
973
974
975
976
977
978
        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
979
            # Register engine with front-end.
980
981
982
            addresses = self.startup_handshake(
                handshake_socket, local_client, headless, parallel_config_to_update
            )
983
984
985
986
987
988
989
990
            yield addresses

            # Send ready message.
            ready_msg = {
                "status": "READY",
                "local": local_client,
                "headless": headless,
            }
991
            # Include config hash for DP configuration validation
992
993
994
995
996
997
            if vllm_config.parallel_config.data_parallel_size > 1:
                ready_msg["parallel_config_hash"] = (
                    vllm_config.parallel_config.compute_hash()
                )

            handshake_socket.send(msgspec.msgpack.encode(ready_msg))
Rui Qiao's avatar
Rui Qiao committed
998

999
    @staticmethod
1000
    def startup_handshake(
1001
1002
        handshake_socket: zmq.Socket,
        local_client: bool,
1003
        headless: bool,
1004
        parallel_config: ParallelConfig | None = None,
1005
    ) -> EngineZmqAddresses:
1006
        # Send registration message.
1007
        handshake_socket.send(
1008
1009
1010
1011
1012
1013
1014
1015
            msgspec.msgpack.encode(
                {
                    "status": "HELLO",
                    "local": local_client,
                    "headless": headless,
                }
            )
        )
1016
1017

        # Receive initialization message.
1018
        logger.debug("Waiting for init message from front-end.")
1019
        if not handshake_socket.poll(timeout=HANDSHAKE_TIMEOUT_MINS * 60_000):
1020
1021
1022
1023
1024
            raise RuntimeError(
                "Did not receive response from front-end "
                f"process within {HANDSHAKE_TIMEOUT_MINS} "
                f"minutes"
            )
1025
1026
        init_bytes = handshake_socket.recv()
        init_message: EngineHandshakeMetadata = msgspec.msgpack.decode(
1027
1028
            init_bytes, type=EngineHandshakeMetadata
        )
1029
1030
        logger.debug("Received init message: %s", init_message)

1031
1032
1033
        if parallel_config is not None:
            for key, value in init_message.parallel_config.items():
                setattr(parallel_config, key, value)
1034

1035
        return init_message.addresses
1036
1037

    @staticmethod
1038
    def run_engine_core(*args, dp_rank: int = 0, local_dp_rank: int = 0, **kwargs):
1039
1040
        """Launch EngineCore busy loop in background process."""

1041
1042
1043
        # Ensure we can serialize transformer config after spawning
        maybe_register_config_serialize_by_value()

1044
        engine_core: EngineCoreProc | None = None
1045
        signal_callback: SignalCallback | None = None
1046
        try:
1047
1048
1049
1050
1051
            vllm_config: VllmConfig = kwargs["vllm_config"]
            parallel_config: ParallelConfig = vllm_config.parallel_config
            data_parallel = parallel_config.data_parallel_size > 1 or dp_rank > 0
            if data_parallel:
                parallel_config.data_parallel_rank_local = local_dp_rank
1052
                process_title = f"EngineCore_DP{dp_rank}"
1053
            else:
1054
1055
1056
                process_title = "EngineCore"
            set_process_title(process_title)
            maybe_init_worker_tracer("vllm.engine_core", "engine_core", process_title)
1057
            decorate_logs()
1058
1059
            if parallel_config.numa_bind:
                numa_utils.log_current_affinity_state(process_title)
1060

1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
            if data_parallel and 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}"
                )
                logger.debug(
                    "Setting kv_transfer_config.engine_id to %s",
                    vllm_config.kv_transfer_config.engine_id,
                )

1072
1073
            parallel_config.data_parallel_index = dp_rank
            if data_parallel and vllm_config.model_config.is_moe:
1074
1075
1076
1077
                # Set data parallel rank for this engine process.
                parallel_config.data_parallel_rank = dp_rank
                engine_core = DPEngineCoreProc(*args, **kwargs)
            else:
1078
1079
1080
1081
1082
1083
1084
                # Non-MoE DP ranks are completely independent, so treat like DP=1.
                # Note that parallel_config.data_parallel_index will still reflect
                # the original DP rank.
                parallel_config.data_parallel_size = 1
                parallel_config.data_parallel_size_local = 1
                parallel_config.data_parallel_rank = 0
                engine_core = EngineCoreProc(*args, engine_index=dp_rank, **kwargs)
1085

1086
            assert engine_core is not None
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102

            def wakeup_engine():
                # Wakes up idle engine via input_queue when shutdown is requested
                # Not safe in a signal handler - we may interrupt the main thread
                # while it is holding the non-reentrant input_queue.mutex
                engine_core.input_queue.put_nowait((EngineCoreRequestType.WAKEUP, None))

            signal_callback = SignalCallback(wakeup_engine)

            def signal_handler(signum, frame):
                engine_core.shutdown_state = EngineShutdownState.REQUESTED
                signal_callback.trigger()

            signal.signal(signal.SIGTERM, signal_handler)
            signal.signal(signal.SIGINT, signal_handler)

1103
1104
            engine_core.run_busy_loop()

1105
        except SystemExit:
1106
            logger.debug("EngineCore exiting.")
1107
            raise
1108
1109
1110
1111
1112
1113
1114
        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
1115
        finally:
1116
1117
1118
1119
            signal.signal(signal.SIGTERM, signal.SIG_DFL)
            signal.signal(signal.SIGINT, signal.SIG_DFL)
            if signal_callback is not None:
                signal_callback.stop()
1120
1121
1122
            if engine_core is not None:
                engine_core.shutdown()

1123
1124
1125
    def _init_data_parallel(self, vllm_config: VllmConfig):
        pass

1126
1127
1128
1129
1130
1131
1132
1133
    def has_work(self) -> bool:
        """Returns true if the engine should be stepped."""
        return (
            self.engines_running
            or self.scheduler.has_requests()
            or bool(self.batch_queue)
        )

1134
1135
1136
1137
    def is_running(self) -> bool:
        """Returns true if shutdown has not been requested."""
        return self.shutdown_state == EngineShutdownState.RUNNING

1138
1139
    def run_busy_loop(self):
        """Core busy loop of the EngineCore."""
1140
        while self._handle_shutdown():
1141
            # 1) Poll the input queue until there is work to do.
1142
1143
            self._process_input_queue()
            # 2) Step the engine core and return the outputs.
1144
            self._process_engine_step()
1145

1146
1147
        raise SystemExit

1148
1149
1150
1151
    def _process_input_queue(self):
        """Exits when an engine step needs to be performed."""

        waited = False
1152
        while not self.has_work() and self.is_running():
1153
1154
            # Notify callbacks waiting for engine to become idle.
            self._notify_idle_state_callbacks()
1155
1156
1157
1158
1159
1160
1161
            if self.input_queue.empty():
                # Drain aborts queue; all aborts are also processed via input_queue.
                with self.aborts_queue.mutex:
                    self.aborts_queue.queue.clear()
                if logger.isEnabledFor(DEBUG):
                    logger.debug("EngineCore waiting for work.")
                    waited = True
1162
1163
1164
1165
1166
1167
1168
1169
            block = self.process_input_queue_block
            try:
                req = self.input_queue.get(block=block)
                self._handle_client_request(*req)
            except queue.Empty:
                break
            if not block:
                break
1170
1171

        if waited:
1172
            logger.debug("EngineCore loop active.")
1173
1174
1175
1176
1177
1178

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

1179
    def _process_engine_step(self) -> bool:
1180
1181
1182
        """Called only when there are unfinished local requests."""

        # Step the engine core.
1183
        outputs, model_executed = self.step_fn()
1184
        # Put EngineCoreOutputs into the output queue.
1185
        for output in outputs.items() if outputs else ():
1186
            self.output_queue.put_nowait(output)
1187
1188
        # Post-step hook.
        self.post_step(model_executed)
1189

1190
1191
1192
1193
1194
1195
1196
        # If no model execution happened but there are waiting requests
        # (e.g., WAITING_FOR_REMOTE_KVS), yield the GIL briefly to allow
        # background threads (like NIXL handshake) to make progress.
        # Without this, the tight polling loop can starve background threads.
        if not model_executed and self.scheduler.has_unfinished_requests():
            time.sleep(0.001)

1197
1198
        return model_executed

1199
1200
1201
1202
    def _notify_idle_state_callbacks(self) -> None:
        while self._idle_state_callbacks:
            callback = self._idle_state_callbacks.pop()
            callback(self)
1203

1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
    def _handle_shutdown(self) -> bool:
        # Check if shutdown was requested and handle it
        if self.shutdown_state == EngineShutdownState.RUNNING:
            return True

        if self.shutdown_state == EngineShutdownState.REQUESTED:
            shutdown_timeout = self.vllm_config.shutdown_timeout

            logger.info("Shutdown initiated (timeout=%d)", shutdown_timeout)

            if shutdown_timeout == 0:
                num_requests = self.scheduler.get_num_unfinished_requests()
                if num_requests > 0:
                    logger.info("Aborting %d requests", num_requests)
                aborted_reqs = self.scheduler.finish_requests(
                    None, RequestStatus.FINISHED_ABORTED
                )
                self._send_abort_outputs(aborted_reqs)
            else:
                num_requests = self.scheduler.get_num_unfinished_requests()
                if num_requests > 0:
                    logger.info(
                        "Draining %d in-flight requests (timeout=%ds)",
                        num_requests,
                        shutdown_timeout,
                    )

            self.shutdown_state = EngineShutdownState.SHUTTING_DOWN

        # Exit when no work remaining
        if not self.has_work():
            logger.info("Shutdown complete")
            return False

        return True

1240
1241
1242
    def _handle_client_request(
        self, request_type: EngineCoreRequestType, request: Any
    ) -> None:
1243
        """Dispatch request from client."""
1244

1245
1246
1247
        if request_type == EngineCoreRequestType.WAKEUP:
            return
        elif request_type == EngineCoreRequestType.ADD:
1248
            req, request_wave = request
1249
1250
            if self._reject_add_in_shutdown(req):
                return
1251
            self.add_request(req, request_wave)
1252
        elif request_type == EngineCoreRequestType.ABORT:
1253
            self.abort_requests(request)
1254
        elif request_type == EngineCoreRequestType.UTILITY:
1255
            client_idx, call_id, method_name, args = request
1256
1257
            if self._reject_utility_in_shutdown(client_idx, call_id, method_name):
                return
1258
            output = UtilityOutput(call_id)
1259
            # Lazily look-up utility method so that failure will be handled/returned.
1260
1261
1262
            get_result = lambda: (
                (method := getattr(self, method_name))
                and method(*self._convert_msgspec_args(method, args))
1263
1264
1265
            )
            enqueue_output = lambda out: self.output_queue.put_nowait(
                (client_idx, EngineCoreOutputs(utility_output=out))
1266
            )
1267
            self._invoke_utility_method(method_name, get_result, output, enqueue_output)
1268
1269
1270
        elif request_type == EngineCoreRequestType.EXECUTOR_FAILED:
            raise RuntimeError("Executor failed.")
        else:
1271
1272
1273
            logger.error(
                "Unrecognized input request type encountered: %s", request_type
            )
1274

1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
    def _reject_add_in_shutdown(self, request: Request) -> bool:
        if self.shutdown_state == EngineShutdownState.RUNNING:
            return False

        logger.info("Rejecting request %s (server shutting down)", request.request_id)
        self._send_abort_outputs_to_client([request.request_id], request.client_index)
        return True

    def _reject_utility_in_shutdown(
        self, client_idx: int, call_id: int, method_name: str
    ) -> bool:
        if self.shutdown_state == EngineShutdownState.RUNNING:
            return False

        logger.warning("Rejecting utility call %s (server shutting down)", method_name)
        output = UtilityOutput(call_id, failure_message="Server shutting down")
        self.output_queue.put_nowait(
            (client_idx, EngineCoreOutputs(utility_output=output))
        )
        return True

1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
    @staticmethod
    def _invoke_utility_method(
        name: str, get_result: Callable, output: UtilityOutput, enqueue_output: Callable
    ):
        try:
            result = get_result()
            if isinstance(result, Future):
                # Defer utility output handling until future completion.
                callback = lambda future: EngineCoreProc._invoke_utility_method(
                    name, future.result, output, enqueue_output
                )
                result.add_done_callback(callback)
                return
            output.result = UtilityResult(result)
        except Exception as e:
            logger.exception("Invocation of %s method failed", name)
            output.failure_message = f"Call to {name} method failed: {str(e)}"
        enqueue_output(output)

1315
1316
1317
    @staticmethod
    def _convert_msgspec_args(method, args):
        """If a provided arg type doesn't match corresponding target method
1318
        arg type, try converting to msgspec object."""
1319
1320
1321
1322
1323
        if not args:
            return args
        arg_types = signature(method).parameters.values()
        assert len(args) <= len(arg_types)
        return tuple(
1324
1325
            msgspec.convert(v, type=p.annotation)
            if isclass(p.annotation)
1326
            and issubclass(p.annotation, msgspec.Struct)
1327
1328
1329
1330
            and not isinstance(v, p.annotation)
            else v
            for v, p in zip(args, arg_types)
        )
1331

1332
1333
1334
1335
1336
1337
1338
1339
1340
    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():
1341
1342
1343
1344
            logger.fatal(
                "vLLM shutdown signal from EngineCore failed "
                "to send. Please report this issue."
            )
1345

1346
1347
1348
    def process_input_sockets(
        self,
        input_addresses: list[str],
1349
        coord_input_address: str | None,
1350
1351
1352
        identity: bytes,
        ready_event: threading.Event,
    ):
1353
1354
        """Input socket IO thread."""

1355
1356
1357
1358
1359
        # Msgpack serialization decoding with optional tensor IPC receiver.
        add_request_decoder = MsgpackDecoder(
            EngineCoreRequest, oob_tensor_provider=self.tensor_ipc_receiver
        )
        generic_decoder = MsgpackDecoder(oob_tensor_provider=self.tensor_ipc_receiver)
1360

1361
1362
1363
        with ExitStack() as stack, zmq.Context() as ctx:
            input_sockets = [
                stack.enter_context(
1364
1365
1366
1367
                    make_zmq_socket(
                        ctx, input_address, zmq.DEALER, identity=identity, bind=False
                    )
                )
1368
1369
1370
1371
1372
1373
                for input_address in input_addresses
            ]
            if coord_input_address is None:
                coord_socket = None
            else:
                coord_socket = stack.enter_context(
1374
1375
1376
1377
1378
1379
1380
1381
                    make_zmq_socket(
                        ctx,
                        coord_input_address,
                        zmq.XSUB,
                        identity=identity,
                        bind=False,
                    )
                )
1382
                # Send subscription message to coordinator.
1383
                coord_socket.send(b"\x01")
1384
1385
1386

            # Register sockets with poller.
            poller = zmq.Poller()
1387
1388
            ready_response = EngineCoreReadyResponse(
                max_model_len=self.vllm_config.model_config.max_model_len,
1389
1390
                num_gpu_blocks=self.vllm_config.cache_config.num_gpu_blocks or 0,
                dp_stats_address=self.frontend_stats_publish_address,
1391
1392
            )
            ready_payload = msgspec.msgpack.encode(ready_response)
1393
1394
1395
1396
            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.
1397
                input_socket.send(ready_payload)
1398
                poller.register(input_socket, zmq.POLLIN)
1399

1400
            if coord_socket is not None:
1401
1402
                # Wait for ready message from coordinator.
                assert coord_socket.recv() == b"READY"
1403
                poller.register(coord_socket, zmq.POLLIN)
1404

1405
1406
            ready_event.set()
            del ready_event
1407
1408
1409
            while True:
                for input_socket, _ in poller.poll():
                    # (RequestType, RequestData)
1410
                    type_frame, *data_frames = input_socket.recv_multipart(copy=False)
1411
1412
1413
1414
1415
                    # NOTE(yongji): ignore READY message sent by DP coordinator
                    # that is used to notify newly started engines
                    if type_frame.buffer == b"READY":
                        assert input_socket == coord_socket
                        continue
1416
                    request_type = EngineCoreRequestType(bytes(type_frame.buffer))
1417
1418

                    # Deserialize the request data.
1419
                    request: Any
1420
                    if request_type == EngineCoreRequestType.ADD:
1421
1422
1423
1424
1425
1426
                        req: EngineCoreRequest = add_request_decoder.decode(data_frames)
                        try:
                            request = self.preprocess_add_request(req)
                        except Exception:
                            self._handle_request_preproc_error(req)
                            continue
1427
1428
                    else:
                        request = generic_decoder.decode(data_frames)
1429

1430
1431
1432
1433
1434
1435
1436
                        if request_type == EngineCoreRequestType.ABORT:
                            # Aborts are added to *both* queues, allows us to eagerly
                            # process aborts while also ensuring ordering in the input
                            # queue to avoid leaking requests. This is ok because
                            # aborting in the scheduler is idempotent.
                            self.aborts_queue.put_nowait(request)

1437
1438
1439
                    # Push to input queue for core busy loop.
                    self.input_queue.put_nowait((request_type, request))

1440
    def process_output_sockets(
1441
        self, output_paths: list[str], coord_output_path: str | None, engine_index: int
1442
    ):
1443
1444
1445
        """Output socket IO thread."""

        # Msgpack serialization encoding.
1446
        encoder = MsgpackEncoder()
1447
1448
1449
1450
1451
1452
        # 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]]()
1453

1454
1455
        # We must set linger to ensure the ENGINE_CORE_DEAD
        # message is sent prior to closing the socket.
1456
1457
1458
        with ExitStack() as stack, zmq.Context() as ctx:
            sockets = [
                stack.enter_context(
1459
1460
                    make_zmq_socket(ctx, output_path, zmq.PUSH, linger=4000)
                )
1461
1462
                for output_path in output_paths
            ]
1463
1464
1465
1466
1467
1468
1469
1470
1471
            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
            )
1472
1473
            max_reuse_bufs = len(sockets) + 1

1474
            while True:
1475
1476
1477
1478
                output = self.output_queue.get()
                if output == EngineCoreProc.ENGINE_CORE_DEAD:
                    for socket in sockets:
                        socket.send(output)
1479
                    break
1480
1481
                assert not isinstance(output, bytes)
                client_index, outputs = output
1482
                outputs.engine_index = engine_index
1483

1484
1485
1486
1487
1488
1489
1490
                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

1491
1492
1493
1494
1495
                # 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()
1496
                buffers = encoder.encode_into(outputs, buffer)
1497
1498
1499
                tracker = sockets[client_index].send_multipart(
                    buffers, copy=False, track=True
                )
1500
1501
1502
                if not tracker.done:
                    ref = outputs if len(buffers) > 1 else None
                    pending.appendleft((tracker, ref, buffer))
1503
1504
                elif len(reuse_buffers) < max_reuse_bufs:
                    # Limit the number of buffers to reuse.
1505
                    reuse_buffers.append(buffer)
1506

1507
1508
1509
1510
1511
1512
1513
    def _handle_request_preproc_error(self, request: EngineCoreRequest) -> None:
        """Log and return a request-scoped error response for exceptions raised
        from the add request preprocessing in the input socket processing thread.
        """
        logger.exception(
            "Unexpected error pre-processing request %s", request.request_id
        )
1514
        self._send_error_outputs_to_client([request.request_id], request.client_index)
1515

1516
1517
1518
1519
1520
    def pause_scheduler(
        self, mode: PauseMode = "abort", clear_cache: bool = True
    ) -> Future | None:
        """Pause generation; behavior depends on mode.

1521
1522
1523
1524
1525
1526
1527
1528
1529
        All pause modes queue new adds -- "abort" and "keep" skip step();
        "wait" allows step() so in-flight requests can drain.

        - ``abort``: Set PAUSED_NEW, abort all requests, wait for abort
          outputs to be sent (when running with output_queue), optionally
          clear caches, then complete the returned Future.
        - ``wait``: Set PAUSED_NEW (queue adds, keep stepping); when drained,
          optionally clear caches, then complete the returned Future.
        - ``keep``: Set PAUSED_ALL; return a Future that completes when the
1530
1531
1532
1533
1534
          output queue is empty.
        """
        if mode not in ("keep", "abort", "wait"):
            raise ValueError(f"Invalid pause mode: {mode}")

1535
        def engine_idle_callback(engine: "EngineCoreProc", future: Future[Any]) -> None:
1536
            if clear_cache:
1537
                engine._reset_caches()
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
            future.set_result(None)

        if mode == "abort":
            aborted_reqs = self.scheduler.finish_requests(
                None, RequestStatus.FINISHED_ABORTED
            )
            self._send_abort_outputs(aborted_reqs)

        pause_state = PauseState.PAUSED_ALL if mode == "keep" else PauseState.PAUSED_NEW
        self.scheduler.set_pause_state(pause_state)
1548
1549
1550
1551
1552
1553
1554
1555
        if not self.has_work():
            if clear_cache:
                self._reset_caches()
            return None

        future = Future[Any]()
        self._idle_state_callbacks.append(partial(engine_idle_callback, future=future))
        return future
1556

1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
    def _send_finish_outputs_to_client(
        self, req_ids: list[str], client_index: int, finish_reason: FinishReason
    ) -> None:
        outputs = [
            EngineCoreOutput(req_id, [], finish_reason=finish_reason)
            for req_id in req_ids
        ]
        eco = EngineCoreOutputs(finished_requests=req_ids, outputs=outputs)
        self.output_queue.put_nowait((client_index, eco))

    def _send_abort_outputs_to_client(
        self, req_ids: list[str], client_index: int
    ) -> None:
        self._send_finish_outputs_to_client(req_ids, client_index, FinishReason.ABORT)

    def _send_error_outputs_to_client(
        self, req_ids: list[str], client_index: int
    ) -> None:
        self._send_finish_outputs_to_client(req_ids, client_index, FinishReason.ERROR)

1577
    def _send_abort_outputs(self, aborted_reqs: list[tuple[str, int]]) -> None:
1578
        # TODO(nick) this will be moved inside the scheduler
1579
1580
1581
1582
1583
1584
        if aborted_reqs:
            # Map client_index to list of request_ids that belong to that client.
            by_client = defaultdict[int, set[str]](set)
            for req_id, client_index in aborted_reqs:
                by_client[client_index].add(req_id)
            for client_index, req_ids in by_client.items():
1585
                self._send_abort_outputs_to_client(list(req_ids), client_index)
1586

1587
1588
1589
1590
1591
1592
1593
1594

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

    def __init__(
        self,
        vllm_config: VllmConfig,
1595
        local_client: bool,
1596
        handshake_address: str,
1597
1598
        executor_class: type[Executor],
        log_stats: bool,
1599
        client_handshake_address: str | None = None,
1600
        tensor_queue: Queue | None = None,
1601
    ):
1602
1603
1604
1605
        assert vllm_config.model_config.is_moe, (
            "DPEngineCoreProc should only be used for MoE models"
        )

1606
1607
        # Counts forward-passes of the model so that we can synchronize
        # finished with DP peers every N steps.
1608
        self.step_counter = 0
1609
        self.current_wave = 0
Rui Qiao's avatar
Rui Qiao committed
1610
        self.last_counts = (0, 0)
1611

1612
1613
1614
1615
        from vllm.distributed.elastic_ep.elastic_state import ElasticEPScalingState

        self.eep_scaling_state: ElasticEPScalingState | None = None

1616
1617
        # Initialize the engine.
        dp_rank = vllm_config.parallel_config.data_parallel_rank
1618
1619
1620
1621
1622
1623
1624
        super().__init__(
            vllm_config,
            local_client,
            handshake_address,
            executor_class,
            log_stats,
            client_handshake_address,
1625
            engine_index=dp_rank,
1626
            tensor_queue=tensor_queue,
1627
        )
1628
1629
1630

    def _init_data_parallel(self, vllm_config: VllmConfig):
        # Configure GPUs and stateless process group for data parallel.
1631
1632
1633
1634
        parallel_config = vllm_config.parallel_config
        dp_rank = parallel_config.data_parallel_rank
        dp_size = parallel_config.data_parallel_size
        local_dp_rank = parallel_config.data_parallel_rank_local
1635
1636

        assert dp_size > 1
1637
        assert local_dp_rank is not None
1638
1639
        assert 0 <= local_dp_rank <= dp_rank < dp_size

1640
        self.dp_rank = dp_rank
1641
1642
        dp_group, dp_store = parallel_config.stateless_init_dp_group(return_store=True)
        self.dp_group, self.dp_store = dp_group, dp_store
1643
1644
1645
1646
1647
1648

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

1649
    def add_request(self, request: Request, request_wave: int = 0):
1650
        super().add_request(request, request_wave)
1651
1652
1653
        if self.has_coordinator and request_wave != self.current_wave:
            if request_wave > self.current_wave:
                self.current_wave = request_wave
1654
1655
1656
1657
1658
            elif (
                not self.engines_running
                and self.scheduler.pause_state == PauseState.UNPAUSED
            ):
                self.engines_running = True
1659
1660
1661
                # Request received for an already-completed wave, notify
                # front-end that we need to start the next one.
                self.output_queue.put_nowait(
1662
1663
                    (-1, EngineCoreOutputs(start_wave=self.current_wave))
                )
1664

1665
1666
    def resume_scheduler(self):
        super().resume_scheduler()
1667
1668
1669
1670
1671
        if (
            self.has_coordinator
            and not self.engines_running
            and self.scheduler.has_unfinished_requests()
        ):
1672
1673
1674
1675
            # Wake up other DP engines.
            self.output_queue.put_nowait(
                (-1, EngineCoreOutputs(start_wave=self.current_wave))
            )
1676

1677
1678
1679
    def _handle_client_request(
        self, request_type: EngineCoreRequestType, request: Any
    ) -> None:
1680
        if request_type == EngineCoreRequestType.START_DP_WAVE:
1681
1682
            new_wave, exclude_eng_index = request
            if exclude_eng_index != self.engine_index and (
1683
1684
                new_wave >= self.current_wave
            ):
1685
1686
                self.current_wave = new_wave
                if not self.engines_running:
1687
                    logger.debug("EngineCore starting idle loop for wave %d.", new_wave)
1688
1689
1690
1691
                    self.engines_running = True
        else:
            super()._handle_client_request(request_type, request)

1692
    def _maybe_publish_request_counts(self):
1693
        if not self.publish_dp_lb_stats:
1694
1695
1696
1697
1698
1699
            return

        # Publish our request counts (if they've changed).
        counts = self.scheduler.get_request_counts()
        if counts != self.last_counts:
            self.last_counts = counts
1700
1701
1702
1703
            stats = SchedulerStats(
                *counts, step_counter=self.step_counter, current_wave=self.current_wave
            )
            self.output_queue.put_nowait((-1, EngineCoreOutputs(scheduler_stats=stats)))
1704

1705
1706
1707
1708
    def run_busy_loop(self):
        """Core busy loop of the EngineCore for data parallel case."""

        # Loop until process is sent a SIGINT or SIGTERM
1709
        while self._handle_shutdown():
1710
1711
1712
            # 1) Poll the input queue until there is work to do.
            self._process_input_queue()

1713
1714
1715
            if self.eep_scaling_state is not None:
                _ = self.eep_scaling_state.progress()
                if self.eep_scaling_state.is_complete():
1716
1717
                    if self.eep_scaling_state.worker_type == "removing":
                        raise SystemExit
1718
1719
1720
                    self.process_input_queue_block = True
                    self.eep_scaling_state = None

1721
            executed = self._process_engine_step()
1722
            self._maybe_publish_request_counts()
1723

1724
            local_unfinished_reqs = self.scheduler.has_unfinished_requests()
1725
1726
            if not executed:
                if not local_unfinished_reqs and not self.engines_running:
1727
1728
1729
                    # All engines are idle.
                    continue

1730
1731
                # We are in a running state and so must execute a dummy pass
                # if the model didn't execute any ready requests.
1732
1733
1734
                self.execute_dummy_batch()

            # 3) All-reduce operation to determine global unfinished reqs.
1735
            self.engines_running = self._has_global_unfinished_reqs(
1736
1737
                local_unfinished_reqs
            )
1738

1739
            if not self.engines_running:
1740
                if self.dp_rank == 0 or not self.has_coordinator:
1741
                    # Notify client that we are pausing the loop.
1742
1743
1744
                    logger.debug(
                        "Wave %d finished, pausing engine loop.", self.current_wave
                    )
1745
1746
1747
1748
                    # 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
1749
                    self.output_queue.put_nowait(
1750
1751
1752
1753
1754
                        (
                            client_index,
                            EngineCoreOutputs(wave_complete=self.current_wave),
                        )
                    )
1755
                # Increment wave count and reset step counter.
1756
                self.current_wave += 1
1757
                self.step_counter = 0
1758

1759
1760
        raise SystemExit

1761
    def _has_global_unfinished_reqs(self, local_unfinished: bool) -> bool:
1762
        # Optimization - only perform finish-sync all-reduce every 32 steps.
1763
1764
        self.step_counter += 1
        if self.step_counter % 32 != 0:
1765
1766
            return True

1767
        return ParallelConfig.has_unfinished_dp(self.dp_group, local_unfinished)
Rui Qiao's avatar
Rui Qiao committed
1768

1769
    def reinitialize_distributed(
1770
1771
        self, reconfig_request: ReconfigureDistributedRequest
    ) -> None:
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
        from copy import deepcopy

        from vllm.distributed.elastic_ep.elastic_state import ElasticEPScalingState

        new_parallel_config = deepcopy(self.vllm_config.parallel_config)
        old_dp_size = new_parallel_config.data_parallel_size
        new_parallel_config.data_parallel_size = reconfig_request.new_data_parallel_size
        if (
            reconfig_request.new_data_parallel_rank
            != ReconfigureRankType.KEEP_CURRENT_RANK
        ):
            new_parallel_config.data_parallel_rank = (
                reconfig_request.new_data_parallel_rank
            )
        new_parallel_config.data_parallel_master_ip = (
1787
            reconfig_request.new_data_parallel_master_ip
1788
        )
1789
        new_parallel_config.data_parallel_master_port = (
1790
            reconfig_request.new_data_parallel_master_port
1791
        )
1792
1793
        new_parallel_config._data_parallel_master_port_list = (
            reconfig_request.new_data_parallel_master_port_list
1794
        )
1795
        new_parallel_config._coord_store_port = reconfig_request.coord_store_port
1796

1797
1798
        is_scale_down = reconfig_request.new_data_parallel_size < old_dp_size
        is_shutdown = (
1799
1800
            reconfig_request.new_data_parallel_rank
            == ReconfigureRankType.SHUTDOWN_CURRENT_RANK
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
        )

        self.eep_scaling_state = ElasticEPScalingState(
            model_executor=self.model_executor,
            engine_core=self,
            vllm_config=self.vllm_config,
            new_parallel_config=new_parallel_config,
            worker_type="removing" if is_shutdown else "existing",
            scale_type="scale_down" if is_scale_down else "scale_up",
            reconfig_request=reconfig_request,
        )
        self.process_input_queue_block = False
        logger.info(
            "[Elastic EP] Received reconfiguration request and starting scaling up/down"
        )

    def _eep_send_engine_core_notification(
        self,
        notification_type: EEPNotificationType,
        vllm_config: VllmConfig | None = None,
    ):
        """
        Send notifications to EngineCoreClient, which can then forward
        the notifications to other engine core processes. It is used for:
Jiayi Yan's avatar
Jiayi Yan committed
1825
        1) In scale up: new core engines to notify existing core engines
1826
1827
1828
           that they are ready;
        2) In scale down: removing core engines to notify EngineCoreClient
           so EngineCoreClient can release their ray placement groups;
Jiayi Yan's avatar
Jiayi Yan committed
1829
        3) Both scale up/down: to notify EngineCoreClient that existing
1830
1831
1832
1833
           core engines have already switched to the new parallel setup.
        """
        if vllm_config is None:
            dp_rank = self.vllm_config.parallel_config.data_parallel_rank
1834
        else:
1835
1836
1837
1838
1839
1840
            dp_rank = vllm_config.parallel_config.data_parallel_rank
        notification_data = (notification_type.value, dp_rank)
        outputs = EngineCoreOutputs(
            utility_output=UtilityOutput(
                call_id=EEP_NOTIFICATION_CALL_ID,
                result=UtilityResult(notification_data),
1841
            )
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
        )
        outputs.engine_index = self.engine_index

        if hasattr(self, "output_thread") and self.output_thread.is_alive():
            self.output_queue.put_nowait((0, outputs))
        else:
            encoder = MsgpackEncoder()
            with (
                zmq.Context() as ctx,
                make_zmq_socket(
                    ctx, self.addresses.outputs[0], zmq.PUSH, linger=4000
                ) as socket,
            ):
                socket.send_multipart(encoder.encode(outputs))

    def eep_handle_engine_core_notification(
        self, notification_type: str | EEPNotificationType
    ):
        """
        Handle notification received from EngineCoreClient
        (forwarded from new core engines).
        """
        assert self.eep_scaling_state is not None
        if isinstance(notification_type, str):
            notification_type = EEPNotificationType(notification_type)
        self.eep_scaling_state.handle_notification(notification_type)

    def _eep_scale_up_before_kv_init(self):
        from vllm.distributed.elastic_ep.elastic_state import ElasticEPScalingState

        self.eep_scaling_state = ElasticEPScalingState(
            model_executor=self.model_executor,
            engine_core=self,
            vllm_config=self.vllm_config,
            new_parallel_config=self.vllm_config.parallel_config,
            worker_type="new",
            scale_type="scale_up",
            reconfig_request=None,
        )
1881
        self.eep_scaling_state.run_pre_kv_init_states()
1882
        self.process_input_queue_block = False
1883

Rui Qiao's avatar
Rui Qiao committed
1884

1885
class EngineCoreActorMixin:
Rui Qiao's avatar
Rui Qiao committed
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
    """
    Ray actor for running EngineCore in a data parallel context
    """

    def __init__(
        self,
        vllm_config: VllmConfig,
        addresses: EngineZmqAddresses,
        dp_rank: int = 0,
        local_dp_rank: int = 0,
    ):
1897
1898
1899
1900
1901
1902
1903
        # Initialize tracer for distributed tracing if configured.
        maybe_init_worker_tracer(
            instrumenting_module_name="vllm.engine_core",
            process_kind="engine_core",
            process_name=f"DPEngineCoreActor_DP{dp_rank}",
        )

Rui Qiao's avatar
Rui Qiao committed
1904
        self.addresses = addresses
1905
        vllm_config.parallel_config.data_parallel_index = dp_rank
1906
        vllm_config.parallel_config.data_parallel_rank_local = local_dp_rank
Rui Qiao's avatar
Rui Qiao committed
1907

1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
        # 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
1918
1919
1920
1921
1922
1923
1924
        # 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.
1925
        self._set_visible_devices(vllm_config, local_dp_rank)
Rui Qiao's avatar
Rui Qiao committed
1926

1927
    def _set_visible_devices(self, vllm_config: VllmConfig, local_dp_rank: int):
1928
        from vllm.platforms import current_platform
1929

1930
1931
1932
1933
        if current_platform.is_xpu():
            pass
        else:
            device_control_env_var = current_platform.device_control_env_var
1934
1935
1936
            self._set_cuda_visible_devices(
                vllm_config, local_dp_rank, device_control_env_var
            )
1937

1938
1939
1940
    def _set_cuda_visible_devices(
        self, vllm_config: VllmConfig, local_dp_rank: int, device_control_env_var: str
    ):
1941
1942
1943
        world_size = vllm_config.parallel_config.world_size
        # Set CUDA_VISIBLE_DEVICES or equivalent.
        try:
1944
1945
1946
            value = get_device_indices(
                device_control_env_var, local_dp_rank, world_size
            )
1947
            os.environ[device_control_env_var] = value
1948
1949
1950
1951
1952
        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}) "
1953
1954
                f'base value: "{os.getenv(device_control_env_var)}"'
            ) from e
1955

Rui Qiao's avatar
Rui Qiao committed
1956
    @contextmanager
1957
1958
1959
1960
1961
1962
    def _perform_handshakes(
        self,
        handshake_address: str,
        identity: bytes,
        local_client: bool,
        vllm_config: VllmConfig,
1963
        client_handshake_address: str | None,
1964
    ):
Rui Qiao's avatar
Rui Qiao committed
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
        """
        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:
1987
            self.run_busy_loop()  # type: ignore[attr-defined]
Rui Qiao's avatar
Rui Qiao committed
1988
1989
1990
1991
1992
1993
1994
        except SystemExit:
            logger.debug("EngineCore exiting.")
            raise
        except Exception:
            logger.exception("EngineCore encountered a fatal error.")
            raise
        finally:
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
            self.shutdown()  # type: ignore[attr-defined]


class DPMoEEngineCoreActor(EngineCoreActorMixin, DPEngineCoreProc):
    """Used for MoE model data parallel cases."""

    def __init__(
        self,
        vllm_config: VllmConfig,
        local_client: bool,
        addresses: EngineZmqAddresses,
        executor_class: type[Executor],
        log_stats: bool,
        dp_rank: int = 0,
        local_dp_rank: int = 0,
    ):
        vllm_config.parallel_config.data_parallel_rank = dp_rank

        EngineCoreActorMixin.__init__(
            self, vllm_config, addresses, dp_rank, local_dp_rank
        )
        DPEngineCoreProc.__init__(
            self, vllm_config, local_client, "", executor_class, log_stats
        )


class EngineCoreActor(EngineCoreActorMixin, EngineCoreProc):
    """Used for non-MoE and/or non-DP cases."""

    def __init__(
        self,
        vllm_config: VllmConfig,
        local_client: bool,
        addresses: EngineZmqAddresses,
        executor_class: type[Executor],
        log_stats: bool,
        dp_rank: int = 0,
        local_dp_rank: int = 0,
    ):
        vllm_config.parallel_config.data_parallel_size = 1
        vllm_config.parallel_config.data_parallel_size_local = 1
        vllm_config.parallel_config.data_parallel_rank = 0

        EngineCoreActorMixin.__init__(
            self, vllm_config, addresses, dp_rank, local_dp_rank
        )
        EngineCoreProc.__init__(
            self,
            vllm_config,
            local_client,
            "",
            executor_class,
            log_stats,
            engine_index=dp_rank,
        )