async_llm.py 33.4 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
import warnings
8
from collections.abc import AsyncGenerator, Iterable, Mapping
9
from copy import copy
10
from typing import Any, cast
11

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.tokenizers import TokenizerLike, cached_tokenizer_from_config
29
from vllm.tracing import init_tracer
30
from vllm.transformers_utils.config import maybe_register_config_serialize_by_value
31
from vllm.usage.usage_lib import UsageContext
32
33
from vllm.utils.async_utils import cancel_task_threadsafe
from vllm.utils.collection_utils import as_list
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.input_processor import InputProcessor
38
from vllm.v1.engine.output_processor import OutputProcessor, RequestOutputCollector
39
from vllm.v1.engine.parallel_sampling import ParentRequest
40
from vllm.v1.executor import Executor
41
42
43
44
45
from vllm.v1.metrics.loggers import (
    StatLoggerFactory,
    StatLoggerManager,
    load_stat_logger_plugin_factories,
)
46
from vllm.v1.metrics.prometheus import shutdown_prometheus
47
from vllm.v1.metrics.stats import IterationStats
48
49
50
51
52
53
54
55

logger = init_logger(__name__)


class AsyncLLM(EngineClient):
    def __init__(
        self,
        vllm_config: VllmConfig,
56
        executor_class: type[Executor],
57
58
        log_stats: bool,
        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
59
        mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
60
61
62
        use_cached_outputs: bool = False,
        log_requests: bool = True,
        start_engine_loop: bool = True,
63
        stat_loggers: list[StatLoggerFactory] | None = None,
64
        aggregate_engine_logging: bool = False,
65
        client_addresses: dict[str, str] | None = None,
66
        client_count: int = 1,
67
        client_index: int = 0,
68
    ) -> None:
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
        """
        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
        """
89
90
91
        # Ensure we can serialize custom transformer configs
        maybe_register_config_serialize_by_value()

92
        self.model_config = vllm_config.model_config
93
        self.vllm_config = vllm_config
94
        self.observability_config = vllm_config.observability_config
95
        self.log_requests = log_requests
96

97
98
99
100
101
102
        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:
103
            logger.info(
104
105
106
                "AsyncLLM created with log_stats=False, "
                "but custom stat loggers were found; "
                "enabling logging without default stat loggers."
107
            )
108

109
        tokenizer = cached_tokenizer_from_config(self.model_config)
110

111
        self.input_processor = InputProcessor(self.vllm_config, tokenizer)
112
113
        self.io_processor = get_io_processor(
            self.vllm_config,
114
            self.model_config.io_processor_plugin,
115
        )
116

117
        # OutputProcessor (converts EngineCoreOutputs --> RequestOutput).
118
        self.output_processor = OutputProcessor(
119
120
121
            self.tokenizer,
            log_stats=self.log_stats,
            stream_interval=self.vllm_config.scheduler_config.stream_interval,
122
        )
123
124
125
        endpoint = self.observability_config.otlp_traces_endpoint
        if endpoint is not None:
            tracer = init_tracer("vllm.llm_engine", endpoint)
126
            self.output_processor.tracer = tracer
127
128

        # EngineCore (starts the engine in background process).
129
        self.engine_core = EngineCoreClient.make_async_mp_client(
130
131
            vllm_config=vllm_config,
            executor_class=executor_class,
132
            log_stats=self.log_stats,
133
            client_addresses=client_addresses,
134
            client_count=client_count,
135
            client_index=client_index,
136
        )
137
138

        # Loggers.
139
        self.logger_manager: StatLoggerManager | None = None
140
141
142
        if self.log_stats:
            self.logger_manager = StatLoggerManager(
                vllm_config=vllm_config,
143
                engine_idxs=self.engine_core.engine_ranks_managed,
144
                custom_stat_loggers=custom_stat_loggers,
145
                enable_default_loggers=log_stats,
146
                client_count=client_count,
147
                aggregate_engine_logging=aggregate_engine_logging,
148
149
150
            )
            self.logger_manager.log_engine_initialized()

151
152
153
154
        # Pause / resume state for async RL workflows.
        self._pause_cond = asyncio.Condition()
        self._paused = False

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
        if (
164
165
            vllm_config.profiler_config.profiler == "torch"
            and not vllm_config.profiler_config.ignore_frontend
166
        ):
167
            profiler_dir = vllm_config.profiler_config.torch_profiler_dir
168
169
            logger.info(
                "Torch profiler enabled. AsyncLLM CPU traces will be collected under %s",  # noqa: E501
170
                profiler_dir,
171
            )
172
173
174
175
176
            worker_name = f"{socket.gethostname()}_{os.getpid()}.async_llm"
            self.profiler = torch.profiler.profile(
                activities=[
                    torch.profiler.ProfilerActivity.CPU,
                ],
177
                with_stack=vllm_config.profiler_config.torch_profiler_with_stack,
178
                on_trace_ready=torch.profiler.tensorboard_trace_handler(
179
                    profiler_dir,
180
                    worker_name=worker_name,
181
                    use_gzip=vllm_config.profiler_config.torch_profiler_use_gzip,
182
183
                ),
            )
184
185
186
        else:
            self.profiler = None

187
188
    @classmethod
    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
        client_count: int = 1,
        client_index: int = 0,
200
201
202
203
204
205
    ) -> "AsyncLLM":
        # Create the LLMEngine.
        return cls(
            vllm_config=vllm_config,
            executor_class=Executor.get_class(vllm_config),
            start_engine_loop=start_engine_loop,
206
            stat_loggers=stat_loggers,
207
            log_requests=enable_log_requests,
208
            log_stats=not disable_log_stats,
209
            aggregate_engine_logging=aggregate_engine_logging,
210
            usage_context=usage_context,
211
            client_addresses=client_addresses,
212
            client_count=client_count,
213
            client_index=client_index,
214
215
        )

216
217
218
219
220
221
    @classmethod
    def from_engine_args(
        cls,
        engine_args: AsyncEngineArgs,
        start_engine_loop: bool = True,
        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
222
        stat_loggers: list[StatLoggerFactory] | None = None,
223
    ) -> "AsyncLLM":
224
225
226
        """Create an AsyncLLM from the EngineArgs."""

        # Create the engine configs.
227
        vllm_config = engine_args.create_engine_config(usage_context)
228
        executor_class = Executor.get_class(vllm_config)
229
230
231
232
233

        # Create the AsyncLLM.
        return cls(
            vllm_config=vllm_config,
            executor_class=executor_class,
234
            log_requests=engine_args.enable_log_requests,
235
236
237
            log_stats=not engine_args.disable_log_stats,
            start_engine_loop=start_engine_loop,
            usage_context=usage_context,
238
            stat_loggers=stat_loggers,
239
240
        )

241
242
243
    def __del__(self):
        self.shutdown()

244
245
246
    def shutdown(self):
        """Shutdown, cleaning up the background proc and IPC."""

247
248
        shutdown_prometheus()

249
250
        if engine_core := getattr(self, "engine_core", None):
            engine_core.shutdown()
251

252
253
254
        if input_processor := getattr(self, "input_processor", None):
            input_processor.close()

255
256
257
        handler = getattr(self, "output_handler", None)
        if handler is not None:
            cancel_task_threadsafe(handler)
258

259
260
261
    async def get_supported_tasks(self) -> tuple[SupportedTask, ...]:
        return await self.engine_core.get_supported_tasks_async()

262
263
264
    async def add_request(
        self,
        request_id: str,
265
266
267
268
269
270
        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,
271
        priority: int = 0,
272
273
        data_parallel_rank: int | None = None,
        prompt_text: str | None = None,
274
    ) -> RequestOutputCollector:
275
276
        """Add new request to the AsyncLLM."""

277
278
279
        if self.errored:
            raise EngineDeadError()

280
        is_pooling = isinstance(params, PoolingParams)
281

282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
        if (
            self.vllm_config.cache_config.kv_sharing_fast_prefill
            and not is_pooling
            and params.prompt_logprobs
        ):
            raise ValueError(
                "--kv-sharing-fast-prefill produces incorrect logprobs for "
                "prompt tokens, please disable it when the requests need "
                "prompt logprobs"
            )

        if tokenization_kwargs is None:
            tokenization_kwargs = {}
        _validate_truncation_size(
            self.model_config.max_model_len,
            params.truncate_prompt_tokens,
            tokenization_kwargs,
        )

301
        # Convert Input --> Request.
302
303
        if isinstance(prompt, EngineCoreRequest):
            request = prompt
304
305
306
307
308
309
            if request_id != request.request_id:
                logger.warning_once(
                    "AsyncLLM.add_request() was passed a request_id parameter that "
                    "does not match the EngineCoreRequest.request_id attribute. The "
                    "latter will be used, and the former will be ignored."
                )
310
        else:
311
312
313
314
            if prompt_text is not None:
                raise ValueError(
                    "should only provide prompt_text with EngineCoreRequest"
                )
315
            request = self.input_processor.process_inputs(
316
317
318
319
320
321
322
323
324
325
                request_id,
                prompt,
                params,
                arrival_time,
                lora_request,
                tokenization_kwargs,
                trace_headers,
                priority,
                data_parallel_rank,
            )
326
327
328
329
            if isinstance(prompt, str):
                prompt_text = prompt
            elif isinstance(prompt, Mapping):
                prompt_text = cast(str | None, prompt.get("prompt"))
330

331
332
        self.input_processor.assign_request_id(request)

333
334
335
336
337
338
339
340
341
        # We start the output_handler on the first call to add_request() 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()

        # Respect pause state before accepting new requests.
        async with self._pause_cond:
            await self._pause_cond.wait_for(lambda: not self._paused)

342
343
344
        # Create a new output collector for the request.
        queue = RequestOutputCollector(params.output_kind, request.request_id)

345
346
347
        # Use cloned params that may have been updated in process_inputs()
        params = request.params

348
        if is_pooling or params.n == 1:
349
            await self._add_request(request, prompt_text, None, 0, queue)
350
351
            return queue

352
353
        parent_params = params
        assert isinstance(parent_params, SamplingParams)
354

355
        # Fan out child requests (for n>1).
356
        parent_request = ParentRequest(request)
357
358
        for idx in range(parent_params.n):
            request_id, child_params = parent_request.get_child_info(idx)
359
            child_request = request if idx == parent_params.n - 1 else copy(request)
360
            child_request.request_id = request_id
361
            child_request.sampling_params = child_params
362
363
364
            await self._add_request(
                child_request, prompt_text, parent_request, idx, queue
            )
365
        return queue
366

367
368
369
    async def _add_request(
        self,
        request: EngineCoreRequest,
370
371
        prompt: str | None,
        parent_req: ParentRequest | None,
372
373
374
        index: int,
        queue: RequestOutputCollector,
    ):
375
        # Add the request to OutputProcessor (this process).
376
        self.output_processor.add_request(request, prompt, parent_req, index, queue)
377

378
379
        # Add the EngineCoreRequest to EngineCore (separate process).
        await self.engine_core.add_request_async(request)
380

381
382
        if self.log_requests:
            logger.info("Added request %s.", request.request_id)
383
384
385
386
387
388

    # 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.
389
    async def generate(
390
        self,
391
        prompt: EngineCoreRequest | PromptType,
392
393
        sampling_params: SamplingParams,
        request_id: str,
394
        *,
395
396
397
398
        prompt_text: str | None = None,
        lora_request: LoRARequest | None = None,
        tokenization_kwargs: dict[str, Any] | None = None,
        trace_headers: Mapping[str, str] | None = None,
399
        priority: int = 0,
400
        data_parallel_rank: int | None = None,
401
402
403
404
    ) -> AsyncGenerator[RequestOutput, None]:
        """
        Main function called by the API server to kick off a request
            * 1) Making an AsyncStream corresponding to the Request.
405
            * 2) Processing the Input.
406
407
408
            * 3) Adding the Request to the Detokenizer.
            * 4) Adding the Request to the EngineCore (separate process).

409
410
        A separate output_handler loop runs in a background AsyncIO task,
        pulling outputs from EngineCore and putting them into the
411
412
413
414
415
416
        per-request AsyncStream.

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

417
        q: RequestOutputCollector | None = None
418
        try:
419
420
421
422
423
424
425
426
427
428
429
            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,
            )
430

431
432
            # The output_handler task pushes items into the queue.
            # This task pulls from the queue and yields to caller.
433
434
            finished = False
            while not finished:
435
436
                # Note: drain queue without await if possible (avoids
                # task switching under load which helps performance).
437
                out = q.get_nowait() or await q.get()
438

439
                # Note: both OutputProcessor and EngineCore handle their
440
                # own request cleanup based on finished.
441
                finished = out.finished
442
                assert isinstance(out, RequestOutput)
443
444
                yield out

445
        # If the request is disconnected by the client, generate()
446
447
448
        # is cancelled or the generator is garbage collected. So,
        # we abort the request if we end up here.
        except (asyncio.CancelledError, GeneratorExit):
449
450
            if q is not None:
                await self.abort(q.request_id, internal=True)
451
452
            if self.log_requests:
                logger.info("Request %s aborted.", request_id)
453
            raise
454

455
456
457
458
459
        # 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
460

461
        # Request validation error.
462
        except ValueError as e:
463
            if self.log_requests:
464
                logger.info("Request %s failed (bad request): %s.", request_id, e)
465
            raise
466

467
        # Unexpected error in the generate() task (possibly recoverable).
468
        except Exception as e:
469
470
            if q is not None:
                await self.abort(q.request_id, internal=True)
471
            if self.log_requests:
472
473
474
475
476
477
478
479
480
                try:
                    s = f"{e.__class__.__name__}: {e}"
                except Exception as e2:
                    s = (
                        f"{e.__class__.__name__}: "
                        + "error during printing an exception of class"
                        + e2.__class__.__name__
                    )
                logger.info("Request %s failed due to %s.", request_id, s)
481
482
483
484
485
486
487
488
489
490
491
492
493
            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
494
        logger_manager = self.logger_manager
495
        input_processor = self.input_processor
496
        chunk_size = envs.VLLM_V1_OUTPUT_PROC_CHUNK_SIZE
497
498
499
500
501
502
503
504

        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)

505
506
507
                    iteration_stats = (
                        IterationStats() if (log_stats and num_outputs) else None
                    )
508
509
510
511

                    # 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.
512
513
514
515
                    engine_core_outputs = outputs.outputs
                    for start in range(0, num_outputs, chunk_size):
                        end = start + chunk_size
                        outputs_slice = engine_core_outputs[start:end]
516
517
                        # 2) Process EngineCoreOutputs.
                        processed_outputs = output_processor.process_outputs(
518
519
                            outputs_slice, outputs.timestamp, iteration_stats
                        )
520
521
522
523
                        # NOTE: RequestOutputs are pushed to their queues.
                        assert not processed_outputs.request_outputs

                        # Allow other asyncio tasks to run between chunks
524
                        if end < num_outputs:
525
526
527
                            await asyncio.sleep(0)

                        # 3) Abort any reqs that finished due to stop strings.
528
529
530
531
                        if processed_outputs.reqs_to_abort:
                            await engine_core.abort_requests_async(
                                processed_outputs.reqs_to_abort
                            )
532

533
534
                    output_processor.update_scheduler_stats(outputs.scheduler_stats)

535
536
537
                    # 4) Logging.
                    # TODO(rob): make into a coroutine and launch it in
                    # background thread once Prometheus overhead is non-trivial.
538
539
540
                    if logger_manager:
                        logger_manager.record(
                            engine_idx=outputs.engine_index,
541
542
                            scheduler_stats=outputs.scheduler_stats,
                            iteration_stats=iteration_stats,
543
                            mm_cache_stats=input_processor.stat_mm_cache(),
544
545
546
547
548
549
                        )
            except Exception as e:
                logger.exception("AsyncLLM output_handler failed.")
                output_processor.propagate_error(e)

        self.output_handler = asyncio.create_task(output_handler())
550

551
552
553
    async def abort(
        self, request_id: str | Iterable[str], internal: bool = False
    ) -> None:
554
        """Abort RequestId in OutputProcessor and EngineCore."""
555

556
557
558
        request_ids = (
            (request_id,) if isinstance(request_id, str) else as_list(request_id)
        )
559
        all_request_ids = self.output_processor.abort_requests(request_ids, internal)
560
        await self.engine_core.abort_requests_async(all_request_ids)
561

562
        if self.log_requests:
563
            logger.info("Aborted request(s) %s.", ",".join(request_ids))
564

565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
    async def pause_generation(
        self,
        *,
        wait_for_inflight_requests: bool = False,
        clear_cache: bool = True,
    ) -> None:
        """
        Pause generation to allow model weight updates.

        New generation/encoding requests are blocked until resume.

        Args:
            wait_for_inflight_requests: When ``True`` waits for in-flight
                requests to finish before pausing. When ``False`` (default),
                immediately aborts any in-flight requests.
            clear_cache: Whether to clear KV cache and prefix cache after
                draining. Set to ``False`` to preserve cache for faster resume.
                Default is ``True`` (clear caches).
        """

        async with self._pause_cond:
            if self._paused:
                return
            self._paused = True

        if not wait_for_inflight_requests:
            request_ids = list(self.output_processor.request_states.keys())
            if request_ids:
593
                await self.abort(request_ids, internal=True)
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616

        # Wait for running requests to drain before clearing cache.
        if self.output_processor.has_unfinished_requests():
            await self.output_processor.wait_for_requests_to_drain()

        # Clear cache
        if clear_cache:
            await self.reset_prefix_cache()
            await self.reset_mm_cache()

    async def resume_generation(self) -> None:
        """Resume generation after :meth:`pause_generation`."""

        async with self._pause_cond:
            self._paused = False
            self._pause_cond.notify_all()  # Wake up all waiting requests

    async def is_paused(self) -> bool:
        """Return whether the engine is currently paused."""

        async with self._pause_cond:
            return self._paused

617
    async def encode(
618
619
620
621
        self,
        prompt: PromptType,
        pooling_params: PoolingParams,
        request_id: str,
622
623
        lora_request: LoRARequest | None = None,
        trace_headers: Mapping[str, str] | None = None,
624
        priority: int = 0,
625
626
        truncate_prompt_tokens: int | None = None,
        tokenization_kwargs: dict[str, Any] | None = None,
627
628
629
630
631
632
633
634
635
636
637
638
639
    ) -> 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.
640
641
642

        NOTE: truncate_prompt_tokens is deprecated in v0.14.
        TODO: Remove truncate_prompt_tokens in v0.15.
643
644
        """

645
        q: RequestOutputCollector | None = None
646
        try:
647
648
649
650
651
652
653
654
655
            if truncate_prompt_tokens is not None:
                warnings.warn(
                    "The `truncate_prompt_tokens` parameter in `AsyncLLM.encode()` "
                    "is deprecated and will be removed in v0.15. "
                    "Please use `pooling_params.truncate_prompt_tokens` instead.",
                    DeprecationWarning,
                    stacklevel=2,
                )

656
657
658
659
660
            q = await self.add_request(
                request_id,
                prompt,
                pooling_params,
                lora_request=lora_request,
661
                tokenization_kwargs=tokenization_kwargs,
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
                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:
682
683
            if q is not None:
                await self.abort(q.request_id, internal=True)
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
            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:
702
703
            if q is not None:
                await self.abort(q.request_id, internal=True)
704
705
706
            if self.log_requests:
                logger.info("Request %s failed.", request_id)
            raise EngineGenerateError() from e
707

708
    @property
709
    def tokenizer(self) -> TokenizerLike | None:
710
        return self.input_processor.tokenizer
711

712
    async def get_tokenizer(self) -> TokenizerLike:
713
        if self.tokenizer is None:
714
            raise ValueError(
715
                "Unable to get tokenizer because `skip_tokenizer_init=True`"
716
            )
717

718
        return self.tokenizer
719
720

    async def is_tracing_enabled(self) -> bool:
721
        return self.observability_config.otlp_traces_endpoint is not None  # type: ignore
722

723
    async def do_log_stats(self) -> None:
724
725
        if self.logger_manager:
            self.logger_manager.log()
726
727
728

    async def check_health(self) -> None:
        logger.debug("Called check_health.")
729
730
        if self.errored:
            raise self.dead_error
731
732

    async def start_profile(self) -> None:
733
734
735
736
        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)
737
738

    async def stop_profile(self) -> None:
739
740
741
742
        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)
743

744
    async def reset_mm_cache(self) -> None:
745
        self.input_processor.clear_mm_cache()
746
747
        await self.engine_core.reset_mm_cache_async()

748
749
750
751
752
753
    async def reset_prefix_cache(
        self, reset_running_requests: bool = False, reset_connector: bool = False
    ) -> bool:
        return await self.engine_core.reset_prefix_cache_async(
            reset_running_requests, reset_connector
        )
754

755
    async def sleep(self, level: int = 1) -> None:
756
        await self.reset_prefix_cache()
757
758
        await self.engine_core.sleep_async(level)

759
760
761
        if self.logger_manager is not None:
            self.logger_manager.record_sleep_state(1, level)

762
    async def wake_up(self, tags: list[str] | None = None) -> None:
763
        await self.engine_core.wake_up_async(tags)
764

765
766
767
        if self.logger_manager is not None:
            self.logger_manager.record_sleep_state(0, 0)

768
769
770
    async def is_sleeping(self) -> bool:
        return await self.engine_core.is_sleeping_async()

771
    async def add_lora(self, lora_request: LoRARequest) -> bool:
772
        """Load a new LoRA adapter into the engine for future requests."""
773
774
775
776
777
778
        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)

779
    async def list_loras(self) -> set[int]:
780
781
782
783
784
785
        """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)
786

787
788
789
    async def collective_rpc(
        self,
        method: str,
790
        timeout: float | None = None,
791
        args: tuple = (),
792
        kwargs: dict | None = None,
793
    ):
794
795
796
797
        """
        Perform a collective RPC call to the given path.
        """
        return await self.engine_core.collective_rpc_async(
798
799
            method, timeout, args, kwargs
        )
800

801
802
803
804
805
806
807
808
    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

809
            logger.info("Engines are still running, waiting for requests to drain...")
810
811
            await asyncio.sleep(1)  # Wait 1 second before checking again

812
813
814
815
        raise TimeoutError(
            f"Timeout reached after {drain_timeout} seconds "
            "waiting for requests to drain."
        )
816

817
818
819
    async def scale_elastic_ep(
        self, new_data_parallel_size: int, drain_timeout: int = 300
    ):
820
821
822
823
824
825
826
827
        """
        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)
        """
828
        old_data_parallel_size = self.vllm_config.parallel_config.data_parallel_size
829
        if old_data_parallel_size == new_data_parallel_size:
830
831
832
833
            logger.info(
                "Data parallel size is already %s, skipping scale",
                new_data_parallel_size,
            )
834
835
            return
        logger.info(
836
837
838
            "Waiting for requests to drain before scaling up to %s engines...",
            new_data_parallel_size,
        )
839
840
        await self.wait_for_requests_to_drain(drain_timeout)
        logger.info(
841
842
843
            "Requests have been drained, proceeding with scale to %s engines",
            new_data_parallel_size,
        )
844
        await self.engine_core.scale_elastic_ep(new_data_parallel_size)
845
        self.vllm_config.parallel_config.data_parallel_size = new_data_parallel_size
846
847

        # recreate stat loggers
848
849
850
851
852
853
        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(
854
                vllm_config=self.vllm_config,
855
                engine_idxs=list(range(new_data_parallel_size)),
856
857
858
                custom_stat_loggers=None,
            )

859
860
    @property
    def is_running(self) -> bool:
861
862
        # Is None before the loop is started.
        return self.output_handler is None or not self.output_handler.done()
863
864
865

    @property
    def is_stopped(self) -> bool:
866
        return self.errored
867
868
869

    @property
    def errored(self) -> bool:
870
        return self.engine_core.resources.engine_dead or not self.is_running
871
872
873

    @property
    def dead_error(self) -> BaseException:
874
        return EngineDeadError()