async_llm.py 30.1 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
import asyncio
4
5
import os
import socket
6
import time
7
from collections.abc import AsyncGenerator, Iterable, Mapping
8
from copy import copy
9
from typing import Any, Optional, Union
10

11
import numpy as np
12
import torch
13

14
import vllm.envs as envs
15
16
17
from vllm.config import ModelConfig, VllmConfig
from vllm.engine.arg_utils import AsyncEngineArgs
from vllm.engine.protocol import EngineClient
18
from vllm.entrypoints.utils import _validate_truncation_size
19
from vllm.envs import VLLM_V1_OUTPUT_PROC_CHUNK_SIZE
20
from vllm.inputs import PromptType
21
from vllm.inputs.preprocess import InputPreprocessor
22
23
from vllm.logger import init_logger
from vllm.lora.request import LoRARequest
24
from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalRegistry
25
from vllm.outputs import PoolingRequestOutput, RequestOutput
26
from vllm.pooling_params import PoolingParams
27
from vllm.sampling_params import SamplingParams
28
from vllm.tasks import SupportedTask
29
from vllm.tracing import init_tracer
30
from vllm.transformers_utils.config import maybe_register_config_serialize_by_value
31
from vllm.transformers_utils.tokenizer import AnyTokenizer
32
from vllm.usage.usage_lib import UsageContext
33
from vllm.utils import Device, as_list, cancel_task_threadsafe, cdiv, deprecate_kwargs
34
from vllm.v1.engine import EngineCoreRequest
35
from vllm.v1.engine.core_client import EngineCoreClient
36
from vllm.v1.engine.exceptions import EngineDeadError, EngineGenerateError
37
from vllm.v1.engine.output_processor import OutputProcessor, RequestOutputCollector
38
from vllm.v1.engine.parallel_sampling import ParentRequest
39
from vllm.v1.engine.processor import Processor
40
from vllm.v1.executor.abstract import Executor
41
from vllm.v1.metrics.loggers import StatLoggerFactory, StatLoggerManager
42
from vllm.v1.metrics.prometheus import shutdown_prometheus
43
from vllm.v1.metrics.stats import IterationStats
44
45
46
47
48
49
50
51

logger = init_logger(__name__)


class AsyncLLM(EngineClient):
    def __init__(
        self,
        vllm_config: VllmConfig,
52
        executor_class: type[Executor],
53
54
        log_stats: bool,
        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
55
        mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
56
57
58
        use_cached_outputs: bool = False,
        log_requests: bool = True,
        start_engine_loop: bool = True,
59
        stat_loggers: Optional[list[StatLoggerFactory]] = None,
60
        client_addresses: Optional[dict[str, str]] = None,
61
        client_count: int = 1,
62
        client_index: int = 0,
63
    ) -> None:
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
        """
        Create an AsyncLLM.

        Args:
            vllm_config: global configuration.
            executor_class: an Executor impl, e.g. MultiprocExecutor.
            log_stats: Whether to log stats.
            usage_context: Usage context of the LLM.
            mm_registry: Multi-modal registry.
            use_cached_outputs: Whether to use cached outputs.
            log_requests: Whether to log requests.
            start_engine_loop: Whether to start the engine loop.
            stat_loggers: customized stat loggers for the engine.
                If not provided, default stat loggers will be used.
                PLEASE BE AWARE THAT STAT LOGGER IS NOT STABLE
                IN V1, AND ITS BASE CLASS INTERFACE MIGHT CHANGE.

        Returns:
            None
        """
84
85
86
87
88
        if not envs.VLLM_USE_V1:
            raise ValueError(
                "Using V1 AsyncLLMEngine, but envs.VLLM_USE_V1=False. "
                "This should not happen. As a workaround, try using "
                "AsyncLLMEngine.from_vllm_config(...) or explicitly set "
89
90
                "VLLM_USE_V1=0 or 1 and report this issue on Github."
            )
91

92
93
94
        # Ensure we can serialize custom transformer configs
        maybe_register_config_serialize_by_value()

95
        self.model_config = vllm_config.model_config
96
        self.vllm_config = vllm_config
97
        self.observability_config = vllm_config.observability_config
98
        self.log_requests = log_requests
99
100
101
102
103

        self.log_stats = log_stats or (stat_loggers is not None)
        if not log_stats and stat_loggers is not None:
            logger.info(
                "AsyncLLM created with log_stats=False and non-empty custom "
104
105
                "logger list; enabling logging without default stat loggers"
            )
106

107
        # Processor (converts Inputs --> EngineCoreRequests).
108
        self.processor = Processor(vllm_config, mm_registry=mm_registry)
109

110
        # OutputProcessor (converts EngineCoreOutputs --> RequestOutput).
111
112
113
        self.output_processor = OutputProcessor(
            self.tokenizer, log_stats=self.log_stats
        )
114
115
        if self.observability_config.otlp_traces_endpoint is not None:
            tracer = init_tracer(
116
117
                "vllm.llm_engine", self.observability_config.otlp_traces_endpoint
            )
118
            self.output_processor.tracer = tracer
119
120

        # EngineCore (starts the engine in background process).
121
        self.engine_core = EngineCoreClient.make_async_mp_client(
122
123
            vllm_config=vllm_config,
            executor_class=executor_class,
124
            log_stats=self.log_stats,
125
            client_addresses=client_addresses,
126
            client_count=client_count,
127
            client_index=client_index,
128
        )
129
130
131
132
133
134

        # Loggers.
        self.logger_manager: Optional[StatLoggerManager] = None
        if self.log_stats:
            self.logger_manager = StatLoggerManager(
                vllm_config=vllm_config,
135
                engine_idxs=self.engine_core.engine_ranks_managed,
136
                custom_stat_loggers=stat_loggers,
137
                enable_default_loggers=log_stats,
138
                client_count=client_count,
139
140
141
            )
            self.logger_manager.log_engine_initialized()

142
        self.output_handler: Optional[asyncio.Task] = None
143
144
145
146
147
148
        try:
            # Start output handler eagerly if we are in the asyncio eventloop.
            asyncio.get_running_loop()
            self._run_output_handler()
        except RuntimeError:
            pass
149

150
151
152
        if envs.VLLM_TORCH_PROFILER_DIR:
            logger.info(
                "Torch profiler enabled. AsyncLLM CPU traces will be collected under %s",  # noqa: E501
153
154
                envs.VLLM_TORCH_PROFILER_DIR,
            )
155
156
157
158
159
160
161
            worker_name = f"{socket.gethostname()}_{os.getpid()}.async_llm"
            self.profiler = torch.profiler.profile(
                activities=[
                    torch.profiler.ProfilerActivity.CPU,
                ],
                with_stack=envs.VLLM_TORCH_PROFILER_WITH_STACK,
                on_trace_ready=torch.profiler.tensorboard_trace_handler(
162
163
164
                    envs.VLLM_TORCH_PROFILER_DIR, worker_name=worker_name, use_gzip=True
                ),
            )
165
166
167
        else:
            self.profiler = None

168
    @classmethod
169
170
    @deprecate_kwargs(
        "disable_log_requests",
171
172
173
        additional_message=(
            "This argument will have no effect. Use `enable_log_requests` instead."
        ),
174
    )
175
    def from_vllm_config(
176
177
178
179
180
181
182
183
184
185
186
        cls,
        vllm_config: VllmConfig,
        start_engine_loop: bool = True,
        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
        stat_loggers: Optional[list[StatLoggerFactory]] = None,
        enable_log_requests: bool = False,
        disable_log_stats: bool = False,
        client_addresses: Optional[dict[str, str]] = None,
        client_count: int = 1,
        client_index: int = 0,
        disable_log_requests: bool = True,  # Deprecated, will be removed
187
188
189
190
191
192
    ) -> "AsyncLLM":
        if not envs.VLLM_USE_V1:
            raise ValueError(
                "Using V1 AsyncLLMEngine, but envs.VLLM_USE_V1=False. "
                "This should not happen. As a workaround, try using "
                "AsyncLLMEngine.from_vllm_config(...) or explicitly set "
193
194
                "VLLM_USE_V1=0 or 1 and report this issue on Github."
            )
195
196
197
198
199
200

        # Create the LLMEngine.
        return cls(
            vllm_config=vllm_config,
            executor_class=Executor.get_class(vllm_config),
            start_engine_loop=start_engine_loop,
201
            stat_loggers=stat_loggers,
202
            log_requests=enable_log_requests,
203
204
            log_stats=not disable_log_stats,
            usage_context=usage_context,
205
            client_addresses=client_addresses,
206
            client_count=client_count,
207
            client_index=client_index,
208
209
        )

210
211
212
213
214
215
    @classmethod
    def from_engine_args(
        cls,
        engine_args: AsyncEngineArgs,
        start_engine_loop: bool = True,
        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
216
        stat_loggers: Optional[list[StatLoggerFactory]] = None,
217
    ) -> "AsyncLLM":
218
219
220
        """Create an AsyncLLM from the EngineArgs."""

        # Create the engine configs.
221
        vllm_config = engine_args.create_engine_config(usage_context)
222
        executor_class = Executor.get_class(vllm_config)
223
224
225
226
227

        # Create the AsyncLLM.
        return cls(
            vllm_config=vllm_config,
            executor_class=executor_class,
228
            log_requests=engine_args.enable_log_requests,
229
230
231
            log_stats=not engine_args.disable_log_stats,
            start_engine_loop=start_engine_loop,
            usage_context=usage_context,
232
            stat_loggers=stat_loggers,
233
234
        )

235
236
237
    def __del__(self):
        self.shutdown()

238
239
240
    def shutdown(self):
        """Shutdown, cleaning up the background proc and IPC."""

241
242
        shutdown_prometheus()

243
244
        if engine_core := getattr(self, "engine_core", None):
            engine_core.shutdown()
245

246
        cancel_task_threadsafe(getattr(self, "output_handler", None))
247

248
249
250
251
    @property
    def tokenizer(self) -> Optional[AnyTokenizer]:
        return self.processor.tokenizer

252
253
254
    async def get_supported_tasks(self) -> tuple[SupportedTask, ...]:
        return await self.engine_core.get_supported_tasks_async()

255
256
257
    async def add_request(
        self,
        request_id: str,
258
        prompt: Union[EngineCoreRequest, PromptType],
259
260
261
        params: Union[SamplingParams, PoolingParams],
        arrival_time: Optional[float] = None,
        lora_request: Optional[LoRARequest] = None,
262
        tokenization_kwargs: Optional[dict[str, Any]] = None,
263
264
        trace_headers: Optional[Mapping[str, str]] = None,
        priority: int = 0,
265
        data_parallel_rank: Optional[int] = None,
266
        prompt_text: Optional[str] = None,
267
    ) -> RequestOutputCollector:
268
269
        """Add new request to the AsyncLLM."""

270
271
272
        if self.errored:
            raise EngineDeadError()

273
        is_pooling = isinstance(params, PoolingParams)
274
275
276

        # Create a new output collector for the request.
        queue = RequestOutputCollector(output_kind=params.output_kind)
277

278
        # Convert Input --> Request.
279
280
281
282
283
284
        if isinstance(prompt, EngineCoreRequest):
            request = prompt
        else:
            assert prompt_text is None
            logger.warning_once(
                "Processor has been moved under OpenAIServing and will "
285
286
287
288
289
290
291
292
293
294
295
296
297
298
                "be removed from AsyncLLM in v0.13."
            )
            request = self.processor.process_inputs(
                request_id,
                prompt,
                params,
                arrival_time,
                lora_request,
                tokenization_kwargs,
                trace_headers,
                priority,
                data_parallel_rank,
            )
            prompt_text = prompt if isinstance(prompt, str) else prompt.get("prompt")
299

300
        if is_pooling or params.n == 1:
301
            await self._add_request(request, prompt_text, None, 0, queue)
302
303
            return queue

304
305
306
307
308
        # Get the updated SamplingParams from the request, which
        # were cloned/updated in processor.process_inputs above.
        parent_params = request.sampling_params
        assert parent_params is not None

309
        # Fan out child requests (for n>1).
310
311
312
        parent_request = ParentRequest(request_id, parent_params)
        for idx in range(parent_params.n):
            request_id, child_params = parent_request.get_child_info(idx)
313
            child_request = request if idx == parent_params.n - 1 else copy(request)
314
            child_request.request_id = request_id
315
            child_request.sampling_params = child_params
316
317
318
            await self._add_request(
                child_request, prompt_text, parent_request, idx, queue
            )
319
        return queue
320

321
322
323
324
325
326
327
328
    async def _add_request(
        self,
        request: EngineCoreRequest,
        prompt: Optional[str],
        parent_req: Optional[ParentRequest],
        index: int,
        queue: RequestOutputCollector,
    ):
329
        # Add the request to OutputProcessor (this process).
330
        self.output_processor.add_request(request, prompt, parent_req, index, queue)
331

332
333
        # Add the EngineCoreRequest to EngineCore (separate process).
        await self.engine_core.add_request_async(request)
334

335
336
        if self.log_requests:
            logger.info("Added request %s.", request.request_id)
337
338
339
340
341
342

    # TODO: we should support multiple prompts in one call, as you
    # can do with LLM.generate. So that for multi-prompt completion
    # requests we don't need to send multiple messages to core proc,
    # and so we don't need multiple streams which then get
    # re-multiplexed in the API server anyhow.
343
    async def generate(
344
        self,
345
        prompt: Union[EngineCoreRequest, PromptType],
346
347
        sampling_params: SamplingParams,
        request_id: str,
348
349
        *,
        prompt_text: Optional[str] = None,
350
        lora_request: Optional[LoRARequest] = None,
351
        tokenization_kwargs: Optional[dict[str, Any]] = None,
352
353
        trace_headers: Optional[Mapping[str, str]] = None,
        priority: int = 0,
354
        data_parallel_rank: Optional[int] = None,
355
356
357
358
    ) -> AsyncGenerator[RequestOutput, None]:
        """
        Main function called by the API server to kick off a request
            * 1) Making an AsyncStream corresponding to the Request.
359
            * 2) Processing the Input.
360
361
362
            * 3) Adding the Request to the Detokenizer.
            * 4) Adding the Request to the EngineCore (separate process).

363
364
        A separate output_handler loop runs in a background AsyncIO task,
        pulling outputs from EngineCore and putting them into the
365
366
367
368
369
370
        per-request AsyncStream.

        The caller of generate() iterates the returned AsyncGenerator,
        returning the RequestOutput back to the caller.
        """

371
372
373
374
        if (
            self.vllm_config.cache_config.kv_sharing_fast_prefill
            and sampling_params.prompt_logprobs
        ):
375
376
377
            raise ValueError(
                "--kv-sharing-fast-prefill produces incorrect logprobs for "
                "prompt tokens, please disable it when the requests need "
378
379
                "prompt logprobs"
            )
380

381
382
383
384
        try:
            # We start the output_handler on the first call to generate() so
            # we can call __init__ before the event loop, which enables us
            # to handle startup failure gracefully in the OpenAI server.
385
            self._run_output_handler()
386

387
388
389
390
391
392
393
394
395
396
            if tokenization_kwargs is None:
                tokenization_kwargs = {}
                truncate_prompt_tokens = sampling_params.truncate_prompt_tokens

                _validate_truncation_size(
                    self.model_config.max_model_len,
                    truncate_prompt_tokens,
                    tokenization_kwargs,
                )

397
398
399
400
401
402
403
404
405
406
407
            q = await self.add_request(
                request_id,
                prompt,
                sampling_params,
                lora_request=lora_request,
                tokenization_kwargs=tokenization_kwargs,
                trace_headers=trace_headers,
                priority=priority,
                data_parallel_rank=data_parallel_rank,
                prompt_text=prompt_text,
            )
408

409
410
            # The output_handler task pushes items into the queue.
            # This task pulls from the queue and yields to caller.
411
412
            finished = False
            while not finished:
413
414
                # Note: drain queue without await if possible (avoids
                # task switching under load which helps performance).
415
                out = q.get_nowait() or await q.get()
416

417
                # Note: both OutputProcessor and EngineCore handle their
418
                # own request cleanup based on finished.
419
                finished = out.finished
420
421
                yield out

422
        # If the request is disconnected by the client, generate()
423
424
425
        # is cancelled or the generator is garbage collected. So,
        # we abort the request if we end up here.
        except (asyncio.CancelledError, GeneratorExit):
426
            await self.abort(request_id)
427
428
            if self.log_requests:
                logger.info("Request %s aborted.", request_id)
429
            raise
430

431
432
433
434
435
        # Engine is dead. Do not abort since we shut down.
        except EngineDeadError:
            if self.log_requests:
                logger.info("Request %s failed (engine dead).", request_id)
            raise
436

437
438
439
440
441
        # Request validation error.
        except ValueError:
            if self.log_requests:
                logger.info("Request %s failed (bad request).", request_id)
            raise
442

443
        # Unexpected error in the generate() task (possibly recoverable).
444
        except Exception as e:
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
            await self.abort(request_id)
            if self.log_requests:
                logger.info("Request %s failed.", request_id)
            raise EngineGenerateError() from e

    def _run_output_handler(self):
        """Background loop: pulls from EngineCore and pushes to AsyncStreams."""

        if self.output_handler is not None:
            return

        # Ensure that the task doesn't have a circular ref back to the AsyncLLM
        # object, or else it won't be garbage collected and cleaned up properly.
        engine_core = self.engine_core
        output_processor = self.output_processor
        log_stats = self.log_stats
461
        logger_manager = self.logger_manager
462
463
464
465
466
467
468
469

        async def output_handler():
            try:
                while True:
                    # 1) Pull EngineCoreOutputs from the EngineCore.
                    outputs = await engine_core.get_output_async()
                    num_outputs = len(outputs.outputs)

470
471
472
                    iteration_stats = (
                        IterationStats() if (log_stats and num_outputs) else None
                    )
473
474
475
476
477

                    # Split outputs into chunks of at most
                    # VLLM_V1_OUTPUT_PROC_CHUNK_SIZE, so that we don't block the
                    # event loop for too long.
                    if num_outputs <= VLLM_V1_OUTPUT_PROC_CHUNK_SIZE:
478
                        slices = (outputs.outputs,)
479
480
481
                    else:
                        slices = np.array_split(
                            outputs.outputs,
482
483
                            cdiv(num_outputs, VLLM_V1_OUTPUT_PROC_CHUNK_SIZE),
                        )
484
485
486
487

                    for i, outputs_slice in enumerate(slices):
                        # 2) Process EngineCoreOutputs.
                        processed_outputs = output_processor.process_outputs(
488
489
                            outputs_slice, outputs.timestamp, iteration_stats
                        )
490
491
492
493
494
495
496
497
498
                        # NOTE: RequestOutputs are pushed to their queues.
                        assert not processed_outputs.request_outputs

                        # Allow other asyncio tasks to run between chunks
                        if i + 1 < len(slices):
                            await asyncio.sleep(0)

                        # 3) Abort any reqs that finished due to stop strings.
                        await engine_core.abort_requests_async(
499
500
                            processed_outputs.reqs_to_abort
                        )
501
502
503
504

                    # 4) Logging.
                    # TODO(rob): make into a coroutine and launch it in
                    # background thread once Prometheus overhead is non-trivial.
505
506
507
                    if logger_manager:
                        logger_manager.record(
                            engine_idx=outputs.engine_index,
508
509
510
511
512
513
514
515
                            scheduler_stats=outputs.scheduler_stats,
                            iteration_stats=iteration_stats,
                        )
            except Exception as e:
                logger.exception("AsyncLLM output_handler failed.")
                output_processor.propagate_error(e)

        self.output_handler = asyncio.create_task(output_handler())
516

517
    async def abort(self, request_id: Union[str, Iterable[str]]) -> None:
518
        """Abort RequestId in OutputProcessor and EngineCore."""
519

520
521
522
        request_ids = (
            (request_id,) if isinstance(request_id, str) else as_list(request_id)
        )
523
524
        all_request_ids = self.output_processor.abort_requests(request_ids)
        await self.engine_core.abort_requests_async(all_request_ids)
525

526
        if self.log_requests:
527
            logger.info("Aborted request(s) %s.", ",".join(request_ids))
528

529
    async def encode(
530
531
532
533
534
535
536
        self,
        prompt: PromptType,
        pooling_params: PoolingParams,
        request_id: str,
        lora_request: Optional[LoRARequest] = None,
        trace_headers: Optional[Mapping[str, str]] = None,
        priority: int = 0,
537
        truncate_prompt_tokens: Optional[int] = None,
538
        tokenization_kwargs: Optional[dict[str, Any]] = None,
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
    ) -> AsyncGenerator[PoolingRequestOutput, None]:
        """
        Main function called by the API server to kick off a request
            * 1) Making an AsyncStream corresponding to the Request.
            * 2) Processing the Input.
            * 3) Adding the Request to the EngineCore (separate process).

        A separate output_handler loop runs in a background AsyncIO task,
        pulling outputs from EngineCore and putting them into the
        per-request AsyncStream.

        The caller of generate() iterates the returned AsyncGenerator,
        returning the RequestOutput back to the caller.
        """

        try:
            # We start the output_handler on the first call to generate() so
            # we can call __init__ before the event loop, which enables us
            # to handle startup failure gracefully in the OpenAI server.
            self._run_output_handler()

560
            if tokenization_kwargs is None:
561
                tokenization_kwargs = {}
562
563
564
565
566
567
            _validate_truncation_size(
                self.model_config.max_model_len,
                truncate_prompt_tokens,
                tokenization_kwargs,
            )

568
569
570
571
572
            q = await self.add_request(
                request_id,
                prompt,
                pooling_params,
                lora_request=lora_request,
573
                tokenization_kwargs=tokenization_kwargs,
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
                trace_headers=trace_headers,
                priority=priority,
            )

            # The output_handler task pushes items into the queue.
            # This task pulls from the queue and yields to caller.
            finished = False
            while not finished:
                # Note: drain queue without await if possible (avoids
                # task switching under load which helps performance).
                out = q.get_nowait() or await q.get()
                assert isinstance(out, PoolingRequestOutput)
                # Note: both OutputProcessor and EngineCore handle their
                # own request cleanup based on finished.
                finished = out.finished
                yield out

        # If the request is disconnected by the client, generate()
        # is cancelled. So, we abort the request if we end up here.
        except asyncio.CancelledError:
            await self.abort(request_id)
            if self.log_requests:
                logger.info("Request %s aborted.", request_id)
            raise

        # Engine is dead. Do not abort since we shut down.
        except EngineDeadError:
            if self.log_requests:
                logger.info("Request %s failed (engine dead).", request_id)
            raise

        # Request validation error.
        except ValueError:
            if self.log_requests:
                logger.info("Request %s failed (bad request).", request_id)
            raise

        # Unexpected error in the generate() task (possibly recoverable).
        except Exception as e:
            await self.abort(request_id)
            if self.log_requests:
                logger.info("Request %s failed.", request_id)
            raise EngineGenerateError() from e
617

618
619
620
    async def get_vllm_config(self) -> VllmConfig:
        return self.vllm_config

621
622
623
    async def get_model_config(self) -> ModelConfig:
        return self.model_config

624
625
626
    async def get_input_preprocessor(self) -> InputPreprocessor:
        return self.processor.input_preprocessor

627
    async def get_tokenizer(self) -> AnyTokenizer:
628
        if self.tokenizer is None:
629
630
631
            raise ValueError(
                "Unable to get tokenizer because skip_tokenizer_init is True"
            )
632

633
        return self.tokenizer
634
635

    async def is_tracing_enabled(self) -> bool:
636
        return self.observability_config.otlp_traces_endpoint is not None
637

638
    async def do_log_stats(self) -> None:
639
640
        if self.logger_manager:
            self.logger_manager.log()
641
642
643

    async def check_health(self) -> None:
        logger.debug("Called check_health.")
644
645
        if self.errored:
            raise self.dead_error
646
647

    async def start_profile(self) -> None:
648
649
650
651
        coros = [self.engine_core.profile_async(True)]
        if self.profiler is not None:
            coros.append(asyncio.to_thread(self.profiler.start))
        await asyncio.gather(*coros)
652
653

    async def stop_profile(self) -> None:
654
655
656
657
        coros = [self.engine_core.profile_async(False)]
        if self.profiler is not None:
            coros.append(asyncio.to_thread(self.profiler.stop))
        await asyncio.gather(*coros)
658

659
    async def reset_mm_cache(self) -> None:
660
        self.processor.clear_cache()
661
662
        await self.engine_core.reset_mm_cache_async()

663
    async def reset_prefix_cache(self, device: Optional[Device] = None) -> None:
664
665
        if device == Device.CPU:
            raise ValueError("Not supported on CPU.")
666
667
        await self.engine_core.reset_prefix_cache_async()

668
    async def sleep(self, level: int = 1) -> None:
669
        await self.reset_prefix_cache()
670
671
        await self.engine_core.sleep_async(level)

672
673
    async def wake_up(self, tags: Optional[list[str]] = None) -> None:
        await self.engine_core.wake_up_async(tags)
674

675
676
677
    async def is_sleeping(self) -> bool:
        return await self.engine_core.is_sleeping_async()

678
    async def add_lora(self, lora_request: LoRARequest) -> bool:
679
        """Load a new LoRA adapter into the engine for future requests."""
680
681
682
683
684
685
        return await self.engine_core.add_lora_async(lora_request)

    async def remove_lora(self, lora_id: int) -> bool:
        """Remove an already loaded LoRA adapter."""
        return await self.engine_core.remove_lora_async(lora_id)

686
    async def list_loras(self) -> set[int]:
687
688
689
690
691
692
        """List all registered adapters."""
        return await self.engine_core.list_loras_async()

    async def pin_lora(self, lora_id: int) -> bool:
        """Prevent an adapter from being evicted."""
        return await self.engine_core.pin_lora_async(lora_id)
693

694
695
696
697
698
699
700
    async def collective_rpc(
        self,
        method: str,
        timeout: Optional[float] = None,
        args: tuple = (),
        kwargs: Optional[dict] = None,
    ):
701
702
703
704
        """
        Perform a collective RPC call to the given path.
        """
        return await self.engine_core.collective_rpc_async(
705
706
            method, timeout, args, kwargs
        )
707

708
709
710
711
712
713
714
715
    async def wait_for_requests_to_drain(self, drain_timeout: int = 300):
        """Wait for all requests to be drained."""
        start_time = time.time()
        while time.time() - start_time < drain_timeout:
            if not self.engine_core.dp_engines_running():
                logger.info("Engines are idle, requests have been drained")
                return

716
            logger.info("Engines are still running, waiting for requests to drain...")
717
718
            await asyncio.sleep(1)  # Wait 1 second before checking again

719
720
721
722
        raise TimeoutError(
            f"Timeout reached after {drain_timeout} seconds "
            "waiting for requests to drain."
        )
723

724
725
726
    async def scale_elastic_ep(
        self, new_data_parallel_size: int, drain_timeout: int = 300
    ):
727
728
729
730
731
732
733
734
        """
        Scale up or down the data parallel size by adding or removing
        engine cores.
        Args:
            new_data_parallel_size: The new number of data parallel workers
            drain_timeout:
                Maximum time to wait for requests to drain (seconds)
        """
735
        old_data_parallel_size = self.vllm_config.parallel_config.data_parallel_size
736
        if old_data_parallel_size == new_data_parallel_size:
737
738
739
740
            logger.info(
                "Data parallel size is already %s, skipping scale",
                new_data_parallel_size,
            )
741
742
            return
        logger.info(
743
744
745
            "Waiting for requests to drain before scaling up to %s engines...",
            new_data_parallel_size,
        )
746
747
        await self.wait_for_requests_to_drain(drain_timeout)
        logger.info(
748
749
750
            "Requests have been drained, proceeding with scale to %s engines",
            new_data_parallel_size,
        )
751
        await self.engine_core.scale_elastic_ep(new_data_parallel_size)
752
        self.vllm_config.parallel_config.data_parallel_size = new_data_parallel_size
753
754

        # recreate stat loggers
755
756
757
758
759
760
        if new_data_parallel_size > old_data_parallel_size and self.log_stats:
            # TODO(rob): fix this after talking with Ray team.
            # This resets all the prometheus metrics since we
            # unregister during initialization. Need to understand
            # the intended behavior here better.
            self.logger_manager = StatLoggerManager(
761
                vllm_config=self.vllm_config,
762
                engine_idxs=list(range(new_data_parallel_size)),
763
764
765
                custom_stat_loggers=None,
            )

766
767
    @property
    def is_running(self) -> bool:
768
769
        # Is None before the loop is started.
        return self.output_handler is None or not self.output_handler.done()
770
771
772

    @property
    def is_stopped(self) -> bool:
773
        return self.errored
774
775
776

    @property
    def errored(self) -> bool:
777
        return self.engine_core.resources.engine_dead or not self.is_running
778
779
780

    @property
    def dead_error(self) -> BaseException:
781
        return EngineDeadError()