async_llm_engine.py 25.8 KB
Newer Older
1
import asyncio
2
import os
3
import time
Antoni Baum's avatar
Antoni Baum committed
4
from functools import partial
5
6
from typing import (Callable, Dict, Iterable, List, Optional, Set, Tuple, Type,
                    Union, AsyncIterator)
7

8
9
from transformers import PreTrainedTokenizer

10
from vllm.lora.request import LoRARequest
11
from vllm.config import ModelConfig
Woosuk Kwon's avatar
Woosuk Kwon committed
12
13
from vllm.engine.arg_utils import AsyncEngineArgs
from vllm.engine.llm_engine import LLMEngine
14
from vllm.engine.ray_utils import initialize_ray_cluster, ray
Woosuk Kwon's avatar
Woosuk Kwon committed
15
16
17
from vllm.logger import init_logger
from vllm.outputs import RequestOutput
from vllm.sampling_params import SamplingParams
18
19

logger = init_logger(__name__)
20
21
ENGINE_ITERATION_TIMEOUT_S = int(
    os.environ.get("VLLM_ENGINE_ITERATION_TIMEOUT_S", "60"))
22

Antoni Baum's avatar
Antoni Baum committed
23

24
25
26
27
class AsyncEngineDeadError(RuntimeError):
    pass


28
29
30
def _raise_exception_on_finish(
        task: asyncio.Task, error_callback: Callable[[Exception],
                                                     None]) -> None:
31
32
    msg = ("Task finished unexpectedly. This should never happen! "
           "Please open an issue on Github.")
33
34

    exception = None
35
    try:
36
37
        task.result()
        # NOTE: This will be thrown if task exits normally (which it should not)
38
        raise AsyncEngineDeadError(msg)
39
40
41
42
43
44
    except Exception as e:
        exception = e
        logger.error("Engine background task failed", exc_info=e)
        error_callback(exception)
        raise AsyncEngineDeadError(
            msg + " See stack trace above for the actual cause.") from e
45
46


Antoni Baum's avatar
Antoni Baum committed
47
48
49
50
51
52
53
54
55
class AsyncStream:
    """A stream of RequestOutputs for a request that can be
    iterated over asynchronously."""

    def __init__(self, request_id: str) -> None:
        self.request_id = request_id
        self._queue = asyncio.Queue()
        self._finished = False

56
    def put(self, item: Union[RequestOutput, Exception]) -> None:
Antoni Baum's avatar
Antoni Baum committed
57
58
59
60
61
        if self._finished:
            return
        self._queue.put_nowait(item)

    def finish(self) -> None:
62
        self._queue.put_nowait(StopAsyncIteration())
Antoni Baum's avatar
Antoni Baum committed
63
64
65
66
67
68
69
70
71
72
73
        self._finished = True

    @property
    def finished(self) -> bool:
        return self._finished

    def __aiter__(self):
        return self

    async def __anext__(self) -> RequestOutput:
        result = await self._queue.get()
74
        if isinstance(result, Exception):
75
            raise result
Antoni Baum's avatar
Antoni Baum committed
76
77
78
        return result


79
80
81
82
83
84
85
86
class RequestTracker:
    """Synchronous abstraction for tracking requests."""

    def __init__(self) -> None:
        self._request_streams: Dict[str, AsyncStream] = {}
        self._finished_requests: asyncio.Queue[str] = asyncio.Queue()
        self._new_requests: asyncio.Queue[Tuple[AsyncStream,
                                                dict]] = asyncio.Queue()
87
        self.new_requests_event = asyncio.Event()
88
89
90
91

    def __contains__(self, item):
        return item in self._request_streams

92
93
    def __len__(self) -> int:
        return len(self._request_streams)
94
95
96
97
98
99
100
101

    def propagate_exception(self,
                            exc: Exception,
                            request_id: Optional[str] = None) -> None:
        """Propagate an exception to request streams
        (all if request_id is None)."""
        if request_id is not None:
            self._request_streams[request_id].put(exc)
102
            self.abort_request(request_id)
103
        else:
104
            for rid, stream in self._request_streams.items():
105
                stream.put(exc)
106
                self.abort_request(rid)
107
108
109
110
111
112
113
114
115
116
117
118
119
120

    def process_request_output(self,
                               request_output: RequestOutput,
                               *,
                               verbose: bool = False) -> None:
        """Process a request output from the engine."""
        request_id = request_output.request_id

        self._request_streams[request_id].put(request_output)
        if request_output.finished:
            if verbose:
                logger.info(f"Finished request {request_id}.")
            self.abort_request(request_id)

121
122
123
124
125
126
127
128
129
130
131
    def process_exception(self,
                          request_id: str,
                          exception: Exception,
                          *,
                          verbose: bool = False) -> None:
        """Propagate an exception from the engine."""
        self._request_streams[request_id].put(exception)
        if verbose:
            logger.info(f"Finished request {request_id}.")
        self.abort_request(request_id)

132
133
134
135
136
137
138
139
140
141
142
143
    def add_request(self, request_id: str,
                    **engine_add_request_kwargs) -> AsyncStream:
        """Add a request to be sent to the engine on the next background
        loop iteration."""
        if request_id in self._request_streams:
            raise KeyError(f"Request {request_id} already exists.")

        stream = AsyncStream(request_id)
        self._new_requests.put_nowait((stream, {
            "request_id": request_id,
            **engine_add_request_kwargs
        }))
144
145
146

        self.new_requests_event.set()

147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
        return stream

    def abort_request(self, request_id: str, *, verbose: bool = False) -> None:
        """Abort a request during next background loop iteration."""
        if verbose:
            logger.info(f"Aborted request {request_id}.")

        self._finished_requests.put_nowait(request_id)

        if request_id not in self._request_streams or self._request_streams[
                request_id].finished:
            # The request has already finished or been aborted.
            return

        self._request_streams[request_id].finish()

163
    def get_new_and_finished_requests(self) -> Tuple[List[Dict], Set[str]]:
164
165
        """Get the new requests and finished requests to be
        sent to the engine."""
166
        new_requests: List[Dict] = []
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
        finished_requests: Set[str] = set()

        while not self._finished_requests.empty():
            request_id = self._finished_requests.get_nowait()
            finished_requests.add(request_id)
            self._request_streams.pop(request_id, None)

        while not self._new_requests.empty():
            stream, new_request = self._new_requests.get_nowait()
            if stream.request_id in finished_requests:
                # The request has already been aborted.
                stream.finish()
                continue
            self._request_streams[stream.request_id] = stream
            new_requests.append(new_request)

        return new_requests, finished_requests
Antoni Baum's avatar
Antoni Baum committed
184

185
    async def wait_for_new_requests(self):
186
187
188
189
190
191
        if not self.has_new_requests():
            await self.new_requests_event.wait()
        self.new_requests_event.clear()

    def has_new_requests(self):
        return not self._new_requests.empty()
192

Antoni Baum's avatar
Antoni Baum committed
193
194
195
196
197
198
199
200
201
202
203
204
205
206

class _AsyncLLMEngine(LLMEngine):
    """Extension of LLMEngine to add async methods."""

    async def step_async(self) -> List[RequestOutput]:
        """Performs one decoding iteration and returns newly generated results.
        The workers are ran asynchronously if possible.

        This function performs one decoding iteration of the engine. It first
        schedules the sequences to be executed in the next iteration and the
        token blocks to be swapped in/out/copy. Then, it executes the model
        and updates the scheduler with the model outputs. Finally, it decodes
        the sequences and returns the newly generated results.
        """
207
        seq_group_metadata_list, scheduler_outputs = self.scheduler.schedule()
Antoni Baum's avatar
Antoni Baum committed
208

209
210
        if not scheduler_outputs.is_empty():
            # Execute the model.
211
212
213
214
            output = await self.model_executor.execute_model_async(
                seq_group_metadata_list, scheduler_outputs.blocks_to_swap_in,
                scheduler_outputs.blocks_to_swap_out,
                scheduler_outputs.blocks_to_copy)
215
216
        else:
            output = []
Antoni Baum's avatar
Antoni Baum committed
217

218
        return self._process_model_outputs(output, scheduler_outputs)
Antoni Baum's avatar
Antoni Baum committed
219

220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
    async def encode_request_async(
        self,
        request_id: str,  # pylint: disable=unused-argument
        prompt: Optional[str],
        prompt_token_ids: Optional[List[int]] = None,
        lora_request: Optional[LoRARequest] = None,
    ):
        if prompt_token_ids is None:
            assert prompt is not None
            prompt_token_ids = await self.tokenizer.encode_async(
                request_id=request_id,
                prompt=prompt,
                lora_request=lora_request)
        return prompt_token_ids

    async def add_request_async(
        self,
        request_id: str,
        prompt: Optional[str],
        sampling_params: SamplingParams,
        prompt_token_ids: Optional[List[int]] = None,
        arrival_time: Optional[float] = None,
        lora_request: Optional[LoRARequest] = None,
    ) -> None:
        if lora_request is not None and not self.lora_config:
            raise ValueError(f"Got lora_request {lora_request} but LoRA is "
                             "not enabled!")
        if arrival_time is None:
            arrival_time = time.time()
        prompt_token_ids = await self.encode_request_async(
            request_id=request_id,
            prompt=prompt,
            prompt_token_ids=prompt_token_ids,
            lora_request=lora_request)

        return self.add_request(
            request_id,
            prompt=prompt,
            prompt_token_ids=prompt_token_ids,
            sampling_params=sampling_params,
            arrival_time=arrival_time,
            lora_request=lora_request,
        )

264
265
    async def check_health_async(self) -> None:
        self.model_executor.check_health()
266

267

268
269
class AsyncLLMEngine:
    """An asynchronous wrapper for LLMEngine.
270

271
    This class is used to wrap the LLMEngine class to make it asynchronous. It
272
    uses asyncio to create a background loop that keeps processing incoming
273
    requests. The LLMEngine is kicked by the generate method when there
274
    are requests in the waiting queue. The generate method yields the outputs
275
    from the LLMEngine to the caller.
276

277
    NOTE: For the comprehensive list of arguments, see `LLMEngine`.
278
279
280
281
282

    Args:
        worker_use_ray: Whether to use Ray for model workers. Required for
            distributed execution. Should be the same as
            `parallel_config.worker_use_ray`.
Zhuohan Li's avatar
Zhuohan Li committed
283
        engine_use_ray: Whether to make LLMEngine a Ray actor. If so, the
284
285
            async frontend will be executed in a separate process as the
            model workers.
286
        log_requests: Whether to log the requests.
zspo's avatar
zspo committed
287
288
        max_log_len: Maximum number of prompt characters or prompt ID numbers
            being printed in log.
289
290
        start_engine_loop: If True, the background task to run the engine
            will be automatically started in the generate call.
291
292
        *args: Arguments for LLMEngine.
        *kwargs: Arguments for LLMEngine.
293
    """
294

Antoni Baum's avatar
Antoni Baum committed
295
296
    _engine_class: Type[_AsyncLLMEngine] = _AsyncLLMEngine

297
298
299
300
301
    def __init__(self,
                 worker_use_ray: bool,
                 engine_use_ray: bool,
                 *args,
                 log_requests: bool = True,
302
                 max_log_len: Optional[int] = None,
303
                 start_engine_loop: bool = True,
304
                 **kwargs) -> None:
305
        self.worker_use_ray = worker_use_ray
Zhuohan Li's avatar
Zhuohan Li committed
306
        self.engine_use_ray = engine_use_ray
307
        self.log_requests = log_requests
308
        self.max_log_len = max_log_len
Antoni Baum's avatar
Antoni Baum committed
309
310
311
        self.engine = self._init_engine(*args, **kwargs)

        self.background_loop = None
312
313
314
315
        # We need to keep a reference to unshielded
        # task as well to prevent it from being garbage
        # collected
        self._background_loop_unshielded = None
316
        self.start_engine_loop = start_engine_loop
317
318
        self._request_tracker: Optional[RequestTracker] = None
        self._errored_with: Optional[BaseException] = None
Antoni Baum's avatar
Antoni Baum committed
319

320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
    @classmethod
    def from_engine_args(cls,
                         engine_args: AsyncEngineArgs,
                         start_engine_loop: bool = True) -> "AsyncLLMEngine":
        """Creates an async LLM engine from the engine arguments."""
        # Create the engine configs.
        engine_configs = engine_args.create_engine_configs()
        parallel_config = engine_configs[2]
        if parallel_config.worker_use_ray or engine_args.engine_use_ray:
            initialize_ray_cluster(parallel_config)
            from vllm.executor.ray_gpu_executor import RayGPUExecutorAsync
            executor_class = RayGPUExecutorAsync
        else:
            assert parallel_config.world_size == 1, (
                "Ray is required if parallel_config.world_size > 1.")
            from vllm.executor.gpu_executor import GPUExecutorAsync
            executor_class = GPUExecutorAsync
        # Create the async LLM engine.
        engine = cls(parallel_config.worker_use_ray,
                     engine_args.engine_use_ray,
                     *engine_configs,
                     executor_class,
                     log_requests=not engine_args.disable_log_requests,
                     log_stats=not engine_args.disable_log_stats,
                     max_log_len=engine_args.max_log_len,
                     start_engine_loop=start_engine_loop)
        return engine

348
349
    @property
    def is_running(self) -> bool:
350
        return (self.background_loop is not None
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
                and not self._background_loop_unshielded.done())

    @property
    def is_stopped(self) -> bool:
        return self.errored or (self.background_loop is not None
                                and self._background_loop_unshielded.done())

    @property
    def errored(self) -> bool:
        return self._errored_with is not None

    def set_errored(self, exc: Exception) -> None:
        self._errored_with = exc

    def _error_callback(self, exc: Exception) -> None:
        self.set_errored(exc)
        self._request_tracker.propagate_exception(exc)
368

369
370
371
372
373
    async def get_tokenizer(self) -> "PreTrainedTokenizer":
        if self.engine_use_ray:
            return await self.engine.get_tokenizer.remote()
        else:
            return self.engine.get_tokenizer()
374

375
    def start_background_loop(self) -> None:
Antoni Baum's avatar
Antoni Baum committed
376
        """Start the background loop."""
377
378
379
        if self.errored:
            raise AsyncEngineDeadError(
                "Background loop has errored already.") from self._errored_with
380
        if self.is_running:
Antoni Baum's avatar
Antoni Baum committed
381
            raise RuntimeError("Background loop is already running.")
382
383
        # Initialize the RequestTracker here so it uses the right event loop.
        self._request_tracker = RequestTracker()
384
385
386
387

        self._background_loop_unshielded = asyncio.get_event_loop(
        ).create_task(self.run_engine_loop())
        self._background_loop_unshielded.add_done_callback(
388
            partial(_raise_exception_on_finish,
389
                    error_callback=self._error_callback))
390
        self.background_loop = asyncio.shield(self._background_loop_unshielded)
Antoni Baum's avatar
Antoni Baum committed
391
392
393

    def _init_engine(self, *args,
                     **kwargs) -> Union[_AsyncLLMEngine, "ray.ObjectRef"]:
Zhuohan Li's avatar
Zhuohan Li committed
394
        if not self.engine_use_ray:
Antoni Baum's avatar
Antoni Baum committed
395
            engine_class = self._engine_class
396
        elif self.worker_use_ray:
Antoni Baum's avatar
Antoni Baum committed
397
            engine_class = ray.remote(num_cpus=0)(self._engine_class).remote
398
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
399
400
401
402
403
404
405
406
407
408
            # FIXME(woosuk): This is a bit hacky. Be careful when changing the
            # order of the arguments.
            cache_config = args[1]
            parallel_config = args[2]
            if parallel_config.tensor_parallel_size == 1:
                num_gpus = cache_config.gpu_memory_utilization
            else:
                num_gpus = 1
            engine_class = ray.remote(num_gpus=num_gpus)(
                self._engine_class).remote
Antoni Baum's avatar
Antoni Baum committed
409
410
        return engine_class(*args, **kwargs)

411
412
413
414
    async def engine_step(self) -> bool:
        """Kick the engine to process the waiting requests.

        Returns True if there are in-progress requests."""
415
416

        new_requests, finished_requests = (
417
            self._request_tracker.get_new_and_finished_requests())
418
419
420
421

        for new_request in new_requests:
            # Add the request into the vLLM engine's waiting queue.
            # TODO: Maybe add add_request_batch to reduce Ray overhead
422
423
424
425
426
427
428
429
430
431
432
433
            try:
                if self.engine_use_ray:
                    await self.engine.add_request.remote(**new_request)
                else:
                    await self.engine.add_request_async(**new_request)
            except ValueError as e:
                # TODO: use a vLLM specific error for failed validation
                self._request_tracker.process_exception(
                    new_request["request_id"],
                    e,
                    verbose=self.log_requests,
                )
434
435
436
437

        if finished_requests:
            await self._engine_abort(finished_requests)

Zhuohan Li's avatar
Zhuohan Li committed
438
439
        if self.engine_use_ray:
            request_outputs = await self.engine.step.remote()
440
        else:
Antoni Baum's avatar
Antoni Baum committed
441
            request_outputs = await self.engine.step_async()
442

Antoni Baum's avatar
Antoni Baum committed
443
        # Put the outputs into the corresponding streams.
444
        for request_output in request_outputs:
445
            self._request_tracker.process_request_output(
446
                request_output, verbose=self.log_requests)
Antoni Baum's avatar
Antoni Baum committed
447

448
449
        return len(request_outputs) > 0

Antoni Baum's avatar
Antoni Baum committed
450
451
452
453
454
455
456
    async def _engine_abort(self, request_ids: Iterable[str]):
        if self.engine_use_ray:
            await self.engine.abort_request.remote(request_ids)
        else:
            self.engine.abort_request(request_ids)

    async def run_engine_loop(self):
457
        has_requests_in_progress = False
Antoni Baum's avatar
Antoni Baum committed
458
        while True:
459
            if not has_requests_in_progress:
460
                logger.debug("Waiting for new requests...")
461
                await self._request_tracker.wait_for_new_requests()
462
463
464
465
466
467
468
469
470
471
472
473
                logger.debug("Got new requests!")

            # Abort if iteration takes too long due to unrecoverable errors
            # (eg. NCCL timeouts).
            try:
                has_requests_in_progress = await asyncio.wait_for(
                    self.engine_step(), ENGINE_ITERATION_TIMEOUT_S)
            except asyncio.TimeoutError as exc:
                logger.error(
                    "Engine iteration timed out. This should never happen!")
                self.set_errored(exc)
                raise
Antoni Baum's avatar
Antoni Baum committed
474
475
476
477
478
479
480
481
482
            await asyncio.sleep(0)

    async def add_request(
        self,
        request_id: str,
        prompt: Optional[str],
        sampling_params: SamplingParams,
        prompt_token_ids: Optional[List[int]] = None,
        arrival_time: Optional[float] = None,
483
        lora_request: Optional[LoRARequest] = None,
Antoni Baum's avatar
Antoni Baum committed
484
485
    ) -> AsyncStream:
        if self.log_requests:
486
487
488
489
490
491
492
493
            shortened_prompt = prompt
            shortened_token_ids = prompt_token_ids
            if self.max_log_len is not None:
                if shortened_prompt is not None:
                    shortened_prompt = shortened_prompt[:self.max_log_len]
                if shortened_token_ids is not None:
                    shortened_token_ids = shortened_token_ids[:self.
                                                              max_log_len]
Antoni Baum's avatar
Antoni Baum committed
494
            logger.info(f"Received request {request_id}: "
495
                        f"prompt: {shortened_prompt!r}, "
zspo's avatar
zspo committed
496
497
                        f"sampling_params: {sampling_params}, "
                        f"prompt_token_ids: {shortened_token_ids}, "
498
                        f"lora_request: {lora_request}.")
Antoni Baum's avatar
Antoni Baum committed
499

500
        if not self.is_running:
501
502
503
504
505
506
507
508
            if self.start_engine_loop:
                self.start_background_loop()
            else:
                raise AsyncEngineDeadError(
                    "Background loop is not running. If it was running, "
                    "inspect the output to find the stacktrace of the "
                    "error that caused the background loop to stop "
                    "(AsyncEngineDeadError).")
Antoni Baum's avatar
Antoni Baum committed
509

510
511
        if arrival_time is None:
            arrival_time = time.time()
512
513
514
515
516
517
518
519
520
521
522
523
524

        if self.engine_use_ray:
            prompt_token_ids = await self.engine.encode_request_async.remote(
                request_id=request_id,
                prompt=prompt,
                prompt_token_ids=prompt_token_ids,
                lora_request=lora_request)
        else:
            prompt_token_ids = await self.engine.encode_request_async(
                request_id=request_id,
                prompt=prompt,
                prompt_token_ids=prompt_token_ids,
                lora_request=lora_request)
525

526
        stream = self._request_tracker.add_request(
527
528
529
530
            request_id,
            prompt=prompt,
            sampling_params=sampling_params,
            prompt_token_ids=prompt_token_ids,
531
            arrival_time=arrival_time,
532
            lora_request=lora_request)
Antoni Baum's avatar
Antoni Baum committed
533
534

        return stream
535

536
    async def generate(
537
538
539
540
        self,
        prompt: Optional[str],
        sampling_params: SamplingParams,
        request_id: str,
541
        prompt_token_ids: Optional[List[int]] = None,
542
        lora_request: Optional[LoRARequest] = None,
543
    ) -> AsyncIterator[RequestOutput]:
544
545
546
        """Generate outputs for a request.

        Generate outputs for a request. This method is a coroutine. It adds the
547
548
        request into the waiting queue of the LLMEngine and streams the outputs
        from the LLMEngine to the caller.
549
550
551
552
553
554
555
556

        Args:
            prompt: The prompt string. Can be None if prompt_token_ids is
                provided.
            sampling_params: The sampling parameters of the request.
            request_id: The unique id of the request.
            prompt_token_ids: The token IDs of the prompt. If None, we
                use the tokenizer to convert the prompts to token IDs.
557
            lora_request: LoRA request to use for generation, if any.
558
559

        Yields:
560
            The output `RequestOutput` objects from the LLMEngine for the
561
            request.
562
563
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
593
594
595
596
597
598
599
600
601
602
603
604

        Details:
            - If the engine is not running, start the background loop,
              which iteratively invokes
              :meth:`~vllm.engine.async_llm_engine.AsyncLLMEngine.engine_step`
              to process the waiting requests.
            - Add the request to the engine's `RequestTracker`.
              On the next background loop, this request will be sent to
              the underlying engine.
              Also, a corresponding `AsyncStream` will be created.
            - Wait for the request outputs from `AsyncStream` and yield them.

        Example:
            >>> # Please refer to entrypoints/api_server.py for
            >>> # the complete example.
            >>>
            >>> # initialize the engine and the example input
            >>> engine = AsyncLLMEngine.from_engine_args(engine_args)
            >>> example_input = {
            >>>     "prompt": "What is LLM?",
            >>>     "stream": False, # assume the non-streaming case
            >>>     "temperature": 0.0,
            >>>     "request_id": 0,
            >>> }
            >>>
            >>> # start the generation
            >>> results_generator = engine.generate(
            >>>    example_input["prompt"],
            >>>    SamplingParams(temperature=example_input["temperature"]),
            >>>    example_input["request_id"])
            >>>
            >>> # get the results
            >>> final_output = None
            >>> async for request_output in results_generator:
            >>>     if await request.is_disconnected():
            >>>         # Abort the request if the client disconnects.
            >>>         await engine.abort(request_id)
            >>>         # Return or raise an error
            >>>         ...
            >>>     final_output = request_output
            >>>
            >>> # Process and return the final output
            >>> ...
605
        """
606
        # Preprocess the request.
607
        arrival_time = time.time()
608

Antoni Baum's avatar
Antoni Baum committed
609
        try:
610
611
612
613
614
615
616
617
            stream = await self.add_request(
                request_id,
                prompt,
                sampling_params,
                prompt_token_ids=prompt_token_ids,
                arrival_time=arrival_time,
                lora_request=lora_request,
            )
618

Antoni Baum's avatar
Antoni Baum committed
619
620
            async for request_output in stream:
                yield request_output
621
622
623
        except (Exception, asyncio.CancelledError) as e:
            # If there is an exception or coroutine is cancelled, abort the
            # request.
Antoni Baum's avatar
Antoni Baum committed
624
625
            self._abort(request_id)
            raise e
626

Antoni Baum's avatar
Antoni Baum committed
627
628
    async def abort(self, request_id: str) -> None:
        """Abort a request.
629

Antoni Baum's avatar
Antoni Baum committed
630
631
        Abort a submitted request. If the request is finished or not found,
        this method will be a no-op.
632

Antoni Baum's avatar
Antoni Baum committed
633
634
635
        Args:
            request_id: The unique id of the request.
        """
636
637
638
639
640
641
642
        if not self.is_running:
            raise AsyncEngineDeadError(
                "Background loop is not running. If it was running, "
                "inspect the output to find the stacktrace of the "
                "error that caused the background loop to stop "
                "(AsyncEngineDeadError).")

Antoni Baum's avatar
Antoni Baum committed
643
        return self._abort(request_id)
644

Antoni Baum's avatar
Antoni Baum committed
645
    def _abort(self, request_id: str) -> None:
646
647
648
649
650
651
652
653
        """Abort a request.

        Abort a submitted request. If the request is finished or not found,
        this method will be a no-op.

        Args:
            request_id: The unique id of the request.
        """
654
655
        self._request_tracker.abort_request(request_id,
                                            verbose=self.log_requests)
656

657
658
659
660
661
662
663
    async def get_model_config(self) -> ModelConfig:
        """Get the model configuration of the vLLM engine."""
        if self.engine_use_ray:
            return await self.engine.get_model_config.remote()
        else:
            return self.engine.get_model_config()

664
665
666
667
668
    async def do_log_stats(self) -> None:
        if self.engine_use_ray:
            await self.engine.do_log_stats.remote()
        else:
            self.engine.do_log_stats()
669

670
    async def check_health(self) -> None:
671
672
673
674
675
676
677
678
679
680
681
682
683
684
        """Raises an error if engine is unhealthy."""
        t = time.perf_counter()
        logger.debug("Starting health check...")
        if self.is_stopped:
            raise AsyncEngineDeadError("Background loop is stopped.")

        if self.engine_use_ray:
            try:
                await self.engine.check_health.remote()
            except ray.exceptions.RayActorError as e:
                raise RuntimeError("Engine is dead.") from e
        else:
            await self.engine.check_health_async()
        logger.debug(f"Health check took {time.perf_counter()-t}s")