async_llm.py 30.5 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, cast
10

11
import numpy as np
12
import torch
13

14
import vllm.envs as envs
15
from vllm.config import VllmConfig
16
from vllm.engine.arg_utils import AsyncEngineArgs
17
from vllm.engine.protocol import EngineClient
18
from vllm.entrypoints.utils import _validate_truncation_size
19
from vllm.inputs import PromptType
20
21
from vllm.logger import init_logger
from vllm.lora.request import LoRARequest
22
from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalRegistry
23
from vllm.outputs import PoolingRequestOutput, RequestOutput
24
from vllm.plugins.io_processors import get_io_processor
25
from vllm.pooling_params import PoolingParams
26
from vllm.sampling_params import SamplingParams
27
from vllm.tasks import SupportedTask
28
from vllm.tracing import init_tracer
29
from vllm.transformers_utils.config import maybe_register_config_serialize_by_value
30
from vllm.transformers_utils.tokenizer import AnyTokenizer, init_tokenizer_from_configs
31
from vllm.usage.usage_lib import UsageContext
32
33
34
from vllm.utils.async_utils import cancel_task_threadsafe
from vllm.utils.collection_utils import as_list
from vllm.utils.func_utils import deprecate_kwargs
35
from vllm.utils.math_utils import cdiv
36
from vllm.v1.engine import EngineCoreRequest
37
from vllm.v1.engine.core_client import EngineCoreClient
38
from vllm.v1.engine.exceptions import EngineDeadError, EngineGenerateError
39
from vllm.v1.engine.output_processor import OutputProcessor, RequestOutputCollector
40
from vllm.v1.engine.parallel_sampling import ParentRequest
41
from vllm.v1.engine.processor import Processor
42
from vllm.v1.executor import Executor
43
44
45
46
47
from vllm.v1.metrics.loggers import (
    StatLoggerFactory,
    StatLoggerManager,
    load_stat_logger_plugin_factories,
)
48
from vllm.v1.metrics.prometheus import shutdown_prometheus
49
from vllm.v1.metrics.stats import IterationStats
50
51
52
53
54
55
56
57

logger = init_logger(__name__)


class AsyncLLM(EngineClient):
    def __init__(
        self,
        vllm_config: VllmConfig,
58
        executor_class: type[Executor],
59
60
        log_stats: bool,
        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
61
        mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
62
63
64
        use_cached_outputs: bool = False,
        log_requests: bool = True,
        start_engine_loop: bool = True,
65
        stat_loggers: list[StatLoggerFactory] | None = None,
66
        aggregate_engine_logging: bool = False,
67
        client_addresses: dict[str, str] | None = None,
68
        client_count: int = 1,
69
        client_index: int = 0,
70
    ) -> None:
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
        """
        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
        """
91
92
93
        # Ensure we can serialize custom transformer configs
        maybe_register_config_serialize_by_value()

94
        self.model_config = vllm_config.model_config
95
        self.vllm_config = vllm_config
96
        self.observability_config = vllm_config.observability_config
97
        self.log_requests = log_requests
98

99
100
101
102
103
104
        custom_stat_loggers = list(stat_loggers or [])
        custom_stat_loggers.extend(load_stat_logger_plugin_factories())

        has_custom_loggers = bool(custom_stat_loggers)
        self.log_stats = log_stats or has_custom_loggers
        if not log_stats and has_custom_loggers:
105
            logger.info(
106
107
108
                "AsyncLLM created with log_stats=False, "
                "but custom stat loggers were found; "
                "enabling logging without default stat loggers."
109
            )
110

111
112
113
114
115
116
117
118
119
120
        if self.model_config.skip_tokenizer_init:
            tokenizer = None
        else:
            tokenizer = init_tokenizer_from_configs(self.model_config)

        self.processor = Processor(self.vllm_config, tokenizer)
        self.io_processor = get_io_processor(
            self.vllm_config,
            self.model_config.io_processor_plugin,
        )
121

122
        # OutputProcessor (converts EngineCoreOutputs --> RequestOutput).
123
        stream_interval = self.vllm_config.scheduler_config.stream_interval
124
        self.output_processor = OutputProcessor(
125
            self.tokenizer, log_stats=self.log_stats, stream_interval=stream_interval
126
        )
127
128
129
        endpoint = self.observability_config.otlp_traces_endpoint
        if endpoint is not None:
            tracer = init_tracer("vllm.llm_engine", endpoint)
130
            self.output_processor.tracer = tracer
131
132

        # EngineCore (starts the engine in background process).
133
        self.engine_core = EngineCoreClient.make_async_mp_client(
134
135
            vllm_config=vllm_config,
            executor_class=executor_class,
136
            log_stats=self.log_stats,
137
            client_addresses=client_addresses,
138
            client_count=client_count,
139
            client_index=client_index,
140
        )
141
142

        # Loggers.
143
        self.logger_manager: StatLoggerManager | None = None
144
145
146
        if self.log_stats:
            self.logger_manager = StatLoggerManager(
                vllm_config=vllm_config,
147
                engine_idxs=self.engine_core.engine_ranks_managed,
148
                custom_stat_loggers=custom_stat_loggers,
149
                enable_default_loggers=log_stats,
150
                client_count=client_count,
151
                aggregate_engine_logging=aggregate_engine_logging,
152
153
154
            )
            self.logger_manager.log_engine_initialized()

155
        self.output_handler: asyncio.Task | None = None
156
157
158
159
160
161
        try:
            # Start output handler eagerly if we are in the asyncio eventloop.
            asyncio.get_running_loop()
            self._run_output_handler()
        except RuntimeError:
            pass
162

163
164
165
        if envs.VLLM_TORCH_PROFILER_DIR:
            logger.info(
                "Torch profiler enabled. AsyncLLM CPU traces will be collected under %s",  # noqa: E501
166
167
                envs.VLLM_TORCH_PROFILER_DIR,
            )
168
169
170
171
172
173
174
            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(
175
176
177
                    envs.VLLM_TORCH_PROFILER_DIR, worker_name=worker_name, use_gzip=True
                ),
            )
178
179
180
        else:
            self.profiler = None

181
    @classmethod
182
183
    @deprecate_kwargs(
        "disable_log_requests",
184
185
186
        additional_message=(
            "This argument will have no effect. Use `enable_log_requests` instead."
        ),
187
    )
188
    def from_vllm_config(
189
190
191
192
        cls,
        vllm_config: VllmConfig,
        start_engine_loop: bool = True,
        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
193
        stat_loggers: list[StatLoggerFactory] | None = None,
194
        enable_log_requests: bool = False,
195
        aggregate_engine_logging: bool = False,
196
        disable_log_stats: bool = False,
197
        client_addresses: dict[str, str] | None = None,
198
199
200
        client_count: int = 1,
        client_index: int = 0,
        disable_log_requests: bool = True,  # Deprecated, will be removed
201
202
203
204
205
206
    ) -> "AsyncLLM":
        # Create the LLMEngine.
        return cls(
            vllm_config=vllm_config,
            executor_class=Executor.get_class(vllm_config),
            start_engine_loop=start_engine_loop,
207
            stat_loggers=stat_loggers,
208
            log_requests=enable_log_requests,
209
            log_stats=not disable_log_stats,
210
            aggregate_engine_logging=aggregate_engine_logging,
211
            usage_context=usage_context,
212
            client_addresses=client_addresses,
213
            client_count=client_count,
214
            client_index=client_index,
215
216
        )

217
218
219
220
221
222
    @classmethod
    def from_engine_args(
        cls,
        engine_args: AsyncEngineArgs,
        start_engine_loop: bool = True,
        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
223
        stat_loggers: list[StatLoggerFactory] | None = None,
224
    ) -> "AsyncLLM":
225
226
227
        """Create an AsyncLLM from the EngineArgs."""

        # Create the engine configs.
228
        vllm_config = engine_args.create_engine_config(usage_context)
229
        executor_class = Executor.get_class(vllm_config)
230
231
232
233
234

        # Create the AsyncLLM.
        return cls(
            vllm_config=vllm_config,
            executor_class=executor_class,
235
            log_requests=engine_args.enable_log_requests,
236
237
238
            log_stats=not engine_args.disable_log_stats,
            start_engine_loop=start_engine_loop,
            usage_context=usage_context,
239
            stat_loggers=stat_loggers,
240
241
        )

242
243
244
    def __del__(self):
        self.shutdown()

245
246
247
    def shutdown(self):
        """Shutdown, cleaning up the background proc and IPC."""

248
249
        shutdown_prometheus()

250
251
        if engine_core := getattr(self, "engine_core", None):
            engine_core.shutdown()
252

253
254
255
        handler = getattr(self, "output_handler", None)
        if handler is not None:
            cancel_task_threadsafe(handler)
256

257
258
259
    async def get_supported_tasks(self) -> tuple[SupportedTask, ...]:
        return await self.engine_core.get_supported_tasks_async()

260
261
262
    async def add_request(
        self,
        request_id: str,
263
264
265
266
267
268
        prompt: EngineCoreRequest | PromptType,
        params: SamplingParams | PoolingParams,
        arrival_time: float | None = None,
        lora_request: LoRARequest | None = None,
        tokenization_kwargs: dict[str, Any] | None = None,
        trace_headers: Mapping[str, str] | None = None,
269
        priority: int = 0,
270
271
        data_parallel_rank: int | None = None,
        prompt_text: str | None = None,
272
    ) -> RequestOutputCollector:
273
274
        """Add new request to the AsyncLLM."""

275
276
277
        if self.errored:
            raise EngineDeadError()

278
        is_pooling = isinstance(params, PoolingParams)
279
280
281

        # Create a new output collector for the request.
        queue = RequestOutputCollector(output_kind=params.output_kind)
282

283
        # Convert Input --> Request.
284
285
286
287
288
289
        if isinstance(prompt, EngineCoreRequest):
            request = prompt
        else:
            assert prompt_text is None
            logger.warning_once(
                "Processor has been moved under OpenAIServing and will "
290
291
292
293
294
295
296
297
298
299
300
301
302
                "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,
            )
303
304
305
306
            if isinstance(prompt, str):
                prompt_text = prompt
            elif isinstance(prompt, Mapping):
                prompt_text = cast(str | None, prompt.get("prompt"))
307

308
        if is_pooling or params.n == 1:
309
            await self._add_request(request, prompt_text, None, 0, queue)
310
311
            return queue

312
313
314
315
316
        # 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

317
        # Fan out child requests (for n>1).
318
319
320
        parent_request = ParentRequest(request_id, parent_params)
        for idx in range(parent_params.n):
            request_id, child_params = parent_request.get_child_info(idx)
321
            child_request = request if idx == parent_params.n - 1 else copy(request)
322
            child_request.request_id = request_id
323
            child_request.sampling_params = child_params
324
325
326
            await self._add_request(
                child_request, prompt_text, parent_request, idx, queue
            )
327
        return queue
328

329
330
331
    async def _add_request(
        self,
        request: EngineCoreRequest,
332
333
        prompt: str | None,
        parent_req: ParentRequest | None,
334
335
336
        index: int,
        queue: RequestOutputCollector,
    ):
337
        # Add the request to OutputProcessor (this process).
338
        self.output_processor.add_request(request, prompt, parent_req, index, queue)
339

340
341
        # Add the EngineCoreRequest to EngineCore (separate process).
        await self.engine_core.add_request_async(request)
342

343
344
        if self.log_requests:
            logger.info("Added request %s.", request.request_id)
345
346
347
348
349
350

    # 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.
351
    async def generate(
352
        self,
353
        prompt: EngineCoreRequest | PromptType,
354
355
        sampling_params: SamplingParams,
        request_id: str,
356
        *,
357
358
359
360
        prompt_text: str | None = None,
        lora_request: LoRARequest | None = None,
        tokenization_kwargs: dict[str, Any] | None = None,
        trace_headers: Mapping[str, str] | None = None,
361
        priority: int = 0,
362
        data_parallel_rank: int | None = None,
363
364
365
366
    ) -> AsyncGenerator[RequestOutput, None]:
        """
        Main function called by the API server to kick off a request
            * 1) Making an AsyncStream corresponding to the Request.
367
            * 2) Processing the Input.
368
369
370
            * 3) Adding the Request to the Detokenizer.
            * 4) Adding the Request to the EngineCore (separate process).

371
372
        A separate output_handler loop runs in a background AsyncIO task,
        pulling outputs from EngineCore and putting them into the
373
374
375
376
377
378
        per-request AsyncStream.

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

379
380
381
382
        if (
            self.vllm_config.cache_config.kv_sharing_fast_prefill
            and sampling_params.prompt_logprobs
        ):
383
384
385
            raise ValueError(
                "--kv-sharing-fast-prefill produces incorrect logprobs for "
                "prompt tokens, please disable it when the requests need "
386
387
                "prompt logprobs"
            )
388

389
390
391
392
        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.
393
            self._run_output_handler()
394

395
396
397
398
399
400
401
402
403
404
            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,
                )

405
406
407
408
409
410
411
412
413
414
415
            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,
            )
416

417
418
            # The output_handler task pushes items into the queue.
            # This task pulls from the queue and yields to caller.
419
420
            finished = False
            while not finished:
421
422
                # Note: drain queue without await if possible (avoids
                # task switching under load which helps performance).
423
                out = q.get_nowait() or await q.get()
424

425
                # Note: both OutputProcessor and EngineCore handle their
426
                # own request cleanup based on finished.
427
                finished = out.finished
428
                assert isinstance(out, RequestOutput)
429
430
                yield out

431
        # If the request is disconnected by the client, generate()
432
433
434
        # is cancelled or the generator is garbage collected. So,
        # we abort the request if we end up here.
        except (asyncio.CancelledError, GeneratorExit):
435
            await self.abort(request_id)
436
437
            if self.log_requests:
                logger.info("Request %s aborted.", request_id)
438
            raise
439

440
441
442
443
444
        # 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
445

446
447
448
449
450
        # Request validation error.
        except ValueError:
            if self.log_requests:
                logger.info("Request %s failed (bad request).", request_id)
            raise
451

452
        # Unexpected error in the generate() task (possibly recoverable).
453
        except Exception as e:
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
            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
470
        logger_manager = self.logger_manager
471
        processor = self.processor
472
473
474
475
476
477
478
479

        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)

480
481
482
                    iteration_stats = (
                        IterationStats() if (log_stats and num_outputs) else None
                    )
483
484
485
486

                    # 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.
487
                    if num_outputs <= envs.VLLM_V1_OUTPUT_PROC_CHUNK_SIZE:
488
                        slices = (outputs.outputs,)
489
490
491
                    else:
                        slices = np.array_split(
                            outputs.outputs,
492
                            cdiv(num_outputs, envs.VLLM_V1_OUTPUT_PROC_CHUNK_SIZE),
493
                        )
494
495
496
497

                    for i, outputs_slice in enumerate(slices):
                        # 2) Process EngineCoreOutputs.
                        processed_outputs = output_processor.process_outputs(
498
499
                            outputs_slice, outputs.timestamp, iteration_stats
                        )
500
501
502
503
504
505
506
507
508
                        # 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(
509
510
                            processed_outputs.reqs_to_abort
                        )
511

512
513
                    output_processor.update_scheduler_stats(outputs.scheduler_stats)

514
515
516
                    # 4) Logging.
                    # TODO(rob): make into a coroutine and launch it in
                    # background thread once Prometheus overhead is non-trivial.
517
518
519
                    if logger_manager:
                        logger_manager.record(
                            engine_idx=outputs.engine_index,
520
521
                            scheduler_stats=outputs.scheduler_stats,
                            iteration_stats=iteration_stats,
522
                            mm_cache_stats=processor.stat_mm_cache(),
523
524
525
526
527
528
                        )
            except Exception as e:
                logger.exception("AsyncLLM output_handler failed.")
                output_processor.propagate_error(e)

        self.output_handler = asyncio.create_task(output_handler())
529

530
    async def abort(self, request_id: str | Iterable[str]) -> None:
531
        """Abort RequestId in OutputProcessor and EngineCore."""
532

533
534
535
        request_ids = (
            (request_id,) if isinstance(request_id, str) else as_list(request_id)
        )
536
537
        all_request_ids = self.output_processor.abort_requests(request_ids)
        await self.engine_core.abort_requests_async(all_request_ids)
538

539
        if self.log_requests:
540
            logger.info("Aborted request(s) %s.", ",".join(request_ids))
541

542
    async def encode(
543
544
545
546
        self,
        prompt: PromptType,
        pooling_params: PoolingParams,
        request_id: str,
547
548
        lora_request: LoRARequest | None = None,
        trace_headers: Mapping[str, str] | None = None,
549
        priority: int = 0,
550
551
        truncate_prompt_tokens: int | None = None,
        tokenization_kwargs: dict[str, Any] | None = None,
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
    ) -> 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()

573
            if tokenization_kwargs is None:
574
                tokenization_kwargs = {}
575
576
577
578
579
580
            _validate_truncation_size(
                self.model_config.max_model_len,
                truncate_prompt_tokens,
                tokenization_kwargs,
            )

581
582
583
584
585
            q = await self.add_request(
                request_id,
                prompt,
                pooling_params,
                lora_request=lora_request,
586
                tokenization_kwargs=tokenization_kwargs,
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
617
618
619
620
621
622
623
624
625
626
627
628
629
                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
630

631
    @property
632
    def tokenizer(self) -> AnyTokenizer | None:
633
        return self.processor.tokenizer
634

635
    @tokenizer.setter
636
    def tokenizer(self, tokenizer: AnyTokenizer | None) -> None:
637
        self.processor.tokenizer = tokenizer
638

639
    async def get_tokenizer(self) -> AnyTokenizer:
640
        if self.tokenizer is None:
641
642
643
            raise ValueError(
                "Unable to get tokenizer because skip_tokenizer_init is True"
            )
644

645
        return self.tokenizer
646
647

    async def is_tracing_enabled(self) -> bool:
648
        return self.observability_config.otlp_traces_endpoint is not None  # type: ignore
649

650
    async def do_log_stats(self) -> None:
651
652
        if self.logger_manager:
            self.logger_manager.log()
653
654
655

    async def check_health(self) -> None:
        logger.debug("Called check_health.")
656
657
        if self.errored:
            raise self.dead_error
658
659

    async def start_profile(self) -> None:
660
661
662
663
        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)
664
665

    async def stop_profile(self) -> None:
666
667
668
669
        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)
670

671
    async def reset_mm_cache(self) -> None:
672
        self.processor.clear_mm_cache()
673
674
        await self.engine_core.reset_mm_cache_async()

675
    async def reset_prefix_cache(self) -> None:
676
677
        await self.engine_core.reset_prefix_cache_async()

678
    async def sleep(self, level: int = 1) -> None:
679
        await self.reset_prefix_cache()
680
681
        await self.engine_core.sleep_async(level)

682
683
684
        if self.logger_manager is not None:
            self.logger_manager.record_sleep_state(1, level)

685
    async def wake_up(self, tags: list[str] | None = None) -> None:
686
        await self.engine_core.wake_up_async(tags)
687

688
689
690
        if self.logger_manager is not None:
            self.logger_manager.record_sleep_state(0, 0)

691
692
693
    async def is_sleeping(self) -> bool:
        return await self.engine_core.is_sleeping_async()

694
    async def add_lora(self, lora_request: LoRARequest) -> bool:
695
        """Load a new LoRA adapter into the engine for future requests."""
696
697
698
699
700
701
        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)

702
    async def list_loras(self) -> set[int]:
703
704
705
706
707
708
        """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)
709

710
711
712
    async def collective_rpc(
        self,
        method: str,
713
        timeout: float | None = None,
714
        args: tuple = (),
715
        kwargs: dict | None = None,
716
    ):
717
718
719
720
        """
        Perform a collective RPC call to the given path.
        """
        return await self.engine_core.collective_rpc_async(
721
722
            method, timeout, args, kwargs
        )
723

724
725
726
727
728
729
730
731
    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

732
            logger.info("Engines are still running, waiting for requests to drain...")
733
734
            await asyncio.sleep(1)  # Wait 1 second before checking again

735
736
737
738
        raise TimeoutError(
            f"Timeout reached after {drain_timeout} seconds "
            "waiting for requests to drain."
        )
739

740
741
742
    async def scale_elastic_ep(
        self, new_data_parallel_size: int, drain_timeout: int = 300
    ):
743
744
745
746
747
748
749
750
        """
        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)
        """
751
        old_data_parallel_size = self.vllm_config.parallel_config.data_parallel_size
752
        if old_data_parallel_size == new_data_parallel_size:
753
754
755
756
            logger.info(
                "Data parallel size is already %s, skipping scale",
                new_data_parallel_size,
            )
757
758
            return
        logger.info(
759
760
761
            "Waiting for requests to drain before scaling up to %s engines...",
            new_data_parallel_size,
        )
762
763
        await self.wait_for_requests_to_drain(drain_timeout)
        logger.info(
764
765
766
            "Requests have been drained, proceeding with scale to %s engines",
            new_data_parallel_size,
        )
767
        await self.engine_core.scale_elastic_ep(new_data_parallel_size)
768
        self.vllm_config.parallel_config.data_parallel_size = new_data_parallel_size
769
770

        # recreate stat loggers
771
772
773
774
775
776
        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(
777
                vllm_config=self.vllm_config,
778
                engine_idxs=list(range(new_data_parallel_size)),
779
780
781
                custom_stat_loggers=None,
            )

782
783
    @property
    def is_running(self) -> bool:
784
785
        # Is None before the loop is started.
        return self.output_handler is None or not self.output_handler.done()
786
787
788

    @property
    def is_stopped(self) -> bool:
789
        return self.errored
790
791
792

    @property
    def errored(self) -> bool:
793
        return self.engine_core.resources.engine_dead or not self.is_running
794
795
796

    @property
    def dead_error(self) -> BaseException:
797
        return EngineDeadError()