core.py 42.7 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
import sys
7
8
import threading
import time
9
from collections import deque
Rui Qiao's avatar
Rui Qiao committed
10
from collections.abc import Generator
11
from concurrent.futures import Future
Rui Qiao's avatar
Rui Qiao committed
12
from contextlib import ExitStack, contextmanager
13
from inspect import isclass, signature
14
from logging import DEBUG
15
from typing import Any, Callable, Optional, TypeVar, Union
16

17
import msgspec
18
19
import zmq

20
21
22
from vllm.config import ParallelConfig, VllmConfig
from vllm.distributed import stateless_destroy_torch_distributed_process_group
from vllm.executor.multiproc_worker_utils import _add_prefix
23
from vllm.logger import init_logger
24
from vllm.logging_utils.dump_input import dump_engine_exception
25
from vllm.lora.request import LoRARequest
26
27
from vllm.transformers_utils.config import (
    maybe_register_config_serialize_by_value)
28
from vllm.utils import make_zmq_socket, resolve_obj_by_qualname
29
30
from vllm.v1.core.kv_cache_utils import (get_kv_cache_config,
                                         unify_kv_cache_configs)
31
from vllm.v1.core.sched.interface import SchedulerInterface
32
33
from vllm.v1.core.sched.output import SchedulerOutput
from vllm.v1.core.sched.scheduler import Scheduler as V1Scheduler
34
from vllm.v1.engine import (EngineCoreOutputs, EngineCoreRequest,
35
                            EngineCoreRequestType, UtilityOutput)
36
from vllm.v1.engine.mm_input_cache import MirroredProcessingCache
37
from vllm.v1.engine.utils import EngineHandshakeMetadata, EngineZmqAddresses
38
from vllm.v1.executor.abstract import Executor
39
from vllm.v1.kv_cache_interface import KVCacheConfig
40
from vllm.v1.metrics.stats import SchedulerStats
41
from vllm.v1.outputs import ModelRunnerOutput
42
from vllm.v1.request import Request, RequestStatus
43
from vllm.v1.serial_utils import MsgpackDecoder, MsgpackEncoder
44
from vllm.v1.structured_output import StructuredOutputManager
45
46
47
48
from vllm.version import __version__ as VLLM_VERSION

logger = init_logger(__name__)

49
POLLING_TIMEOUT_S = 2.5
50
HANDSHAKE_TIMEOUT_MINS = 5
51

52
53
_R = TypeVar('_R')  # Return type for collective_rpc

54
55
56
57

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

58
59
60
61
62
    def __init__(self,
                 vllm_config: VllmConfig,
                 executor_class: type[Executor],
                 log_stats: bool,
                 executor_fail_callback: Optional[Callable] = None):
63

64
65
66
67
        # plugins need to be loaded at the engine/scheduler level too
        from vllm.plugins import load_general_plugins
        load_general_plugins()

68
        self.vllm_config = vllm_config
69
        logger.info("Initializing a V1 LLM engine (v%s) with config: %s",
70
71
                    VLLM_VERSION, vllm_config)

72
73
        self.log_stats = log_stats

74
75
        # Setup Model.
        self.model_executor = executor_class(vllm_config)
76
77
78
        if executor_fail_callback is not None:
            self.model_executor.register_failure_callback(
                executor_fail_callback)
79
80

        # Setup KV Caches and update CacheConfig after profiling.
81
82
83
        num_gpu_blocks, num_cpu_blocks, kv_cache_config = \
            self._initialize_kv_caches(vllm_config)

84
85
        vllm_config.cache_config.num_gpu_blocks = num_gpu_blocks
        vllm_config.cache_config.num_cpu_blocks = num_cpu_blocks
86
87
        self.collective_rpc("initialize_cache",
                            args=(num_gpu_blocks, num_cpu_blocks))
88

89
90
        self.structured_output_manager = StructuredOutputManager(vllm_config)

91
        # Setup scheduler.
92
        if isinstance(vllm_config.scheduler_config.scheduler_cls, str):
93
94
95
96
97
98
99
100
101
            Scheduler = resolve_obj_by_qualname(
                vllm_config.scheduler_config.scheduler_cls)
        else:
            Scheduler = vllm_config.scheduler_config.scheduler_cls

        # This warning can be removed once the V1 Scheduler interface is
        # finalized and we can maintain support for scheduler classes that
        # implement it
        if Scheduler is not V1Scheduler:
102
103
104
105
106
            logger.warning(
                "Using configured V1 scheduler class %s. "
                "This scheduler interface is not public and "
                "compatibility may not be maintained.",
                vllm_config.scheduler_config.scheduler_cls)
107

108
        self.scheduler: SchedulerInterface = Scheduler(
109
            vllm_config=vllm_config,
110
111
            kv_cache_config=kv_cache_config,
            structured_output_manager=self.structured_output_manager,
112
113
            include_finished_set=vllm_config.parallel_config.data_parallel_size
            > 1,
114
            log_stats=self.log_stats,
115
        )
116

117
        # Setup MM Input Mapper.
118
        self.mm_input_cache_server = MirroredProcessingCache(
119
            vllm_config.model_config)
120

121
122
123
124
125
        # 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
126
        self.batch_queue: Optional[queue.Queue[tuple[Future[ModelRunnerOutput],
127
128
129
130
131
132
                                                     SchedulerOutput]]] = None
        if self.batch_queue_size > 1:
            logger.info("Batch queue is enabled with size %d",
                        self.batch_queue_size)
            self.batch_queue = queue.Queue(self.batch_queue_size)

133
134
    def _initialize_kv_caches(
            self, vllm_config: VllmConfig) -> tuple[int, int, KVCacheConfig]:
135
        start = time.time()
136

137
        # Get all kv cache needed by the model
138
        kv_cache_specs = self.model_executor.get_kv_cache_specs()
139
140
141

        # Profiles the peak memory usage of the model to determine how much
        # memory can be allocated for kv cache.
142
        available_gpu_memory = self.model_executor.determine_available_memory()
143

144
        assert len(kv_cache_specs) == len(available_gpu_memory)
145
        # Get the kv cache tensor size
146
147
148
149
150
151
152
153
154
155
156
157
158
        kv_cache_configs = [
            get_kv_cache_config(vllm_config, kv_cache_spec_one_worker,
                                available_gpu_memory_one_worker)
            for kv_cache_spec_one_worker, available_gpu_memory_one_worker in
            zip(kv_cache_specs, available_gpu_memory)
        ]

        # Since we use a shared centralized controller, we need the
        # `kv_cache_config` to be consistent across all workers to make sure
        # all the memory operators can be applied to all workers.
        unify_kv_cache_configs(kv_cache_configs)

        # All workers have the same kv_cache_config except layer names, so use
159
        # an arbitrary one to initialize the scheduler.
160
161
162
163
164
        assert all([
            cfg.num_blocks == kv_cache_configs[0].num_blocks
            for cfg in kv_cache_configs
        ])
        num_gpu_blocks = kv_cache_configs[0].num_blocks
165
        num_cpu_blocks = 0
166
        scheduler_kv_cache_config = kv_cache_configs[0]
167
168

        # Initialize kv cache and warmup the execution
169
        self.model_executor.initialize_from_config(kv_cache_configs)
170

171
172
173
        elapsed = time.time() - start
        logger.info(("init engine (profile, create kv cache, "
                     "warmup model) took %.2f seconds"), elapsed)
174
        return num_gpu_blocks, num_cpu_blocks, scheduler_kv_cache_config
175
176
177

    def add_request(self, request: EngineCoreRequest):
        """Add request to the scheduler."""
178
179

        if request.mm_hashes is not None:
180
181
182
183
184
            # Here, if hash exists for a multimodal input, then it will be
            # fetched from the cache, else it will be added to the cache.
            # Note that the cache here is mirrored with the client cache, so
            # anything that has a hash must have a HIT cache entry here
            # as well.
185
            assert request.mm_inputs is not None
186
            request.mm_inputs = self.mm_input_cache_server.get_and_update_p1(
187
                request.mm_inputs, request.mm_hashes)
188

189
        req = Request.from_engine_core_request(request)
190
191
        if req.use_structured_output:
            # Start grammar compilation asynchronously
192
            self.structured_output_manager.grammar_init(req)
193

194
195
196
197
        if req.kv_transfer_params is not None and (
                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
198

199
200
        self.scheduler.add_request(req)

201
    def abort_requests(self, request_ids: list[str]):
202
203
204
205
206
207
208
209
        """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).
        self.scheduler.finish_requests(request_ids,
                                       RequestStatus.FINISHED_ABORTED)

210
211
212
    def execute_model(self, scheduler_output: SchedulerOutput):
        try:
            return self.model_executor.execute_model(scheduler_output)
213
214
215
216
217
        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.

218
219
220
221
222
            # NOTE: This method is exception-free
            dump_engine_exception(self.vllm_config, scheduler_output,
                                  self.scheduler.make_stats())
            raise err

223
    def step(self) -> tuple[dict[int, EngineCoreOutputs], bool]:
224
225
226
227
228
        """Schedule, execute, and make output.

        Returns tuple of outputs and a flag indicating whether the model
        was executed.
        """
229

230
231
232
        # Check for any requests remaining in the scheduler - unfinished,
        # or finished and not yet removed from the batch.
        if not self.scheduler.has_requests():
233
            return {}, False
234
        scheduler_output = self.scheduler.schedule()
235
        model_output = self.execute_model(scheduler_output)
236
        engine_core_outputs = self.scheduler.update_from_output(
237
            scheduler_output, model_output)  # type: ignore
238

239
240
        return (engine_core_outputs,
                scheduler_output.total_num_scheduled_tokens > 0)
241

242
    def step_with_batch_queue(
243
            self) -> tuple[Optional[dict[int, EngineCoreOutputs]], bool]:
244
245
246
247
        """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:
248
249
250
251
        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.
252
253
254
255
256
257
258
259
260
        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.
        """
        assert self.batch_queue is not None

        engine_core_outputs = None
        scheduler_output = None
261
262
263
264
        # 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.
        if not self.batch_queue.full():
265
266
267
268
269
270
            scheduler_output = self.scheduler.schedule()
            if scheduler_output.total_num_scheduled_tokens > 0:
                future = self.model_executor.execute_model(scheduler_output)
                self.batch_queue.put_nowait(
                    (future, scheduler_output))  # type: ignore

271
272
273
274
        scheduled_batch = (scheduler_output is not None
                           and scheduler_output.total_num_scheduled_tokens > 0)

        # If no more requests can be scheduled and the job queue is not empty,
275
        # block until the first batch in the job queue is finished.
276
277
278
279
        # TODO(comaniac): Ideally we should peek the first batch in the
        # job queue to check if it's finished before scheduling a new batch,
        # but peeking the first element in a queue is not thread-safe,
        # so we need more work.
280
281
282
283
284
        if not scheduled_batch and not self.batch_queue.empty():
            future, scheduler_output = self.batch_queue.get_nowait()
            # Blocking until the first result is available.
            model_output = future.result()
            self.batch_queue.task_done()
285
286
            engine_core_outputs = (self.scheduler.update_from_output(
                scheduler_output, model_output))
287

288
        return engine_core_outputs, scheduled_batch
289

290
    def shutdown(self):
291
        self.structured_output_manager.clear_backend()
292
293
        if self.model_executor:
            self.model_executor.shutdown()
294
295
        if self.scheduler:
            self.scheduler.shutdown()
296

297
    def profile(self, is_start: bool = True):
298
        self.model_executor.profile(is_start)
299

300
301
302
    def reset_mm_cache(self):
        # NOTE: Since this is mainly for debugging, we don't attempt to
        # re-sync the internal caches (P0 processor, P0 mirror, P1 mirror)
303
        if self.scheduler.has_unfinished_requests():
304
305
306
307
308
            logger.warning("Resetting the multi-modal cache when requests are "
                           "in progress may lead to desynced internal caches.")

        self.mm_input_cache_server.reset()

309
310
311
    def reset_prefix_cache(self):
        self.scheduler.reset_prefix_cache()

312
313
314
    def sleep(self, level: int = 1):
        self.model_executor.sleep(level)

315
316
    def wake_up(self, tags: Optional[list[str]] = None):
        self.model_executor.wake_up(tags)
317

318
319
320
    def is_sleeping(self) -> bool:
        return self.model_executor.is_sleeping

321
322
323
    def execute_dummy_batch(self):
        self.model_executor.collective_rpc("execute_dummy_batch")

324
325
326
327
328
329
    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)

330
    def list_loras(self) -> set[int]:
331
332
333
334
        return self.model_executor.list_loras()

    def pin_lora(self, lora_id: int) -> bool:
        return self.model_executor.pin_lora(lora_id)
335

336
337
338
339
340
341
342
343
344
345
    def save_sharded_state(
        self,
        path: str,
        pattern: Optional[str] = None,
        max_size: Optional[int] = None,
    ) -> None:
        self.model_executor.save_sharded_state(path=path,
                                               pattern=pattern,
                                               max_size=max_size)

346
347
348
349
350
351
352
353
    def collective_rpc(self,
                       method: Union[str, Callable[..., _R]],
                       timeout: Optional[float] = None,
                       args: tuple = (),
                       kwargs: Optional[dict[str, Any]] = None) -> list[_R]:
        return self.model_executor.collective_rpc(method, timeout, args,
                                                  kwargs)

354
355
356
357
358
359
360
    def save_tensorized_model(
        self,
        tensorizer_config,
    ) -> None:
        self.model_executor.save_tensorized_model(
            tensorizer_config=tensorizer_config, )

361
362
363
364

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

365
366
    ENGINE_CORE_DEAD = b'ENGINE_CORE_DEAD'

367
368
    def __init__(
        self,
369
        vllm_config: VllmConfig,
370
        local_client: bool,
371
        handshake_address: str,
372
        executor_class: type[Executor],
373
        log_stats: bool,
374
        client_handshake_address: Optional[str] = None,
375
        engine_index: int = 0,
376
    ):
Rui Qiao's avatar
Rui Qiao committed
377
378
379
380
        self.input_queue = queue.Queue[tuple[EngineCoreRequestType, Any]]()
        self.output_queue = queue.Queue[Union[tuple[int, EngineCoreOutputs],
                                              bytes]]()
        executor_fail_callback = lambda: self.input_queue.put_nowait(
381
382
            (EngineCoreRequestType.EXECUTOR_FAILED, b''))

Rui Qiao's avatar
Rui Qiao committed
383
384
385
        self.engine_index = engine_index
        identity = self.engine_index.to_bytes(length=2, byteorder="little")
        self.engines_running = False
386

387
388
389
        with self._perform_handshakes(handshake_address, identity,
                                      local_client, vllm_config,
                                      client_handshake_address) as addresses:
390
            self.client_count = len(addresses.outputs)
391
392

            # Set up data parallel environment.
393
            self.has_coordinator = addresses.coordinator_output is not None
394
395
396
397
398
399
400
401
            self.frontend_stats_publish_address = (
                addresses.frontend_stats_publish_address)
            # Only publish request queue stats to coordinator for "internal"
            # LB mode.
            self.publish_dp_lb_stats = (
                self.has_coordinator
                and not vllm_config.parallel_config.data_parallel_external_lb)

402
403
404
405
406
            self._init_data_parallel(vllm_config)

            super().__init__(vllm_config, executor_class, log_stats,
                             executor_fail_callback)

Rui Qiao's avatar
Rui Qiao committed
407
408
        self.step_fn = (self.step if self.batch_queue is None else
                        self.step_with_batch_queue)
409

410
411
412
413
414
415
416
417
418
419
420
421
        # 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.
        threading.Thread(target=self.process_input_sockets,
                         args=(addresses.inputs, addresses.coordinator_input,
                               identity),
                         daemon=True).start()
        self.output_thread = threading.Thread(
            target=self.process_output_sockets,
            args=(addresses.outputs, addresses.coordinator_output,
Rui Qiao's avatar
Rui Qiao committed
422
                  self.engine_index),
423
424
            daemon=True)
        self.output_thread.start()
425

Rui Qiao's avatar
Rui Qiao committed
426
    @contextmanager
427
428
429
430
431
432
433
    def _perform_handshakes(
        self,
        handshake_address: str,
        identity: bytes,
        local_client: bool,
        vllm_config: VllmConfig,
        client_handshake_address: Optional[str],
Rui Qiao's avatar
Rui Qiao committed
434
    ) -> Generator[EngineZmqAddresses, None, None]:
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
        """
        Perform startup handshakes.

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

        For DP>1 with internal loadbalancing this is with the shared front-end
        process which may reside on a different node.

        For DP>1 with external loadbalancing, two handshakes are performed:
            - 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.
        with the exception of the rank 0 engine itself which doesn't require
        the second handshake.

        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
456
        input_ctx = zmq.Context()
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
        is_local = local_client and client_handshake_address is None
        handshake = self._perform_handshake(input_ctx, handshake_address,
                                            identity, is_local, vllm_config,
                                            vllm_config.parallel_config)
        if client_handshake_address is None:
            with handshake as addresses:
                yield addresses
        else:
            local_handshake = self._perform_handshake(
                input_ctx, client_handshake_address, identity, local_client,
                vllm_config)
            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,
        vllm_config: VllmConfig,
        parallel_config_to_update: Optional[ParallelConfig] = None,
    ) -> Generator[EngineZmqAddresses, None, None]:
        with make_zmq_socket(ctx,
Rui Qiao's avatar
Rui Qiao committed
487
488
489
490
491
492
                             handshake_address,
                             zmq.DEALER,
                             identity=identity,
                             linger=5000,
                             bind=False) as handshake_socket:
            # Register engine with front-end.
493
494
            addresses = self.startup_handshake(handshake_socket, local_client,
                                               parallel_config_to_update)
Rui Qiao's avatar
Rui Qiao committed
495
496
497
498
            yield addresses

            # Send ready message.
            num_gpu_blocks = vllm_config.cache_config.num_gpu_blocks
499
500
501
502
            # 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
Rui Qiao's avatar
Rui Qiao committed
503
504
505
            handshake_socket.send(
                msgspec.msgpack.encode({
                    "status": "READY",
506
                    "local": local_client,
Rui Qiao's avatar
Rui Qiao committed
507
                    "num_gpu_blocks": num_gpu_blocks,
508
                    "dp_stats_address": dp_stats_address,
Rui Qiao's avatar
Rui Qiao committed
509
510
                }))

511
    @staticmethod
512
    def startup_handshake(
513
514
515
516
        handshake_socket: zmq.Socket,
        local_client: bool,
        parallel_config: Optional[ParallelConfig] = None,
    ) -> EngineZmqAddresses:
517
518

        # Send registration message.
519
        handshake_socket.send(
520
521
            msgspec.msgpack.encode({
                "status": "HELLO",
522
                "local": local_client,
523
524
525
526
            }))

        # Receive initialization message.
        logger.info("Waiting for init message from front-end.")
527
        if not handshake_socket.poll(timeout=HANDSHAKE_TIMEOUT_MINS * 60_000):
528
529
530
            raise RuntimeError("Did not receive response from front-end "
                               f"process within {HANDSHAKE_TIMEOUT_MINS} "
                               f"minutes")
531
532
533
        init_bytes = handshake_socket.recv()
        init_message: EngineHandshakeMetadata = msgspec.msgpack.decode(
            init_bytes, type=EngineHandshakeMetadata)
534
535
        logger.debug("Received init message: %s", init_message)

536
537
538
        if parallel_config is not None:
            for key, value in init_message.parallel_config.items():
                setattr(parallel_config, key, value)
539

540
        return init_message.addresses
541
542

    @staticmethod
543
544
545
546
    def run_engine_core(*args,
                        dp_rank: int = 0,
                        local_dp_rank: int = 0,
                        **kwargs):
547
548
        """Launch EngineCore busy loop in background process."""

549
550
551
552
553
        # Signal handler used for graceful termination.
        # SystemExit exception is only raised once to allow this and worker
        # processes to terminate without error
        shutdown_requested = False

554
555
556
        # Ensure we can serialize transformer config after spawning
        maybe_register_config_serialize_by_value()

557
558
559
560
561
562
563
564
565
566
        def signal_handler(signum, frame):
            nonlocal shutdown_requested
            if not shutdown_requested:
                shutdown_requested = True
                raise SystemExit()

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

567
        engine_core: Optional[EngineCoreProc] = None
568
        try:
569
570
            parallel_config: ParallelConfig = kwargs[
                "vllm_config"].parallel_config
571
            if parallel_config.data_parallel_size > 1 or dp_rank > 0:
572
573
574
575
576
577
578
                # Set data parallel rank for this engine process.
                parallel_config.data_parallel_rank = dp_rank
                parallel_config.data_parallel_rank_local = local_dp_rank
                engine_core = DPEngineCoreProc(*args, **kwargs)
            else:
                engine_core = EngineCoreProc(*args, **kwargs)

579
580
            engine_core.run_busy_loop()

581
        except SystemExit:
582
            logger.debug("EngineCore exiting.")
583
            raise
584
585
586
587
588
589
590
        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
591
592
593
594
        finally:
            if engine_core is not None:
                engine_core.shutdown()

595
596
597
    def _init_data_parallel(self, vllm_config: VllmConfig):
        pass

598
599
600
    def run_busy_loop(self):
        """Core busy loop of the EngineCore."""

601
602
        # Loop until process is sent a SIGINT or SIGTERM
        while True:
603
            # 1) Poll the input queue until there is work to do.
604
605
606
607
608
609
610
611
            self._process_input_queue()
            # 2) Step the engine core and return the outputs.
            self._process_engine_step()

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

        waited = False
612
        while not self.engines_running and not self.scheduler.has_requests():
613
614
615
616
617
618
619
            if logger.isEnabledFor(DEBUG) and self.input_queue.empty():
                logger.debug("EngineCore waiting for work.")
                waited = True
            req = self.input_queue.get()
            self._handle_client_request(*req)

        if waited:
620
            logger.debug("EngineCore loop active.")
621
622
623
624
625
626

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

627
    def _process_engine_step(self) -> bool:
628
629
630
        """Called only when there are unfinished local requests."""

        # Step the engine core.
631
        outputs, model_executed = self.step_fn()
632
        # Put EngineCoreOutputs into the output queue.
633
634
        for output in (outputs.items() if outputs else ()):
            self.output_queue.put_nowait(output)
635

636
637
        return model_executed

638
639
640
    def _handle_client_request(self, request_type: EngineCoreRequestType,
                               request: Any) -> None:
        """Dispatch request from client."""
641

642
        if request_type == EngineCoreRequestType.ADD:
643
            self.add_request(request)
644
        elif request_type == EngineCoreRequestType.ABORT:
645
            self.abort_requests(request)
646
        elif request_type == EngineCoreRequestType.UTILITY:
647
            client_idx, call_id, method_name, args = request
648
649
650
651
652
653
654
655
656
657
            output = UtilityOutput(call_id)
            try:
                method = getattr(self, method_name)
                output.result = method(
                    *self._convert_msgspec_args(method, args))
            except BaseException as e:
                logger.exception("Invocation of %s method failed", method_name)
                output.failure_message = (f"Call to {method_name} method"
                                          f" failed: {str(e)}")
            self.output_queue.put_nowait(
658
                (client_idx, EngineCoreOutputs(utility_output=output)))
659
660
661
662
663
        elif request_type == EngineCoreRequestType.EXECUTOR_FAILED:
            raise RuntimeError("Executor failed.")
        else:
            logger.error("Unrecognized input request type encountered: %s",
                         request_type)
664
665
666
667
668
669
670
671
672
673
674
675
676
677

    @staticmethod
    def _convert_msgspec_args(method, args):
        """If a provided arg type doesn't match corresponding target method
         arg type, try converting to msgspec object."""
        if not args:
            return args
        arg_types = signature(method).parameters.values()
        assert len(args) <= len(arg_types)
        return tuple(
            msgspec.convert(v, type=p.annotation) if isclass(p.annotation)
            and issubclass(p.annotation, msgspec.Struct)
            and not isinstance(v, p.annotation) else v
            for v, p in zip(args, arg_types))
678

679
680
681
682
683
684
685
686
687
688
689
690
    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():
            logger.fatal("vLLM shutdown signal from EngineCore failed "
                         "to send. Please report this issue.")

691
692
693
    def process_input_sockets(self, input_addresses: list[str],
                              coord_input_address: Optional[str],
                              identity: bytes):
694
695
696
        """Input socket IO thread."""

        # Msgpack serialization decoding.
697
698
        add_request_decoder = MsgpackDecoder(EngineCoreRequest)
        generic_decoder = MsgpackDecoder()
699

700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
        with ExitStack() as stack, zmq.Context() as ctx:
            input_sockets = [
                stack.enter_context(
                    make_zmq_socket(ctx,
                                    input_address,
                                    zmq.DEALER,
                                    identity=identity,
                                    bind=False))
                for input_address in input_addresses
            ]
            if coord_input_address is None:
                coord_socket = None
            else:
                coord_socket = stack.enter_context(
                    make_zmq_socket(ctx,
                                    coord_input_address,
                                    zmq.XSUB,
                                    identity=identity,
                                    bind=False))
                # Send subscription message to coordinator.
                coord_socket.send(b'\x01')

            # 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.
                input_socket.send(b'')
                poller.register(input_socket, zmq.POLLIN)
            if coord_socket is not None:
                poller.register(coord_socket, zmq.POLLIN)
732

733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
            while True:
                for input_socket, _ in poller.poll():
                    # (RequestType, RequestData)
                    type_frame, *data_frames = input_socket.recv_multipart(
                        copy=False)
                    request_type = EngineCoreRequestType(
                        bytes(type_frame.buffer))

                    # Deserialize the request data.
                    decoder = add_request_decoder if (
                        request_type
                        == EngineCoreRequestType.ADD) else generic_decoder
                    request = decoder.decode(data_frames)

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

    def process_output_sockets(self, output_paths: list[str],
                               coord_output_path: Optional[str],
                               engine_index: int):
753
754
755
        """Output socket IO thread."""

        # Msgpack serialization encoding.
756
        encoder = MsgpackEncoder()
757
758
759
760
761
762
        # 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]]()
763

764
765
        # We must set linger to ensure the ENGINE_CORE_DEAD
        # message is sent prior to closing the socket.
766
767
768
769
770
771
772
773
774
775
776
777
        with ExitStack() as stack, zmq.Context() as ctx:
            sockets = [
                stack.enter_context(
                    make_zmq_socket(ctx, output_path, zmq.PUSH, linger=4000))
                for output_path in output_paths
            ]
            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
            max_reuse_bufs = len(sockets) + 1

778
            while True:
779
780
781
782
                output = self.output_queue.get()
                if output == EngineCoreProc.ENGINE_CORE_DEAD:
                    for socket in sockets:
                        socket.send(output)
783
                    break
784
785
                assert not isinstance(output, bytes)
                client_index, outputs = output
786
                outputs.engine_index = engine_index
787

788
789
790
791
792
793
794
                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

795
796
797
798
799
                # 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()
800
                buffers = encoder.encode_into(outputs, buffer)
801
802
803
                tracker = sockets[client_index].send_multipart(buffers,
                                                               copy=False,
                                                               track=True)
804
805
806
                if not tracker.done:
                    ref = outputs if len(buffers) > 1 else None
                    pending.appendleft((tracker, ref, buffer))
807
808
                elif len(reuse_buffers) < max_reuse_bufs:
                    # Limit the number of buffers to reuse.
809
                    reuse_buffers.append(buffer)
810
811
812
813
814
815
816
817
818


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

    def __init__(
        self,
        vllm_config: VllmConfig,
819
        local_client: bool,
820
        handshake_address: str,
821
822
        executor_class: type[Executor],
        log_stats: bool,
823
        client_handshake_address: Optional[str] = None,
824
    ):
Rui Qiao's avatar
Rui Qiao committed
825
        self._decorate_logs()
826

827
828
829
        # Counts forward-passes of the model so that we can synchronize
        # finished with DP peers every N steps.
        self.counter = 0
830
        self.current_wave = 0
Rui Qiao's avatar
Rui Qiao committed
831
        self.last_counts = (0, 0)
832
833
834

        # Initialize the engine.
        dp_rank = vllm_config.parallel_config.data_parallel_rank
835
836
837
        super().__init__(vllm_config, local_client, handshake_address,
                         executor_class, log_stats, client_handshake_address,
                         dp_rank)
838

Rui Qiao's avatar
Rui Qiao committed
839
840
841
842
843
844
845
846
847
    def _decorate_logs(self):
        # Add process-specific prefix to stdout and stderr before
        # we initialize the engine.
        from multiprocessing import current_process
        process_name = current_process().name
        pid = os.getpid()
        _add_prefix(sys.stdout, process_name, pid)
        _add_prefix(sys.stderr, process_name, pid)

848
849
850
    def _init_data_parallel(self, vllm_config: VllmConfig):

        # Configure GPUs and stateless process group for data parallel.
851
        dp_rank = vllm_config.parallel_config.data_parallel_rank
852
        dp_size = vllm_config.parallel_config.data_parallel_size
853
854
855
856
857
        local_dp_rank = vllm_config.parallel_config.data_parallel_rank_local

        assert dp_size > 1
        assert 0 <= local_dp_rank <= dp_rank < dp_size

858
859
860
861
862
863
864
865
866
        if vllm_config.kv_transfer_config is not None:
            # modify the engine_id and append the local_dp_rank to it to ensure
            # that the kv_transfer_config is unique for each DP rank.
            vllm_config.kv_transfer_config.engine_id = (
                f"{vllm_config.kv_transfer_config.engine_id}_dp{local_dp_rank}"
            )
            logger.debug("Setting kv_transfer_config.engine_id to %s",
                         vllm_config.kv_transfer_config.engine_id)

867
        from vllm.platforms import current_platform
868
        device_control_env_var = current_platform.device_control_env_var
869
        world_size = vllm_config.parallel_config.world_size
870
871
872
873
874
875
876
877
878
879
880
881
        # Set CUDA_VISIBLE_DEVICES or equivalent.
        try:
            os.environ[device_control_env_var] = ",".join(
                str(current_platform.device_id_to_physical_device_id(i))
                for i in range(local_dp_rank *
                               world_size, (local_dp_rank + 1) * world_size))
        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}) "
                f"base value: \"{os.getenv(device_control_env_var)}\"") from e
882

883
        self.dp_rank = dp_rank
884
885
886
887
888
889
890
        self.dp_group = vllm_config.parallel_config.stateless_init_dp_group()

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

891
    def add_request(self, request: EngineCoreRequest):
892
        if self.has_coordinator and request.current_wave != self.current_wave:
893
894
895
896
897
898
            if request.current_wave > self.current_wave:
                self.current_wave = request.current_wave
            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(
899
                    (-1, EngineCoreOutputs(start_wave=self.current_wave)))
900
901
902
903
904
905

        super().add_request(request)

    def _handle_client_request(self, request_type: EngineCoreRequestType,
                               request: Any) -> None:
        if request_type == EngineCoreRequestType.START_DP_WAVE:
906
907
908
            new_wave, exclude_eng_index = request
            if exclude_eng_index != self.engine_index and (
                    new_wave >= self.current_wave):
909
910
911
912
913
914
915
916
                self.current_wave = new_wave
                if not self.engines_running:
                    logger.debug("EngineCore starting idle loop for wave %d.",
                                 new_wave)
                    self.engines_running = True
        else:
            super()._handle_client_request(request_type, request)

917
    def _maybe_publish_request_counts(self):
918
        if not self.publish_dp_lb_stats:
919
920
921
922
923
924
925
926
927
928
            return

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

929
930
931
932
933
934
935
936
    def run_busy_loop(self):
        """Core busy loop of the EngineCore for data parallel case."""

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

937
938
            # 2) Step the engine core.
            executed = self._process_engine_step()
939
940
            self._maybe_publish_request_counts()

941
            local_unfinished_reqs = self.scheduler.has_unfinished_requests()
942
943
            if not executed:
                if not local_unfinished_reqs and not self.engines_running:
944
945
946
                    # All engines are idle.
                    continue

947
948
                # We are in a running state and so must execute a dummy pass
                # if the model didn't execute any ready requests.
949
950
951
                self.execute_dummy_batch()

            # 3) All-reduce operation to determine global unfinished reqs.
952
            self.engines_running = self._has_global_unfinished_reqs(
953
954
                local_unfinished_reqs)

955
            if not self.engines_running:
956
                if self.dp_rank == 0 or not self.has_coordinator:
957
958
959
                    # Notify client that we are pausing the loop.
                    logger.debug("Wave %d finished, pausing engine loop.",
                                 self.current_wave)
960
961
962
963
                    # 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
964
                    self.output_queue.put_nowait(
965
                        (client_index,
966
                         EngineCoreOutputs(wave_complete=self.current_wave)))
967
                self.current_wave += 1
968
969
970

    def _has_global_unfinished_reqs(self, local_unfinished: bool) -> bool:

971
        # Optimization - only perform finish-sync all-reduce every 32 steps.
972
        self.counter += 1
973
        if self.counter != 32:
974
975
976
977
978
            return True
        self.counter = 0

        return ParallelConfig.has_unfinished_dp(self.dp_group,
                                                local_unfinished)
Rui Qiao's avatar
Rui Qiao committed
979
980
981
982
983
984
985
986
987
988


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

    def __init__(
        self,
        vllm_config: VllmConfig,
989
        local_client: bool,
Rui Qiao's avatar
Rui Qiao committed
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
        addresses: EngineZmqAddresses,
        executor_class: type[Executor],
        log_stats: bool,
        dp_rank: int = 0,
        local_dp_rank: int = 0,
    ):
        self.addresses = addresses
        vllm_config.parallel_config.data_parallel_rank = dp_rank
        vllm_config.parallel_config.data_parallel_rank_local = \
            local_dp_rank

        # Ray sets CUDA_VISIBLE_DEVICES to empty string,
        # we clean this up to be able to properly initialize
        # data parallel groups.
        del os.environ['CUDA_VISIBLE_DEVICES']

1006
        super().__init__(vllm_config, local_client, "", executor_class,
Rui Qiao's avatar
Rui Qiao committed
1007
1008
1009
1010
1011
1012
                         log_stats)

    def _decorate_logs(self):
        pass

    @contextmanager
1013
1014
1015
    def _perform_handshakes(self, handshake_address: str, identity: bytes,
                            local_client: bool, vllm_config: VllmConfig,
                            client_handshake_address: Optional[str]):
Rui Qiao's avatar
Rui Qiao committed
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
        """
        For Ray, we don't need to actually perform handshake.
        All addresses information is known before the actor creation.
        Therefore, we simply yield these addresses.
        """
        yield self.addresses

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

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

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