async_llm.py 31.2 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
166
        if (
            envs.VLLM_TORCH_PROFILER_DIR
            and not envs.VLLM_TORCH_PROFILER_DISABLE_ASYNC_LLM
        ):
167
168
            logger.info(
                "Torch profiler enabled. AsyncLLM CPU traces will be collected under %s",  # noqa: E501
169
170
                envs.VLLM_TORCH_PROFILER_DIR,
            )
171
172
173
174
175
176
177
178
179
            if envs.VLLM_PROFILER_MAX_ITERS > 0 or envs.VLLM_PROFILER_DELAY_ITERS > 0:
                logger.warning_once(
                    "Torch profiler received max_iters or delay_iters setting. These "
                    "are not compatible with the AsyncLLM profiler and will be ignored "
                    "for the AsyncLLM process. Engine process profiling will still "
                    "respect these settings. Consider setting "
                    "VLLM_TORCH_PROFILER_DISABLE_ASYNC_LLM=1 to disable "
                    "AsyncLLM profiling."
                )
180
181
182
183
184
185
186
            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(
187
188
189
                    envs.VLLM_TORCH_PROFILER_DIR, worker_name=worker_name, use_gzip=True
                ),
            )
190
191
192
        else:
            self.profiler = None

193
    @classmethod
194
195
    @deprecate_kwargs(
        "disable_log_requests",
196
197
198
        additional_message=(
            "This argument will have no effect. Use `enable_log_requests` instead."
        ),
199
    )
200
    def from_vllm_config(
201
202
203
204
        cls,
        vllm_config: VllmConfig,
        start_engine_loop: bool = True,
        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
205
        stat_loggers: list[StatLoggerFactory] | None = None,
206
        enable_log_requests: bool = False,
207
        aggregate_engine_logging: bool = False,
208
        disable_log_stats: bool = False,
209
        client_addresses: dict[str, str] | None = None,
210
211
212
        client_count: int = 1,
        client_index: int = 0,
        disable_log_requests: bool = True,  # Deprecated, will be removed
213
214
215
216
217
218
    ) -> "AsyncLLM":
        # Create the LLMEngine.
        return cls(
            vllm_config=vllm_config,
            executor_class=Executor.get_class(vllm_config),
            start_engine_loop=start_engine_loop,
219
            stat_loggers=stat_loggers,
220
            log_requests=enable_log_requests,
221
            log_stats=not disable_log_stats,
222
            aggregate_engine_logging=aggregate_engine_logging,
223
            usage_context=usage_context,
224
            client_addresses=client_addresses,
225
            client_count=client_count,
226
            client_index=client_index,
227
228
        )

229
230
231
232
233
234
    @classmethod
    def from_engine_args(
        cls,
        engine_args: AsyncEngineArgs,
        start_engine_loop: bool = True,
        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
235
        stat_loggers: list[StatLoggerFactory] | None = None,
236
    ) -> "AsyncLLM":
237
238
239
        """Create an AsyncLLM from the EngineArgs."""

        # Create the engine configs.
240
        vllm_config = engine_args.create_engine_config(usage_context)
241
        executor_class = Executor.get_class(vllm_config)
242
243
244
245
246

        # Create the AsyncLLM.
        return cls(
            vllm_config=vllm_config,
            executor_class=executor_class,
247
            log_requests=engine_args.enable_log_requests,
248
249
250
            log_stats=not engine_args.disable_log_stats,
            start_engine_loop=start_engine_loop,
            usage_context=usage_context,
251
            stat_loggers=stat_loggers,
252
253
        )

254
255
256
    def __del__(self):
        self.shutdown()

257
258
259
    def shutdown(self):
        """Shutdown, cleaning up the background proc and IPC."""

260
261
        shutdown_prometheus()

262
263
        if engine_core := getattr(self, "engine_core", None):
            engine_core.shutdown()
264

265
266
267
        handler = getattr(self, "output_handler", None)
        if handler is not None:
            cancel_task_threadsafe(handler)
268

269
270
271
    async def get_supported_tasks(self) -> tuple[SupportedTask, ...]:
        return await self.engine_core.get_supported_tasks_async()

272
273
274
    async def add_request(
        self,
        request_id: str,
275
276
277
278
279
280
        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,
281
        priority: int = 0,
282
283
        data_parallel_rank: int | None = None,
        prompt_text: str | None = None,
284
    ) -> RequestOutputCollector:
285
286
        """Add new request to the AsyncLLM."""

287
288
289
        if self.errored:
            raise EngineDeadError()

290
        is_pooling = isinstance(params, PoolingParams)
291
292
293

        # Create a new output collector for the request.
        queue = RequestOutputCollector(output_kind=params.output_kind)
294

295
        # Convert Input --> Request.
296
297
298
299
300
301
        if isinstance(prompt, EngineCoreRequest):
            request = prompt
        else:
            assert prompt_text is None
            logger.warning_once(
                "Processor has been moved under OpenAIServing and will "
302
303
304
305
306
307
308
309
310
311
312
313
314
                "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,
            )
315
316
317
318
            if isinstance(prompt, str):
                prompt_text = prompt
            elif isinstance(prompt, Mapping):
                prompt_text = cast(str | None, prompt.get("prompt"))
319

320
        if is_pooling or params.n == 1:
321
            await self._add_request(request, prompt_text, None, 0, queue)
322
323
            return queue

324
325
326
327
328
        # 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

329
        # Fan out child requests (for n>1).
330
331
332
        parent_request = ParentRequest(request_id, parent_params)
        for idx in range(parent_params.n):
            request_id, child_params = parent_request.get_child_info(idx)
333
            child_request = request if idx == parent_params.n - 1 else copy(request)
334
            child_request.request_id = request_id
335
            child_request.sampling_params = child_params
336
337
338
            await self._add_request(
                child_request, prompt_text, parent_request, idx, queue
            )
339
        return queue
340

341
342
343
    async def _add_request(
        self,
        request: EngineCoreRequest,
344
345
        prompt: str | None,
        parent_req: ParentRequest | None,
346
347
348
        index: int,
        queue: RequestOutputCollector,
    ):
349
        # Add the request to OutputProcessor (this process).
350
        self.output_processor.add_request(request, prompt, parent_req, index, queue)
351

352
353
        # Add the EngineCoreRequest to EngineCore (separate process).
        await self.engine_core.add_request_async(request)
354

355
356
        if self.log_requests:
            logger.info("Added request %s.", request.request_id)
357
358
359
360
361
362

    # 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.
363
    async def generate(
364
        self,
365
        prompt: EngineCoreRequest | PromptType,
366
367
        sampling_params: SamplingParams,
        request_id: str,
368
        *,
369
370
371
372
        prompt_text: str | None = None,
        lora_request: LoRARequest | None = None,
        tokenization_kwargs: dict[str, Any] | None = None,
        trace_headers: Mapping[str, str] | None = None,
373
        priority: int = 0,
374
        data_parallel_rank: int | None = None,
375
376
377
378
    ) -> AsyncGenerator[RequestOutput, None]:
        """
        Main function called by the API server to kick off a request
            * 1) Making an AsyncStream corresponding to the Request.
379
            * 2) Processing the Input.
380
381
382
            * 3) Adding the Request to the Detokenizer.
            * 4) Adding the Request to the EngineCore (separate process).

383
384
        A separate output_handler loop runs in a background AsyncIO task,
        pulling outputs from EngineCore and putting them into the
385
386
387
388
389
390
        per-request AsyncStream.

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

391
392
393
394
        if (
            self.vllm_config.cache_config.kv_sharing_fast_prefill
            and sampling_params.prompt_logprobs
        ):
395
396
397
            raise ValueError(
                "--kv-sharing-fast-prefill produces incorrect logprobs for "
                "prompt tokens, please disable it when the requests need "
398
399
                "prompt logprobs"
            )
400

401
402
403
404
        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.
405
            self._run_output_handler()
406

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

417
418
419
420
421
422
423
424
425
426
427
            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,
            )
428

429
430
            # The output_handler task pushes items into the queue.
            # This task pulls from the queue and yields to caller.
431
432
            finished = False
            while not finished:
433
434
                # Note: drain queue without await if possible (avoids
                # task switching under load which helps performance).
435
                out = q.get_nowait() or await q.get()
436

437
                # Note: both OutputProcessor and EngineCore handle their
438
                # own request cleanup based on finished.
439
                finished = out.finished
440
                assert isinstance(out, RequestOutput)
441
442
                yield out

443
        # If the request is disconnected by the client, generate()
444
445
446
        # is cancelled or the generator is garbage collected. So,
        # we abort the request if we end up here.
        except (asyncio.CancelledError, GeneratorExit):
447
            await self.abort(request_id)
448
449
            if self.log_requests:
                logger.info("Request %s aborted.", request_id)
450
            raise
451

452
453
454
455
456
        # 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
457

458
459
460
461
462
        # Request validation error.
        except ValueError:
            if self.log_requests:
                logger.info("Request %s failed (bad request).", request_id)
            raise
463

464
        # Unexpected error in the generate() task (possibly recoverable).
465
        except Exception as e:
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
            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
482
        logger_manager = self.logger_manager
483
        processor = self.processor
484
485
486
487
488
489
490
491

        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)

492
493
494
                    iteration_stats = (
                        IterationStats() if (log_stats and num_outputs) else None
                    )
495
496
497
498

                    # 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.
499
                    if num_outputs <= envs.VLLM_V1_OUTPUT_PROC_CHUNK_SIZE:
500
                        slices = (outputs.outputs,)
501
502
503
                    else:
                        slices = np.array_split(
                            outputs.outputs,
504
                            cdiv(num_outputs, envs.VLLM_V1_OUTPUT_PROC_CHUNK_SIZE),
505
                        )
506
507
508
509

                    for i, outputs_slice in enumerate(slices):
                        # 2) Process EngineCoreOutputs.
                        processed_outputs = output_processor.process_outputs(
510
511
                            outputs_slice, outputs.timestamp, iteration_stats
                        )
512
513
514
515
516
517
518
519
520
                        # 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(
521
522
                            processed_outputs.reqs_to_abort
                        )
523

524
525
                    output_processor.update_scheduler_stats(outputs.scheduler_stats)

526
527
528
                    # 4) Logging.
                    # TODO(rob): make into a coroutine and launch it in
                    # background thread once Prometheus overhead is non-trivial.
529
530
531
                    if logger_manager:
                        logger_manager.record(
                            engine_idx=outputs.engine_index,
532
533
                            scheduler_stats=outputs.scheduler_stats,
                            iteration_stats=iteration_stats,
534
                            mm_cache_stats=processor.stat_mm_cache(),
535
536
537
538
539
540
                        )
            except Exception as e:
                logger.exception("AsyncLLM output_handler failed.")
                output_processor.propagate_error(e)

        self.output_handler = asyncio.create_task(output_handler())
541

542
    async def abort(self, request_id: str | Iterable[str]) -> None:
543
        """Abort RequestId in OutputProcessor and EngineCore."""
544

545
546
547
        request_ids = (
            (request_id,) if isinstance(request_id, str) else as_list(request_id)
        )
548
549
        all_request_ids = self.output_processor.abort_requests(request_ids)
        await self.engine_core.abort_requests_async(all_request_ids)
550

551
        if self.log_requests:
552
            logger.info("Aborted request(s) %s.", ",".join(request_ids))
553

554
    async def encode(
555
556
557
558
        self,
        prompt: PromptType,
        pooling_params: PoolingParams,
        request_id: str,
559
560
        lora_request: LoRARequest | None = None,
        trace_headers: Mapping[str, str] | None = None,
561
        priority: int = 0,
562
563
        truncate_prompt_tokens: int | None = None,
        tokenization_kwargs: dict[str, Any] | None = None,
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
    ) -> 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()

585
            if tokenization_kwargs is None:
586
                tokenization_kwargs = {}
587
588
589
590
591
592
            _validate_truncation_size(
                self.model_config.max_model_len,
                truncate_prompt_tokens,
                tokenization_kwargs,
            )

593
594
595
596
597
            q = await self.add_request(
                request_id,
                prompt,
                pooling_params,
                lora_request=lora_request,
598
                tokenization_kwargs=tokenization_kwargs,
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
630
631
632
633
634
635
636
637
638
639
640
641
                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
642

643
    @property
644
    def tokenizer(self) -> AnyTokenizer | None:
645
        return self.processor.tokenizer
646

647
    @tokenizer.setter
648
    def tokenizer(self, tokenizer: AnyTokenizer | None) -> None:
649
        self.processor.tokenizer = tokenizer
650

651
    async def get_tokenizer(self) -> AnyTokenizer:
652
        if self.tokenizer is None:
653
654
655
            raise ValueError(
                "Unable to get tokenizer because skip_tokenizer_init is True"
            )
656

657
        return self.tokenizer
658
659

    async def is_tracing_enabled(self) -> bool:
660
        return self.observability_config.otlp_traces_endpoint is not None  # type: ignore
661

662
    async def do_log_stats(self) -> None:
663
664
        if self.logger_manager:
            self.logger_manager.log()
665
666
667

    async def check_health(self) -> None:
        logger.debug("Called check_health.")
668
669
        if self.errored:
            raise self.dead_error
670
671

    async def start_profile(self) -> None:
672
673
674
675
        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)
676
677

    async def stop_profile(self) -> None:
678
679
680
681
        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)
682

683
    async def reset_mm_cache(self) -> None:
684
        self.processor.clear_mm_cache()
685
686
        await self.engine_core.reset_mm_cache_async()

687
    async def reset_prefix_cache(self) -> None:
688
689
        await self.engine_core.reset_prefix_cache_async()

690
    async def sleep(self, level: int = 1) -> None:
691
        await self.reset_prefix_cache()
692
693
        await self.engine_core.sleep_async(level)

694
695
696
        if self.logger_manager is not None:
            self.logger_manager.record_sleep_state(1, level)

697
    async def wake_up(self, tags: list[str] | None = None) -> None:
698
        await self.engine_core.wake_up_async(tags)
699

700
701
702
        if self.logger_manager is not None:
            self.logger_manager.record_sleep_state(0, 0)

703
704
705
    async def is_sleeping(self) -> bool:
        return await self.engine_core.is_sleeping_async()

706
    async def add_lora(self, lora_request: LoRARequest) -> bool:
707
        """Load a new LoRA adapter into the engine for future requests."""
708
709
710
711
712
713
        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)

714
    async def list_loras(self) -> set[int]:
715
716
717
718
719
720
        """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)
721

722
723
724
    async def collective_rpc(
        self,
        method: str,
725
        timeout: float | None = None,
726
        args: tuple = (),
727
        kwargs: dict | None = None,
728
    ):
729
730
731
732
        """
        Perform a collective RPC call to the given path.
        """
        return await self.engine_core.collective_rpc_async(
733
734
            method, timeout, args, kwargs
        )
735

736
737
738
739
740
741
742
743
    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

744
            logger.info("Engines are still running, waiting for requests to drain...")
745
746
            await asyncio.sleep(1)  # Wait 1 second before checking again

747
748
749
750
        raise TimeoutError(
            f"Timeout reached after {drain_timeout} seconds "
            "waiting for requests to drain."
        )
751

752
753
754
    async def scale_elastic_ep(
        self, new_data_parallel_size: int, drain_timeout: int = 300
    ):
755
756
757
758
759
760
761
762
        """
        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)
        """
763
        old_data_parallel_size = self.vllm_config.parallel_config.data_parallel_size
764
        if old_data_parallel_size == new_data_parallel_size:
765
766
767
768
            logger.info(
                "Data parallel size is already %s, skipping scale",
                new_data_parallel_size,
            )
769
770
            return
        logger.info(
771
772
773
            "Waiting for requests to drain before scaling up to %s engines...",
            new_data_parallel_size,
        )
774
775
        await self.wait_for_requests_to_drain(drain_timeout)
        logger.info(
776
777
778
            "Requests have been drained, proceeding with scale to %s engines",
            new_data_parallel_size,
        )
779
        await self.engine_core.scale_elastic_ep(new_data_parallel_size)
780
        self.vllm_config.parallel_config.data_parallel_size = new_data_parallel_size
781
782

        # recreate stat loggers
783
784
785
786
787
788
        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(
789
                vllm_config=self.vllm_config,
790
                engine_idxs=list(range(new_data_parallel_size)),
791
792
793
                custom_stat_loggers=None,
            )

794
795
    @property
    def is_running(self) -> bool:
796
797
        # Is None before the loop is started.
        return self.output_handler is None or not self.output_handler.done()
798
799
800

    @property
    def is_stopped(self) -> bool:
801
        return self.errored
802
803
804

    @property
    def errored(self) -> bool:
805
        return self.engine_core.resources.engine_dead or not self.is_running
806
807
808

    @property
    def dead_error(self) -> BaseException:
809
        return EngineDeadError()