core.py 81.1 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 typing import Any, TypeVar, cast
17

18
import msgspec
19
20
import zmq

21
import vllm.envs as envs
22
23
from vllm.config import ParallelConfig, VllmConfig
from vllm.distributed import stateless_destroy_torch_distributed_process_group
24
from vllm.envs import enable_envs_cache
25
from vllm.logger import init_logger
26
from vllm.logging_utils.dump_input import dump_engine_exception
27
from vllm.lora.request import LoRARequest
28
from vllm.multimodal import MULTIMODAL_REGISTRY
29
from vllm.tasks import POOLING_TASKS, SupportedTask
30
from vllm.tracing import instrument, maybe_init_worker_tracer
31
from vllm.transformers_utils.config import maybe_register_config_serialize_by_value
32
33
34
35
from vllm.utils.gc_utils import (
    freeze_gc_heap,
    maybe_attach_gc_debug_callback,
)
36
from vllm.utils.hashing import get_hash_fn_by_name
37
from vllm.utils.network_utils import make_zmq_socket
38
from vllm.utils.system_utils import decorate_logs, set_process_title
39
40
41
42
43
44
45
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,
)
46
from vllm.v1.core.sched.interface import PauseState, SchedulerInterface
47
from vllm.v1.core.sched.output import SchedulerOutput
48
from vllm.v1.engine import (
49
50
    EEP_NOTIFICATION_CALL_ID,
    EEPNotificationType,
51
    EngineCoreOutput,
52
53
54
    EngineCoreOutputs,
    EngineCoreRequest,
    EngineCoreRequestType,
55
    FinishReason,
56
    PauseMode,
57
58
59
60
61
62
63
64
    ReconfigureDistributedRequest,
    ReconfigureRankType,
    UtilityOutput,
    UtilityResult,
)
from vllm.v1.engine.utils import (
    EngineHandshakeMetadata,
    EngineZmqAddresses,
65
    SignalCallback,
66
67
    get_device_indices,
)
68
from vllm.v1.executor import Executor
69
from vllm.v1.kv_cache_interface import KVCacheConfig
70
from vllm.v1.metrics.stats import SchedulerStats
71
from vllm.v1.outputs import ModelRunnerOutput
72
from vllm.v1.request import Request, RequestStatus
73
from vllm.v1.serial_utils import MsgpackDecoder, MsgpackEncoder
74
from vllm.v1.structured_output import StructuredOutputManager
75
from vllm.v1.utils import compute_iteration_details
76
77
78
79
from vllm.version import __version__ as VLLM_VERSION

logger = init_logger(__name__)

80
HANDSHAKE_TIMEOUT_MINS = 5
81

82
_R = TypeVar("_R")  # Return type for collective_rpc
83

84
85
86
87

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

88
89
90
91
92
    def __init__(
        self,
        vllm_config: VllmConfig,
        executor_class: type[Executor],
        log_stats: bool,
93
        executor_fail_callback: Callable | None = None,
94
        include_finished_set: bool = False,
95
    ):
96
97
        # plugins need to be loaded at the engine/scheduler level too
        from vllm.plugins import load_general_plugins
98

99
100
        load_general_plugins()

101
        self.vllm_config = vllm_config
102
        if not vllm_config.parallel_config.data_parallel_rank_local:
103
104
105
106
107
            logger.info(
                "Initializing a V1 LLM engine (v%s) with config: %s",
                VLLM_VERSION,
                vllm_config,
            )
108

109
110
        self.log_stats = log_stats

111
112
        # Setup Model.
        self.model_executor = executor_class(vllm_config)
113
        if executor_fail_callback is not None:
114
            self.model_executor.register_failure_callback(executor_fail_callback)
115

116
117
        self.available_gpu_memory_for_kv_cache = -1

118
119
120
        if envs.VLLM_ELASTIC_EP_SCALE_UP_LAUNCH:
            self._eep_scale_up_before_kv_init()

121
        # Setup KV Caches and update CacheConfig after profiling.
122
123
124
        num_gpu_blocks, num_cpu_blocks, kv_cache_config = self._initialize_kv_caches(
            vllm_config
        )
125
126
127
128
129
        if kv_cache_config.kv_cache_groups:
            vllm_config.cache_config.block_size = min(
                g.kv_cache_spec.block_size for g in kv_cache_config.kv_cache_groups
            )
        vllm_config.validate_block_size()
130
131
        vllm_config.cache_config.num_gpu_blocks = num_gpu_blocks
        vllm_config.cache_config.num_cpu_blocks = num_cpu_blocks
132
        self.collective_rpc("initialize_cache", args=(num_gpu_blocks, num_cpu_blocks))
133

134
135
        self.structured_output_manager = StructuredOutputManager(vllm_config)

136
        # Setup scheduler.
137
        Scheduler = vllm_config.scheduler_config.get_scheduler_cls()
138

139
        if len(kv_cache_config.kv_cache_groups) == 0:  # noqa: SIM102
140
141
            # Encoder models without KV cache don't support
            # chunked prefill. But do SSM models?
142
143
144
            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
145

146
147
148
        scheduler_block_size = (
            vllm_config.cache_config.block_size
            * vllm_config.parallel_config.decode_context_parallel_size
149
            * vllm_config.parallel_config.prefill_context_parallel_size
150
151
        )

152
        self.scheduler: SchedulerInterface = Scheduler(
153
            vllm_config=vllm_config,
154
155
            kv_cache_config=kv_cache_config,
            structured_output_manager=self.structured_output_manager,
156
            include_finished_set=include_finished_set,
157
            log_stats=self.log_stats,
158
            block_size=scheduler_block_size,
159
        )
160
        self.use_spec_decode = vllm_config.speculative_config is not None
161
        if self.scheduler.connector is not None:  # type: ignore
162
            self.model_executor.init_kv_output_aggregator(self.scheduler.connector)  # type: ignore
163

164
        self.mm_registry = mm_registry = MULTIMODAL_REGISTRY
165
166
        self.mm_receiver_cache = mm_registry.engine_receiver_cache_from_config(
            vllm_config
167
        )
168

169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
        # 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)

190
191
192
193
194
        # 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
195
        self.batch_queue: (
196
            deque[tuple[Future[ModelRunnerOutput], SchedulerOutput, Future[Any]]] | None
197
        ) = None
198
        if self.batch_queue_size > 1:
199
            logger.debug("Batch queue is enabled with size %d", self.batch_queue_size)
200
            self.batch_queue = deque(maxlen=self.batch_queue_size)
201

202
203
204
        self.is_ec_consumer = (
            vllm_config.ec_transfer_config is None
            or vllm_config.ec_transfer_config.is_ec_consumer
205
        )
206
        self.is_pooling_model = vllm_config.model_config.runner_type == "pooling"
207

208
        self.request_block_hasher: Callable[[Request], list[BlockHash]] | None = None
209
        if vllm_config.cache_config.enable_prefix_caching or kv_connector is not None:
210
            caching_hash_fn = get_hash_fn_by_name(
211
212
                vllm_config.cache_config.prefix_caching_hash_algo
            )
213
214
215
            init_none_hash(caching_hash_fn)

            self.request_block_hasher = get_request_block_hasher(
216
                scheduler_block_size, caching_hash_fn
217
            )
218

219
220
221
        self.step_fn = (
            self.step if self.batch_queue is None else self.step_with_batch_queue
        )
222
        self.async_scheduling = vllm_config.scheduler_config.async_scheduling
223

224
        self.aborts_queue = queue.Queue[list[str]]()
225

226
        self._idle_state_callbacks: list[Callable] = []
227

228
229
230
        # Mark the startup heap as static so that it's ignored by GC.
        # Reduces pause times of oldest generation collections.
        freeze_gc_heap()
231
232
        # If enable, attach GC debugger after static variable freeze.
        maybe_attach_gc_debug_callback()
233
234
235
        # Enable environment variable cache (e.g. assume no more
        # environment variable overrides after this point)
        enable_envs_cache()
236

237
    @instrument(span_name="Prepare model")
238
    def _initialize_kv_caches(
239
240
        self, vllm_config: VllmConfig
    ) -> tuple[int, int, KVCacheConfig]:
241
        start = time.time()
242

243
        # Get all kv cache needed by the model
244
        kv_cache_specs = self.model_executor.get_kv_cache_specs()
245

246
247
        has_kv_cache = any(kv_cache_spec for kv_cache_spec in kv_cache_specs)
        if has_kv_cache:
248
249
250
251
            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
252
253
254
                available_gpu_memory = [self.available_gpu_memory_for_kv_cache] * len(
                    kv_cache_specs
                )
255
256
257
            else:
                # Profiles the peak memory usage of the model to determine how
                # much memory can be allocated for kv cache.
258
259
                available_gpu_memory = self.model_executor.determine_available_memory()
                self.available_gpu_memory_for_kv_cache = available_gpu_memory[0]
260
261
262
        else:
            # Attention free models don't need memory for kv cache
            available_gpu_memory = [0] * len(kv_cache_specs)
263

264
        assert len(kv_cache_specs) == len(available_gpu_memory)
265

266
267
268
        # Track max_model_len before KV cache config to detect auto-fit changes
        max_model_len_before = vllm_config.model_config.max_model_len

269
270
271
        kv_cache_configs = get_kv_cache_configs(
            vllm_config, kv_cache_specs, available_gpu_memory
        )
272
273
274
275
276
277
278
279

        # 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,))

280
        scheduler_kv_cache_config = generate_scheduler_kv_cache_config(kv_cache_configs)
281
        num_gpu_blocks = scheduler_kv_cache_config.num_blocks
282
        num_cpu_blocks = 0
283
284

        # Initialize kv cache and warmup the execution
285
        self.model_executor.initialize_from_config(kv_cache_configs)
286

287
        elapsed = time.time() - start
288
        logger.info_once(
289
            "init engine (profile, create kv cache, warmup model) took %.2f seconds",
290
            elapsed,
291
            scope="local",
292
        )
293
        return num_gpu_blocks, num_cpu_blocks, scheduler_kv_cache_config
294

295
296
297
    def get_supported_tasks(self) -> tuple[SupportedTask, ...]:
        return self.model_executor.supported_tasks

298
299
    def add_request(self, request: Request, request_wave: int = 0):
        """Add request to the scheduler.
300

301
302
303
        `request_wave`: indicate which wave of requests this is expected to
        belong to in DP case
        """
304
305
306
        # Validate the request_id type.
        if not isinstance(request.request_id, str):
            raise TypeError(
307
308
                f"request_id must be a string, got {type(request.request_id)}"
            )
309

310
        if pooling_params := request.pooling_params:
311
            supported_pooling_tasks = [
312
                task for task in self.get_supported_tasks() if task in POOLING_TASKS
313
314
            ]

315
            if pooling_params.task not in supported_pooling_tasks:
316
317
318
319
                raise ValueError(
                    f"Unsupported task: {pooling_params.task!r} "
                    f"Supported tasks: {supported_pooling_tasks}"
                )
320

321
        if request.kv_transfer_params is not None and (
322
323
324
325
326
327
            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
328

329
        self.scheduler.add_request(request)
330

331
    def abort_requests(self, request_ids: list[str]):
332
333
334
335
336
        """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).
337
        self.scheduler.finish_requests(request_ids, RequestStatus.FINISHED_ABORTED)
338

339
340
    @contextmanager
    def log_error_detail(self, scheduler_output: SchedulerOutput):
341
        """Execute the model and log detailed info on failure."""
342
        try:
343
            yield
344
345
346
347
348
        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.

349
            # NOTE: This method is exception-free
350
351
352
            dump_engine_exception(
                self.vllm_config, scheduler_output, self.scheduler.make_stats()
            )
353
354
            raise err

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
382
383
384
    @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

385
    def step(self) -> tuple[dict[int, EngineCoreOutputs], bool]:
386
387
388
389
390
        """Schedule, execute, and make output.

        Returns tuple of outputs and a flag indicating whether the model
        was executed.
        """
391

392
393
394
        # Check for any requests remaining in the scheduler - unfinished,
        # or finished and not yet removed from the batch.
        if not self.scheduler.has_requests():
395
            return {}, False
396
397
398
        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)
399
400
401
402
        with (
            self.log_error_detail(scheduler_output),
            self.log_iteration_details(scheduler_output),
        ):
403
404
405
406
            model_output = future.result()
            if model_output is None:
                model_output = self.model_executor.sample_tokens(grammar_output)

407
408
409
        # Before processing the model output, process any aborts that happened
        # during the model execution.
        self._process_aborts_queue()
410
411
412
        engine_core_outputs = self.scheduler.update_from_output(
            scheduler_output, model_output
        )
413

414
        return engine_core_outputs, scheduler_output.total_num_scheduled_tokens > 0
415

416
    def post_step(self, model_executed: bool) -> None:
417
418
419
420
        # 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:
421
422
423
424
425
            # 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)

426
    def step_with_batch_queue(
427
        self,
428
    ) -> tuple[dict[int, EngineCoreOutputs] | None, bool]:
429
430
431
432
        """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:
433
434
435
436
        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.
437
438
439
440
441
        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.
        """
442

443
444
        batch_queue = self.batch_queue
        assert batch_queue is not None
445

446
447
448
        # 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.
449
        assert len(batch_queue) < self.batch_queue_size
450

451
        model_executed = False
452
        deferred_scheduler_output = None
453
        if self.scheduler.has_requests():
454
455
456
457
            scheduler_output = self.scheduler.schedule()
            exec_future = self.model_executor.execute_model(
                scheduler_output, non_block=True
            )
458
            if self.is_ec_consumer:
459
                model_executed = scheduler_output.total_num_scheduled_tokens > 0
460

461
            if self.is_pooling_model or not model_executed:
462
463
                # No sampling required (no requests scheduled).
                future = cast(Future[ModelRunnerOutput], exec_future)
464
            else:
465
466
467
                if not scheduler_output.pending_structured_output_tokens:
                    # We aren't waiting for any tokens, get any grammar output
                    # and sample immediately.
468
469
470
                    grammar_output = self.scheduler.get_grammar_bitmask(
                        scheduler_output
                    )
471
472
473
                    future = self.model_executor.sample_tokens(
                        grammar_output, non_block=True
                    )
474
                else:
475
476
477
478
479
                    # 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:
480
                # Add this step's future to the queue.
481
                batch_queue.appendleft((future, scheduler_output, exec_future))
482
483
484
485
486
487
488
489
                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
490
491
492
493
494
495

        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
496
497

        # Block until the next result is available.
498
        future, scheduler_output, exec_model_fut = batch_queue.pop()
499
500
501
502
        with (
            self.log_error_detail(scheduler_output),
            self.log_iteration_details(scheduler_output),
        ):
503
            model_output = future.result()
504
505
506
507
508
            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")
509

510
511
512
        # Before processing the model output, process any aborts that happened
        # during the model execution.
        self._process_aborts_queue()
513
514
515
        engine_core_outputs = self.scheduler.update_from_output(
            scheduler_output, model_output
        )
516
517
518
519
520

        # 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:
521
522
523
524
525
526
527
528
529
530
531
532
            # 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
                )
533
534
535
536
537
538
            # 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)
539
            batch_queue.appendleft((future, deferred_scheduler_output, exec_future))
540

541
        return engine_core_outputs, model_executed
542

543
544
545
546
547
    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()
548
549
                # Should be a list here, but also handle string just in case.
                request_ids.extend((ids,) if isinstance(ids, str) else ids)
550
551
552
            # More efficient to abort all as a single batch.
            self.abort_requests(request_ids)

553
    def shutdown(self):
554
        self.structured_output_manager.clear_backend()
555
556
        if self.model_executor:
            self.model_executor.shutdown()
557
558
        if self.scheduler:
            self.scheduler.shutdown()
559

560
561
    def profile(self, is_start: bool = True, profile_prefix: str | None = None):
        self.model_executor.profile(is_start, profile_prefix)
562

563
564
    def reset_mm_cache(self):
        # NOTE: Since this is mainly for debugging, we don't attempt to
565
        # re-sync the internal caches (P0 sender, P1 receiver)
566
        if self.scheduler.has_unfinished_requests():
567
568
569
570
            logger.warning(
                "Resetting the multi-modal cache when requests are "
                "in progress may lead to desynced internal caches."
            )
571

572
        # The cache either exists in EngineCore or WorkerWrapperBase
573
574
        if self.mm_receiver_cache is not None:
            self.mm_receiver_cache.clear_cache()
575

576
577
        self.model_executor.reset_mm_cache()

578
579
580
581
582
583
    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
        )
584

585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
    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()

605
606
607
608
609
    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()

610
611
    def pause_scheduler(
        self, mode: PauseMode = "abort", clear_cache: bool = True
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
637
638
    ) -> 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()

639
640
641
        return None

    def resume_scheduler(self) -> None:
642
643
        """Resume the scheduler and flush any requests queued while paused."""
        self.scheduler.set_pause_state(PauseState.UNPAUSED)
644
645

    def is_scheduler_paused(self) -> bool:
646
647
        """Return whether the scheduler is in any pause state."""
        return self.scheduler.pause_state != PauseState.UNPAUSED
648

649
    def sleep(self, level: int = 1, mode: PauseMode = "abort") -> None | Future:
650
651
652
653
654
655
656
657
        """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.
658
659
            mode: Pause mode - how to deal with any existing requests, see
                documentation of pause_scheduler method.
660
        """
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685

        # 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
686

687
    def wake_up(self, tags: list[str] | None = None):
688
689
690
691
692
693
        """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:
694
695
696
697
            # 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:
698
            self.model_executor.wake_up(tags)
699

700
701
702
        # Resume scheduling (applies to all levels)
        self.resume_scheduler()

703
    def is_sleeping(self) -> bool:
704
        """Check if engine is sleeping at any level."""
705
        return self.is_scheduler_paused() or self.model_executor.is_sleeping
706

707
    def execute_dummy_batch(self):
708
        self.model_executor.execute_dummy_batch()
709

710
711
712
713
714
715
    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)

716
    def list_loras(self) -> set[int]:
717
718
719
720
        return self.model_executor.list_loras()

    def pin_lora(self, lora_id: int) -> bool:
        return self.model_executor.pin_lora(lora_id)
721

722
723
724
    def save_sharded_state(
        self,
        path: str,
725
726
        pattern: str | None = None,
        max_size: int | None = None,
727
    ) -> None:
728
729
730
731
732
733
        self.model_executor.save_sharded_state(
            path=path, pattern=pattern, max_size=max_size
        )

    def collective_rpc(
        self,
734
735
        method: str | Callable[..., _R],
        timeout: float | None = None,
736
        args: tuple = (),
737
        kwargs: dict[str, Any] | None = None,
738
739
    ) -> list[_R]:
        return self.model_executor.collective_rpc(method, timeout, args, kwargs)
740

741
    def preprocess_add_request(self, request: EngineCoreRequest) -> tuple[Request, int]:
742
        """Preprocess the request.
743

744
745
746
        This function could be directly used in input processing thread to allow
        request initialization running in parallel with Model forward
        """
747
748
        # Note on thread safety: no race condition.
        # `mm_receiver_cache` is reset at the end of LLMEngine init,
749
        # and will only be accessed in the input processing thread afterwards.
750
        if self.mm_receiver_cache is not None and request.mm_features:
751
752
753
            request.mm_features = self.mm_receiver_cache.get_and_update_features(
                request.mm_features
            )
754

755
        req = Request.from_engine_core_request(request, self.request_block_hasher)
756
757
758
759
760
761
762
763
764
        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

765
766
767
768
769
770
771
772
773
774
    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

775

776
777
778
779
780
781
class EngineShutdownState(IntEnum):
    RUNNING = 0
    REQUESTED = 1
    SHUTTING_DOWN = 2


782
783
784
class EngineCoreProc(EngineCore):
    """ZMQ-wrapper for running EngineCore in background process."""

785
    ENGINE_CORE_DEAD = b"ENGINE_CORE_DEAD"
786
    addresses: EngineZmqAddresses
787

788
    @instrument(span_name="EngineCoreProc init")
789
790
    def __init__(
        self,
791
        vllm_config: VllmConfig,
792
        local_client: bool,
793
        handshake_address: str,
794
        executor_class: type[Executor],
795
        log_stats: bool,
796
        client_handshake_address: str | None = None,
797
        *,
798
        engine_index: int = 0,
799
    ):
Rui Qiao's avatar
Rui Qiao committed
800
        self.input_queue = queue.Queue[tuple[EngineCoreRequestType, Any]]()
801
        self.output_queue = queue.Queue[tuple[int, EngineCoreOutputs] | bytes]()
Rui Qiao's avatar
Rui Qiao committed
802
        executor_fail_callback = lambda: self.input_queue.put_nowait(
803
804
            (EngineCoreRequestType.EXECUTOR_FAILED, b"")
        )
805

Rui Qiao's avatar
Rui Qiao committed
806
807
808
        self.engine_index = engine_index
        identity = self.engine_index.to_bytes(length=2, byteorder="little")
        self.engines_running = False
809
        self.shutdown_state = EngineShutdownState.RUNNING
810

811
812
813
814
815
816
817
        with self._perform_handshakes(
            handshake_address,
            identity,
            local_client,
            vllm_config,
            client_handshake_address,
        ) as addresses:
818
            self.client_count = len(addresses.outputs)
819
820

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

838
839
840
841
842
843
844
            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,
                )
845
846
            self._init_data_parallel(vllm_config)

847
            super().__init__(
848
849
850
851
852
                vllm_config,
                executor_class,
                log_stats,
                executor_fail_callback,
                internal_dp_balancing,
853
            )
854

855
856
857
858
859
860
            # 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()
861
862
863
864
865
866
867
868
869
870
            input_thread = threading.Thread(
                target=self.process_input_sockets,
                args=(
                    addresses.inputs,
                    addresses.coordinator_input,
                    identity,
                    ready_event,
                ),
                daemon=True,
            )
871
872
873
874
            input_thread.start()

            self.output_thread = threading.Thread(
                target=self.process_output_sockets,
875
876
877
878
879
880
881
                args=(
                    addresses.outputs,
                    addresses.coordinator_output,
                    self.engine_index,
                ),
                daemon=True,
            )
882
883
884
885
886
887
            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():
888
                    raise RuntimeError("Input socket thread died during startup")
889
890
891
                assert addresses.coordinator_input is not None
                logger.info("Waiting for READY message from DP Coordinator...")

Rui Qiao's avatar
Rui Qiao committed
892
    @contextmanager
893
894
895
896
897
898
    def _perform_handshakes(
        self,
        handshake_address: str,
        identity: bytes,
        local_client: bool,
        vllm_config: VllmConfig,
899
        client_handshake_address: str | None,
Rui Qiao's avatar
Rui Qiao committed
900
    ) -> Generator[EngineZmqAddresses, None, None]:
901
902
903
904
905
        """
        Perform startup handshakes.

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

906
        For DP>1 with internal load-balancing this is with the shared front-end
907
908
        process which may reside on a different node.

909
        For DP>1 with external or hybrid load-balancing, two handshakes are
910
        performed:
911
912
913
914
            - 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.
915
916
        with the exception of the rank 0 and colocated engines themselves which
        don't require the second handshake.
917
918
919
920
921
922

        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
923
        input_ctx = zmq.Context()
924
        is_local = local_client and client_handshake_address is None
925
        headless = not local_client
926
927
928
929
930
931
932
933
934
        handshake = self._perform_handshake(
            input_ctx,
            handshake_address,
            identity,
            is_local,
            headless,
            vllm_config,
            vllm_config.parallel_config,
        )
935
936
937
938
        if client_handshake_address is None:
            with handshake as addresses:
                yield addresses
        else:
939
            assert local_client
940
            local_handshake = self._perform_handshake(
941
942
                input_ctx, client_handshake_address, identity, True, False, vllm_config
            )
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
            with handshake as addresses, local_handshake as client_addresses:
                addresses.inputs = client_addresses.inputs
                addresses.outputs = client_addresses.outputs
                yield addresses

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

    @contextmanager
    def _perform_handshake(
        self,
        ctx: zmq.Context,
        handshake_address: str,
        identity: bytes,
        local_client: bool,
958
        headless: bool,
959
        vllm_config: VllmConfig,
960
        parallel_config_to_update: ParallelConfig | None = None,
961
    ) -> Generator[EngineZmqAddresses, None, None]:
962
963
964
965
966
967
968
969
        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
970
            # Register engine with front-end.
971
972
973
            addresses = self.startup_handshake(
                handshake_socket, local_client, headless, parallel_config_to_update
            )
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
            yield addresses

            # Send ready message.
            num_gpu_blocks = vllm_config.cache_config.num_gpu_blocks
            # We pass back the coordinator stats update address here for the
            # external LB case for our colocated front-end to use (coordinator
            # only runs with rank 0).
            dp_stats_address = self.frontend_stats_publish_address

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

            handshake_socket.send(msgspec.msgpack.encode(ready_msg))
Rui Qiao's avatar
Rui Qiao committed
997

998
    @staticmethod
999
    def startup_handshake(
1000
1001
        handshake_socket: zmq.Socket,
        local_client: bool,
1002
        headless: bool,
1003
        parallel_config: ParallelConfig | None = None,
1004
    ) -> EngineZmqAddresses:
1005
        # Send registration message.
1006
        handshake_socket.send(
1007
1008
1009
1010
1011
1012
1013
1014
            msgspec.msgpack.encode(
                {
                    "status": "HELLO",
                    "local": local_client,
                    "headless": headless,
                }
            )
        )
1015
1016

        # Receive initialization message.
1017
        logger.debug("Waiting for init message from front-end.")
1018
        if not handshake_socket.poll(timeout=HANDSHAKE_TIMEOUT_MINS * 60_000):
1019
1020
1021
1022
1023
            raise RuntimeError(
                "Did not receive response from front-end "
                f"process within {HANDSHAKE_TIMEOUT_MINS} "
                f"minutes"
            )
1024
1025
        init_bytes = handshake_socket.recv()
        init_message: EngineHandshakeMetadata = msgspec.msgpack.decode(
1026
1027
            init_bytes, type=EngineHandshakeMetadata
        )
1028
1029
        logger.debug("Received init message: %s", init_message)

1030
1031
1032
        if parallel_config is not None:
            for key, value in init_message.parallel_config.items():
                setattr(parallel_config, key, value)
1033

1034
        return init_message.addresses
1035
1036

    @staticmethod
1037
    def run_engine_core(*args, dp_rank: int = 0, local_dp_rank: int = 0, **kwargs):
1038
1039
        """Launch EngineCore busy loop in background process."""

1040
1041
1042
        # Ensure we can serialize transformer config after spawning
        maybe_register_config_serialize_by_value()

1043
        engine_core: EngineCoreProc | None = None
1044
        signal_callback: SignalCallback | None = None
1045
        try:
1046
1047
1048
1049
1050
            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
1051
1052
1053
1054
1055
                maybe_init_worker_tracer(
                    instrumenting_module_name="vllm.engine_core",
                    process_kind="engine_core",
                    process_name=f"EngineCore_DP{dp_rank}",
                )
1056
                set_process_title("EngineCore", f"DP{dp_rank}")
1057
            else:
1058
1059
1060
1061
1062
                maybe_init_worker_tracer(
                    instrumenting_module_name="vllm.engine_core",
                    process_kind="engine_core",
                    process_name="EngineCore",
                )
1063
1064
1065
                set_process_title("EngineCore")
            decorate_logs()

1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
            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,
                )

1077
1078
            parallel_config.data_parallel_index = dp_rank
            if data_parallel and vllm_config.model_config.is_moe:
1079
1080
1081
1082
                # Set data parallel rank for this engine process.
                parallel_config.data_parallel_rank = dp_rank
                engine_core = DPEngineCoreProc(*args, **kwargs)
            else:
1083
1084
1085
1086
1087
1088
1089
                # 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)
1090

1091
            assert engine_core is not None
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107

            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)

1108
1109
            engine_core.run_busy_loop()

1110
        except SystemExit:
1111
            logger.debug("EngineCore exiting.")
1112
            raise
1113
1114
1115
1116
1117
1118
1119
        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
1120
        finally:
1121
1122
1123
1124
            signal.signal(signal.SIGTERM, signal.SIG_DFL)
            signal.signal(signal.SIGINT, signal.SIG_DFL)
            if signal_callback is not None:
                signal_callback.stop()
1125
1126
1127
            if engine_core is not None:
                engine_core.shutdown()

1128
1129
1130
    def _init_data_parallel(self, vllm_config: VllmConfig):
        pass

1131
1132
1133
1134
1135
1136
1137
1138
    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)
        )

1139
1140
1141
1142
    def is_running(self) -> bool:
        """Returns true if shutdown has not been requested."""
        return self.shutdown_state == EngineShutdownState.RUNNING

1143
1144
    def run_busy_loop(self):
        """Core busy loop of the EngineCore."""
1145
        while self._handle_shutdown():
1146
            # 1) Poll the input queue until there is work to do.
1147
1148
            self._process_input_queue()
            # 2) Step the engine core and return the outputs.
1149
            self._process_engine_step()
1150

1151
1152
        raise SystemExit

1153
1154
1155
1156
    def _process_input_queue(self):
        """Exits when an engine step needs to be performed."""

        waited = False
1157
        while not self.has_work() and self.is_running():
1158
1159
            # Notify callbacks waiting for engine to become idle.
            self._notify_idle_state_callbacks()
1160
1161
1162
1163
1164
1165
1166
            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
1167
1168
1169
1170
1171
1172
1173
1174
            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
1175
1176

        if waited:
1177
            logger.debug("EngineCore loop active.")
1178
1179
1180
1181
1182
1183

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

1184
    def _process_engine_step(self) -> bool:
1185
1186
1187
        """Called only when there are unfinished local requests."""

        # Step the engine core.
1188
        outputs, model_executed = self.step_fn()
1189
        # Put EngineCoreOutputs into the output queue.
1190
        for output in outputs.items() if outputs else ():
1191
            self.output_queue.put_nowait(output)
1192
1193
        # Post-step hook.
        self.post_step(model_executed)
1194

1195
1196
1197
1198
1199
1200
1201
        # 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)

1202
1203
        return model_executed

1204
1205
1206
1207
    def _notify_idle_state_callbacks(self) -> None:
        while self._idle_state_callbacks:
            callback = self._idle_state_callbacks.pop()
            callback(self)
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
1240
1241
1242
1243
1244
    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

1245
1246
1247
    def _handle_client_request(
        self, request_type: EngineCoreRequestType, request: Any
    ) -> None:
1248
        """Dispatch request from client."""
1249

1250
1251
1252
        if request_type == EngineCoreRequestType.WAKEUP:
            return
        elif request_type == EngineCoreRequestType.ADD:
1253
            req, request_wave = request
1254
1255
            if self._reject_add_in_shutdown(req):
                return
1256
            self.add_request(req, request_wave)
1257
        elif request_type == EngineCoreRequestType.ABORT:
1258
            self.abort_requests(request)
1259
        elif request_type == EngineCoreRequestType.UTILITY:
1260
            client_idx, call_id, method_name, args = request
1261
1262
            if self._reject_utility_in_shutdown(client_idx, call_id, method_name):
                return
1263
            output = UtilityOutput(call_id)
1264
1265
1266
1267
1268
1269
            # Lazily look-up utility method so that failure will be handled/returned.
            get_result = lambda: (method := getattr(self, method_name)) and method(
                *self._convert_msgspec_args(method, args)
            )
            enqueue_output = lambda out: self.output_queue.put_nowait(
                (client_idx, EngineCoreOutputs(utility_output=out))
1270
            )
1271
            self._invoke_utility_method(method_name, get_result, output, enqueue_output)
1272
1273
1274
        elif request_type == EngineCoreRequestType.EXECUTOR_FAILED:
            raise RuntimeError("Executor failed.")
        else:
1275
1276
1277
            logger.error(
                "Unrecognized input request type encountered: %s", request_type
            )
1278

1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
    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

1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
    @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)

1319
1320
1321
    @staticmethod
    def _convert_msgspec_args(method, args):
        """If a provided arg type doesn't match corresponding target method
1322
        arg type, try converting to msgspec object."""
1323
1324
1325
1326
1327
        if not args:
            return args
        arg_types = signature(method).parameters.values()
        assert len(args) <= len(arg_types)
        return tuple(
1328
1329
            msgspec.convert(v, type=p.annotation)
            if isclass(p.annotation)
1330
            and issubclass(p.annotation, msgspec.Struct)
1331
1332
1333
1334
            and not isinstance(v, p.annotation)
            else v
            for v, p in zip(args, arg_types)
        )
1335

1336
1337
1338
1339
1340
1341
1342
1343
1344
    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():
1345
1346
1347
1348
            logger.fatal(
                "vLLM shutdown signal from EngineCore failed "
                "to send. Please report this issue."
            )
1349

1350
1351
1352
    def process_input_sockets(
        self,
        input_addresses: list[str],
1353
        coord_input_address: str | None,
1354
1355
1356
        identity: bytes,
        ready_event: threading.Event,
    ):
1357
1358
1359
        """Input socket IO thread."""

        # Msgpack serialization decoding.
1360
1361
        add_request_decoder = MsgpackDecoder(EngineCoreRequest)
        generic_decoder = MsgpackDecoder()
1362

1363
1364
1365
        with ExitStack() as stack, zmq.Context() as ctx:
            input_sockets = [
                stack.enter_context(
1366
1367
1368
1369
                    make_zmq_socket(
                        ctx, input_address, zmq.DEALER, identity=identity, bind=False
                    )
                )
1370
1371
1372
1373
1374
1375
                for input_address in input_addresses
            ]
            if coord_input_address is None:
                coord_socket = None
            else:
                coord_socket = stack.enter_context(
1376
1377
1378
1379
1380
1381
1382
1383
                    make_zmq_socket(
                        ctx,
                        coord_input_address,
                        zmq.XSUB,
                        identity=identity,
                        bind=False,
                    )
                )
1384
                # Send subscription message to coordinator.
1385
                coord_socket.send(b"\x01")
1386
1387
1388
1389
1390
1391
1392

            # Register sockets with poller.
            poller = zmq.Poller()
            for input_socket in input_sockets:
                # Send initial message to each input socket - this is required
                # before the front-end ROUTER socket can send input messages
                # back to us.
1393
                input_socket.send(b"")
1394
                poller.register(input_socket, zmq.POLLIN)
1395

1396
            if coord_socket is not None:
1397
1398
                # Wait for ready message from coordinator.
                assert coord_socket.recv() == b"READY"
1399
                poller.register(coord_socket, zmq.POLLIN)
1400

1401
1402
            ready_event.set()
            del ready_event
1403
1404
1405
            while True:
                for input_socket, _ in poller.poll():
                    # (RequestType, RequestData)
1406
                    type_frame, *data_frames = input_socket.recv_multipart(copy=False)
1407
1408
1409
1410
1411
                    # 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
1412
                    request_type = EngineCoreRequestType(bytes(type_frame.buffer))
1413
1414

                    # Deserialize the request data.
1415
                    request: Any
1416
                    if request_type == EngineCoreRequestType.ADD:
1417
1418
1419
1420
1421
1422
                        req: EngineCoreRequest = add_request_decoder.decode(data_frames)
                        try:
                            request = self.preprocess_add_request(req)
                        except Exception:
                            self._handle_request_preproc_error(req)
                            continue
1423
1424
                    else:
                        request = generic_decoder.decode(data_frames)
1425

1426
1427
1428
1429
1430
1431
1432
                        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)

1433
1434
1435
                    # Push to input queue for core busy loop.
                    self.input_queue.put_nowait((request_type, request))

1436
1437
1438
    def process_output_sockets(
        self,
        output_paths: list[str],
1439
        coord_output_path: str | None,
1440
1441
        engine_index: int,
    ):
1442
1443
1444
        """Output socket IO thread."""

        # Msgpack serialization encoding.
1445
        encoder = MsgpackEncoder()
1446
1447
1448
1449
1450
1451
        # 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]]()
1452

1453
1454
        # We must set linger to ensure the ENGINE_CORE_DEAD
        # message is sent prior to closing the socket.
1455
1456
1457
        with ExitStack() as stack, zmq.Context() as ctx:
            sockets = [
                stack.enter_context(
1458
1459
                    make_zmq_socket(ctx, output_path, zmq.PUSH, linger=4000)
                )
1460
1461
                for output_path in output_paths
            ]
1462
1463
1464
1465
1466
1467
1468
1469
1470
            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
            )
1471
1472
            max_reuse_bufs = len(sockets) + 1

1473
            while True:
1474
1475
1476
1477
                output = self.output_queue.get()
                if output == EngineCoreProc.ENGINE_CORE_DEAD:
                    for socket in sockets:
                        socket.send(output)
1478
                    break
1479
1480
                assert not isinstance(output, bytes)
                client_index, outputs = output
1481
                outputs.engine_index = engine_index
1482

1483
1484
1485
1486
1487
1488
1489
                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

1490
1491
1492
1493
1494
                # 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()
1495
                buffers = encoder.encode_into(outputs, buffer)
1496
1497
1498
                tracker = sockets[client_index].send_multipart(
                    buffers, copy=False, track=True
                )
1499
1500
1501
                if not tracker.done:
                    ref = outputs if len(buffers) > 1 else None
                    pending.appendleft((tracker, ref, buffer))
1502
1503
                elif len(reuse_buffers) < max_reuse_bufs:
                    # Limit the number of buffers to reuse.
1504
                    reuse_buffers.append(buffer)
1505

1506
1507
1508
1509
1510
1511
1512
    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
        )
1513
        self._send_error_outputs_to_client([request.request_id], request.client_index)
1514

1515
1516
1517
1518
1519
    def pause_scheduler(
        self, mode: PauseMode = "abort", clear_cache: bool = True
    ) -> Future | None:
        """Pause generation; behavior depends on mode.

1520
1521
1522
1523
1524
1525
1526
1527
1528
        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
1529
1530
1531
1532
1533
          output queue is empty.
        """
        if mode not in ("keep", "abort", "wait"):
            raise ValueError(f"Invalid pause mode: {mode}")

1534
        def engine_idle_callback(engine: "EngineCoreProc", future: Future[Any]) -> None:
1535
            if clear_cache:
1536
                engine._reset_caches()
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
            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)
1547
1548
1549
1550
1551
1552
1553
1554
        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
1555

1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
    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)

1576
    def _send_abort_outputs(self, aborted_reqs: list[tuple[str, int]]) -> None:
1577
        # TODO(nick) this will be moved inside the scheduler
1578
1579
1580
1581
1582
1583
        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():
1584
                self._send_abort_outputs_to_client(list(req_ids), client_index)
1585

1586
1587
1588
1589
1590
1591
1592
1593

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

    def __init__(
        self,
        vllm_config: VllmConfig,
1594
        local_client: bool,
1595
        handshake_address: str,
1596
1597
        executor_class: type[Executor],
        log_stats: bool,
1598
        client_handshake_address: str | None = None,
1599
    ):
1600
1601
1602
1603
        assert vllm_config.model_config.is_moe, (
            "DPEngineCoreProc should only be used for MoE models"
        )

1604
1605
        # Counts forward-passes of the model so that we can synchronize
        # finished with DP peers every N steps.
1606
        self.step_counter = 0
1607
        self.current_wave = 0
Rui Qiao's avatar
Rui Qiao committed
1608
        self.last_counts = (0, 0)
1609

1610
1611
1612
1613
        from vllm.distributed.elastic_ep.elastic_state import ElasticEPScalingState

        self.eep_scaling_state: ElasticEPScalingState | None = None

1614
1615
        # Initialize the engine.
        dp_rank = vllm_config.parallel_config.data_parallel_rank
1616
1617
1618
1619
1620
1621
1622
        super().__init__(
            vllm_config,
            local_client,
            handshake_address,
            executor_class,
            log_stats,
            client_handshake_address,
1623
            engine_index=dp_rank,
1624
        )
1625
1626
1627

    def _init_data_parallel(self, vllm_config: VllmConfig):
        # Configure GPUs and stateless process group for data parallel.
1628
1629
1630
1631
        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
1632
1633

        assert dp_size > 1
1634
        assert local_dp_rank is not None
1635
1636
        assert 0 <= local_dp_rank <= dp_rank < dp_size

1637
        self.dp_rank = dp_rank
1638
1639
        dp_group, dp_store = parallel_config.stateless_init_dp_group(return_store=True)
        self.dp_group, self.dp_store = dp_group, dp_store
1640
1641
1642
1643
1644
1645

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

1646
    def add_request(self, request: Request, request_wave: int = 0):
1647
        super().add_request(request, request_wave)
1648
1649
1650
        if self.has_coordinator and request_wave != self.current_wave:
            if request_wave > self.current_wave:
                self.current_wave = request_wave
1651
1652
1653
1654
            elif not self.engines_running:
                # Request received for an already-completed wave, notify
                # front-end that we need to start the next one.
                self.output_queue.put_nowait(
1655
1656
                    (-1, EngineCoreOutputs(start_wave=self.current_wave))
                )
1657

1658
1659
    def resume_scheduler(self):
        super().resume_scheduler()
1660
1661
1662
1663
1664
        if (
            self.has_coordinator
            and not self.engines_running
            and self.scheduler.has_unfinished_requests()
        ):
1665
1666
1667
1668
            # Wake up other DP engines.
            self.output_queue.put_nowait(
                (-1, EngineCoreOutputs(start_wave=self.current_wave))
            )
1669

1670
1671
1672
    def _handle_client_request(
        self, request_type: EngineCoreRequestType, request: Any
    ) -> None:
1673
        if request_type == EngineCoreRequestType.START_DP_WAVE:
1674
1675
            new_wave, exclude_eng_index = request
            if exclude_eng_index != self.engine_index and (
1676
1677
                new_wave >= self.current_wave
            ):
1678
1679
                self.current_wave = new_wave
                if not self.engines_running:
1680
                    logger.debug("EngineCore starting idle loop for wave %d.", new_wave)
1681
1682
1683
1684
                    self.engines_running = True
        else:
            super()._handle_client_request(request_type, request)

1685
    def _maybe_publish_request_counts(self):
1686
        if not self.publish_dp_lb_stats:
1687
1688
1689
1690
1691
1692
            return

        # Publish our request counts (if they've changed).
        counts = self.scheduler.get_request_counts()
        if counts != self.last_counts:
            self.last_counts = counts
1693
1694
1695
1696
            stats = SchedulerStats(
                *counts, step_counter=self.step_counter, current_wave=self.current_wave
            )
            self.output_queue.put_nowait((-1, EngineCoreOutputs(scheduler_stats=stats)))
1697

1698
1699
1700
1701
    def run_busy_loop(self):
        """Core busy loop of the EngineCore for data parallel case."""

        # Loop until process is sent a SIGINT or SIGTERM
1702
        while self._handle_shutdown():
1703
1704
1705
            # 1) Poll the input queue until there is work to do.
            self._process_input_queue()

1706
1707
1708
1709
1710
1711
            if self.eep_scaling_state is not None:
                _ = self.eep_scaling_state.progress()
                if self.eep_scaling_state.is_complete():
                    self.process_input_queue_block = True
                    self.eep_scaling_state = None

1712
            executed = self._process_engine_step()
1713
            self._maybe_publish_request_counts()
1714

1715
            local_unfinished_reqs = self.scheduler.has_unfinished_requests()
1716
1717
            if not executed:
                if not local_unfinished_reqs and not self.engines_running:
1718
1719
1720
                    # All engines are idle.
                    continue

1721
1722
                # We are in a running state and so must execute a dummy pass
                # if the model didn't execute any ready requests.
1723
1724
1725
                self.execute_dummy_batch()

            # 3) All-reduce operation to determine global unfinished reqs.
1726
            self.engines_running = self._has_global_unfinished_reqs(
1727
1728
                local_unfinished_reqs
            )
1729

1730
            if not self.engines_running:
1731
                if self.dp_rank == 0 or not self.has_coordinator:
1732
                    # Notify client that we are pausing the loop.
1733
1734
1735
                    logger.debug(
                        "Wave %d finished, pausing engine loop.", self.current_wave
                    )
1736
1737
1738
1739
                    # 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
1740
                    self.output_queue.put_nowait(
1741
1742
1743
1744
1745
                        (
                            client_index,
                            EngineCoreOutputs(wave_complete=self.current_wave),
                        )
                    )
1746
                # Increment wave count and reset step counter.
1747
                self.current_wave += 1
1748
                self.step_counter = 0
1749

1750
1751
        raise SystemExit

1752
    def _has_global_unfinished_reqs(self, local_unfinished: bool) -> bool:
1753
        # Optimization - only perform finish-sync all-reduce every 32 steps.
1754
1755
        self.step_counter += 1
        if self.step_counter % 32 != 0:
1756
1757
            return True

1758
        return ParallelConfig.has_unfinished_dp(self.dp_group, local_unfinished)
Rui Qiao's avatar
Rui Qiao committed
1759

1760
    def reinitialize_distributed(
1761
1762
        self, reconfig_request: ReconfigureDistributedRequest
    ) -> None:
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
        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 = (
1778
            reconfig_request.new_data_parallel_master_ip
1779
        )
1780
        new_parallel_config.data_parallel_master_port = (
1781
            reconfig_request.new_data_parallel_master_port
1782
        )
1783
1784
        new_parallel_config._data_parallel_master_port_list = (
            reconfig_request.new_data_parallel_master_port_list
1785
        )
1786

1787
1788
        is_scale_down = reconfig_request.new_data_parallel_size < old_dp_size
        is_shutdown = (
1789
1790
            reconfig_request.new_data_parallel_rank
            == ReconfigureRankType.SHUTDOWN_CURRENT_RANK
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
        )

        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
1815
        1) In scale up: new core engines to notify existing core engines
1816
1817
1818
           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
1819
        3) Both scale up/down: to notify EngineCoreClient that existing
1820
1821
1822
1823
           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
1824
        else:
1825
1826
1827
1828
1829
1830
            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),
1831
            )
1832
1833
1834
1835
1836
1837
1838
1839
1840
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
1881
1882
1883
1884
1885
        )
        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,
        )
        self.model_executor.collective_rpc("init_device")
        self.model_executor.collective_rpc("load_model")
        self._eep_send_engine_core_notification(
            EEPNotificationType.NEW_CORE_ENGINES_WEIGHTS_INIT_READY
        )
        self.model_executor.collective_rpc(
            "elastic_ep_execute", args=("receive_weights",)
        )
        self.available_gpu_memory_for_kv_cache = (
            ParallelConfig.sync_kv_cache_memory_size(self.dp_group, -1)
        )
        self.model_executor.collective_rpc(
            "elastic_ep_execute", args=("prepare_new_worker",)
        )
        self.process_input_queue_block = False
1886

Rui Qiao's avatar
Rui Qiao committed
1887

1888
class EngineCoreActorMixin:
Rui Qiao's avatar
Rui Qiao committed
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
    """
    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,
    ):
1900
1901
1902
1903
1904
1905
1906
        # 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
1907
        self.addresses = addresses
1908
        vllm_config.parallel_config.data_parallel_index = dp_rank
1909
        vllm_config.parallel_config.data_parallel_rank_local = local_dp_rank
Rui Qiao's avatar
Rui Qiao committed
1910

1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
        # 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
1921
1922
1923
1924
1925
1926
1927
        # 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.
1928
        self._set_visible_devices(vllm_config, local_dp_rank)
Rui Qiao's avatar
Rui Qiao committed
1929

1930
    def _set_visible_devices(self, vllm_config: VllmConfig, local_dp_rank: int):
1931
        from vllm.platforms import current_platform
1932

1933
1934
1935
1936
        if current_platform.is_xpu():
            pass
        else:
            device_control_env_var = current_platform.device_control_env_var
1937
1938
1939
            self._set_cuda_visible_devices(
                vllm_config, local_dp_rank, device_control_env_var
            )
1940

1941
1942
1943
    def _set_cuda_visible_devices(
        self, vllm_config: VllmConfig, local_dp_rank: int, device_control_env_var: str
    ):
1944
1945
1946
        world_size = vllm_config.parallel_config.world_size
        # Set CUDA_VISIBLE_DEVICES or equivalent.
        try:
1947
1948
1949
            value = get_device_indices(
                device_control_env_var, local_dp_rank, world_size
            )
1950
            os.environ[device_control_env_var] = value
1951
1952
1953
1954
1955
        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}) "
1956
1957
                f'base value: "{os.getenv(device_control_env_var)}"'
            ) from e
1958

Rui Qiao's avatar
Rui Qiao committed
1959
    @contextmanager
1960
1961
1962
1963
1964
1965
    def _perform_handshakes(
        self,
        handshake_address: str,
        identity: bytes,
        local_client: bool,
        vllm_config: VllmConfig,
1966
        client_handshake_address: str | None,
1967
    ):
Rui Qiao's avatar
Rui Qiao committed
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
        """
        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:
1990
            self.run_busy_loop()  # type: ignore[attr-defined]
Rui Qiao's avatar
Rui Qiao committed
1991
1992
1993
1994
1995
1996
1997
        except SystemExit:
            logger.debug("EngineCore exiting.")
            raise
        except Exception:
            logger.exception("EngineCore encountered a fatal error.")
            raise
        finally:
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
2050
2051
2052
            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,
        )