llm_engine.py 45.7 KB
Newer Older
Fang li's avatar
Fang li committed
1
import copy
2
from collections import defaultdict
3
import os
Antoni Baum's avatar
Antoni Baum committed
4
import time
5
import pickle
6
import importlib
7
8
from typing import (TYPE_CHECKING, Any, Dict, Iterable, List, Optional, Tuple,
                    Union)
9

10
from vllm.lora.request import LoRARequest
11
12
from vllm.config import (CacheConfig, DeviceConfig, ModelConfig,
                         ParallelConfig, SchedulerConfig, LoRAConfig)
Antoni Baum's avatar
Antoni Baum committed
13
from vllm.core.scheduler import Scheduler, SchedulerOutputs
Woosuk Kwon's avatar
Woosuk Kwon committed
14
from vllm.engine.arg_utils import EngineArgs
15
from vllm.engine.metrics import StatLogger, Stats
16
from vllm.engine.ray_utils import RayWorkerVllm, initialize_cluster, ray
Woosuk Kwon's avatar
Woosuk Kwon committed
17
18
19
from vllm.logger import init_logger
from vllm.outputs import RequestOutput
from vllm.sampling_params import SamplingParams
20
from vllm.sequence import (SamplerOutput, Sequence, SequenceGroup,
21
                           SequenceGroupOutput, SequenceOutput, SequenceStatus)
22
from vllm.transformers_utils.tokenizer import (detokenize_incrementally,
23
                                               TokenizerGroup)
24
25
from vllm.utils import (Counter, set_cuda_visible_devices, get_ip,
                        get_open_port, get_distributed_init_method)
26
27
28
29
30
31

if ray:
    from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy

if TYPE_CHECKING:
    from ray.util.placement_group import PlacementGroup
32
33

logger = init_logger(__name__)
34
_LOCAL_LOGGING_INTERVAL_SEC = 5
Woosuk Kwon's avatar
Woosuk Kwon committed
35

36
37
38
39
40
41
# A map between the device type (in device config) to its worker module.
DEVICE_TO_WORKER_MODULE_MAP = {
    "cuda": "vllm.worker.worker",
    "neuron": "vllm.worker.neuron_worker",
}

42
43
44
45
46
# If the env var is set, it uses the Ray's compiled DAG API
# which optimizes the control plane overhead.
# Run VLLM with VLLM_USE_RAY_COMPILED_DAG=1 to enable it.
USE_RAY_COMPILED_DAG = bool(os.getenv("VLLM_USE_RAY_COMPILED_DAG", 0))

47

48
class LLMEngine:
Zhuohan Li's avatar
Zhuohan Li committed
49
    """An LLM engine that receives requests and generates texts.
50

Woosuk Kwon's avatar
Woosuk Kwon committed
51
    This is the main class for the vLLM engine. It receives requests
52
53
54
55
56
57
58
    from clients and generates texts from the LLM. It includes a tokenizer, a
    language model (possibly distributed across multiple GPUs), and GPU memory
    space allocated for intermediate states (aka KV cache). This class utilizes
    iteration-level scheduling and efficient memory management to maximize the
    serving throughput.

    The `LLM` class wraps this class for offline batched inference and the
59
    `AsyncLLMEngine` class wraps this class for online serving.
60

Zhuohan Li's avatar
Zhuohan Li committed
61
62
    NOTE: The config arguments are derived from the `EngineArgs` class. For the
    comprehensive list of arguments, see `EngineArgs`.
63
64
65
66
67
68
69

    Args:
        model_config: The configuration related to the LLM model.
        cache_config: The configuration related to the KV cache memory
            management.
        parallel_config: The configuration related to distributed execution.
        scheduler_config: The configuration related to the request scheduler.
70
        device_config: The configuration related to the device.
Wang Ran (汪然)'s avatar
Wang Ran (汪然) committed
71
72
        placement_group: Ray placement group for distributed execution.
            Required for distributed execution.
73
74
        log_stats: Whether to log statistics.
    """
75
76
77
78
79
80
81

    def __init__(
        self,
        model_config: ModelConfig,
        cache_config: CacheConfig,
        parallel_config: ParallelConfig,
        scheduler_config: SchedulerConfig,
82
        device_config: DeviceConfig,
83
        lora_config: Optional[LoRAConfig],
84
        placement_group: Optional["PlacementGroup"],
85
        log_stats: bool,
86
87
    ) -> None:
        logger.info(
Zhuohan Li's avatar
Zhuohan Li committed
88
            "Initializing an LLM engine with config: "
89
            f"model={model_config.model!r}, "
90
            f"tokenizer={model_config.tokenizer!r}, "
91
            f"tokenizer_mode={model_config.tokenizer_mode}, "
Jasmond L's avatar
Jasmond L committed
92
            f"revision={model_config.revision}, "
93
            f"tokenizer_revision={model_config.tokenizer_revision}, "
94
            f"trust_remote_code={model_config.trust_remote_code}, "
95
            f"dtype={model_config.dtype}, "
96
            f"max_seq_len={model_config.max_model_len}, "
97
            f"download_dir={model_config.download_dir!r}, "
98
            f"load_format={model_config.load_format}, "
99
            f"tensor_parallel_size={parallel_config.tensor_parallel_size}, "
100
            f"disable_custom_all_reduce={parallel_config.disable_custom_all_reduce}, "
101
            f"quantization={model_config.quantization}, "
102
            f"enforce_eager={model_config.enforce_eager}, "
103
            f"kv_cache_dtype={cache_config.cache_dtype}, "
104
            f"device_config={device_config.device}, "
105
            f"seed={model_config.seed})")
106
107
108
109
        # TODO(woosuk): Print more configs in debug mode.

        self.model_config = model_config
        self.cache_config = cache_config
110
        self.lora_config = lora_config
111
112
        self.parallel_config = parallel_config
        self.scheduler_config = scheduler_config
113
        self.device_config = device_config
114
115
116
        self.log_stats = log_stats
        self._verify_args()

117
        self._init_tokenizer()
118
119
120
        self.seq_counter = Counter()

        # Create the parallel GPU workers.
121
        if self.parallel_config.worker_use_ray:
122
123
124
125
            # Disable Ray usage stats collection.
            ray_usage = os.environ.get("RAY_USAGE_STATS_ENABLED", "0")
            if ray_usage != "1":
                os.environ["RAY_USAGE_STATS_ENABLED"] = "0"
126
127
            self._init_workers_ray(placement_group)
        else:
128
            self._init_workers()
129

130
131
132
133
        # Profile the memory usage and initialize the cache.
        self._init_cache()

        # Create the scheduler.
134
        self.scheduler = Scheduler(scheduler_config, cache_config, lora_config)
Woosuk Kwon's avatar
Woosuk Kwon committed
135

136
137
138
        # Metric Logging.
        if self.log_stats:
            self.stat_logger = StatLogger(
139
140
                local_interval=_LOCAL_LOGGING_INTERVAL_SEC,
                labels=dict(model_name=model_config.model))
141

142
143
144
145
        self.forward_dag = None
        if USE_RAY_COMPILED_DAG:
            self.forward_dag = self._compiled_ray_dag()

146
147
148
    def get_tokenizer_for_seq(self, sequence: Sequence):
        return self.tokenizer.get_lora_tokenizer(sequence.lora_request)

149
150
151
152
153
154
155
    def _dispatch_worker(self):
        worker_module = DEVICE_TO_WORKER_MODULE_MAP[
            self.device_config.device_type]
        imported_worker = importlib.import_module(worker_module)
        Worker = imported_worker.Worker
        return Worker

156
    def _init_workers(self):
157
158
        # Lazy import the Worker to avoid importing torch.cuda/xformers
        # before CUDA_VISIBLE_DEVICES is set in the Worker
159
        Worker = self._dispatch_worker()
160
161
162
163
164

        assert self.parallel_config.world_size == 1, (
            "Ray is required if parallel_config.world_size > 1.")

        self.workers: List[Worker] = []
165
166
        distributed_init_method = get_distributed_init_method(
            get_ip(), get_open_port())
167
        self.driver_worker = Worker(
168
169
170
            self.model_config,
            self.parallel_config,
            self.scheduler_config,
171
            self.device_config,
172
173
174
            local_rank=0,
            rank=0,
            distributed_init_method=distributed_init_method,
175
            lora_config=self.lora_config,
176
            kv_cache_dtype=self.cache_config.cache_dtype,
177
            is_driver_worker=True,
178
        )
179
180
        self._run_workers("init_model")
        self._run_workers("load_model")
181

182
183
184
185
186
187
188
189
190
191
192
193
    def _init_tokenizer(self, **tokenizer_init_kwargs):
        init_kwargs = dict(
            enable_lora=bool(self.lora_config),
            max_num_seqs=self.scheduler_config.max_num_seqs,
            max_input_length=None,
            tokenizer_mode=self.model_config.tokenizer_mode,
            trust_remote_code=self.model_config.trust_remote_code,
            revision=self.model_config.tokenizer_revision)
        init_kwargs.update(tokenizer_init_kwargs)
        self.tokenizer: TokenizerGroup = TokenizerGroup(
            self.model_config.tokenizer, **init_kwargs)

Antoni Baum's avatar
Antoni Baum committed
194
195
    def _init_workers_ray(self, placement_group: "PlacementGroup",
                          **ray_remote_kwargs):
196
197
198
199
        if self.parallel_config.tensor_parallel_size == 1:
            num_gpus = self.cache_config.gpu_memory_utilization
        else:
            num_gpus = 1
200

201
202
203
204
205
        self.driver_dummy_worker: RayWorkerVllm = None
        self.workers: List[RayWorkerVllm] = []

        driver_ip = get_ip()
        for bundle_id, bundle in enumerate(placement_group.bundle_specs):
206
207
            if not bundle.get("GPU", 0):
                continue
208
209
210
211
212
            scheduling_strategy = PlacementGroupSchedulingStrategy(
                placement_group=placement_group,
                placement_group_capture_child_tasks=True,
                placement_group_bundle_index=bundle_id,
            )
213
214
            worker = ray.remote(
                num_cpus=0,
Woosuk Kwon's avatar
Woosuk Kwon committed
215
                num_gpus=num_gpus,
216
                scheduling_strategy=scheduling_strategy,
Antoni Baum's avatar
Antoni Baum committed
217
                **ray_remote_kwargs,
218
            )(RayWorkerVllm).remote(self.model_config.trust_remote_code)
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

            worker_ip = ray.get(worker.get_node_ip.remote())
            if worker_ip == driver_ip and self.driver_dummy_worker is None:
                # If the worker is on the same node as the driver, we use it
                # as the resource holder for the driver process.
                self.driver_dummy_worker = worker
            else:
                self.workers.append(worker)

        if self.driver_dummy_worker is None:
            raise ValueError(
                "Ray does not allocate any GPUs on the driver node. Consider "
                "adjusting the Ray placement group or running the driver on a "
                "GPU node.")

        driver_node_id, driver_gpu_ids = ray.get(
            self.driver_dummy_worker.get_node_and_gpu_ids.remote())
        worker_node_and_gpu_ids = ray.get(
            [worker.get_node_and_gpu_ids.remote() for worker in self.workers])

        node_workers = defaultdict(list)
        node_gpus = defaultdict(list)

        node_workers[driver_node_id].append(0)
        node_gpus[driver_node_id].extend(driver_gpu_ids)
        for i, (node_id, gpu_ids) in enumerate(worker_node_and_gpu_ids,
                                               start=1):
            node_workers[node_id].append(i)
            node_gpus[node_id].extend(gpu_ids)
        for node_id, gpu_ids in node_gpus.items():
            node_gpus[node_id] = sorted(gpu_ids)

        # Set CUDA_VISIBLE_DEVICES for the driver.
        set_cuda_visible_devices(node_gpus[driver_node_id])
        for worker, (node_id, _) in zip(self.workers, worker_node_and_gpu_ids):
            worker.set_cuda_visible_devices.remote(node_gpus[node_id])

256
        distributed_init_method = get_distributed_init_method(
257
            driver_ip, get_open_port())
258
259
260

        # Lazy import the Worker to avoid importing torch.cuda/xformers
        # before CUDA_VISIBLE_DEVICES is set in the Worker
261
        Worker = self._dispatch_worker()
262
263

        # Initialize torch distributed process group for the workers.
Fang li's avatar
Fang li committed
264
265
266
        model_config = copy.deepcopy(self.model_config)
        parallel_config = copy.deepcopy(self.parallel_config)
        scheduler_config = copy.deepcopy(self.scheduler_config)
267
        device_config = copy.deepcopy(self.device_config)
268
269
270
271
272
273
274
275
276
277
278

        for rank, (worker, (node_id,
                            _)) in enumerate(zip(self.workers,
                                                 worker_node_and_gpu_ids),
                                             start=1):
            local_rank = node_workers[node_id].index(rank)
            worker.init_worker.remote(
                lambda rank=rank, local_rank=local_rank: Worker(
                    model_config,
                    parallel_config,
                    scheduler_config,
279
                    device_config,
280
281
282
                    local_rank,
                    rank,
                    distributed_init_method,
283
                    lora_config=self.lora_config,
zhaoyang-star's avatar
zhaoyang-star committed
284
                    kv_cache_dtype=self.cache_config.cache_dtype,
285
286
287
288
289
290
291
292
                ))

        driver_rank = 0
        driver_local_rank = node_workers[driver_node_id].index(driver_rank)
        self.driver_worker = Worker(
            model_config,
            parallel_config,
            scheduler_config,
293
            device_config,
294
295
296
            driver_local_rank,
            driver_rank,
            distributed_init_method,
297
            lora_config=self.lora_config,
zhaoyang-star's avatar
zhaoyang-star committed
298
            kv_cache_dtype=self.cache_config.cache_dtype,
299
            is_driver_worker=True,
300
        )
301

302
303
304
305
        # don't use cupy for eager mode
        self._run_workers("init_model",
                          cupy_port=get_open_port()
                          if not model_config.enforce_eager else None)
306
307
308
309
310
        self._run_workers(
            "load_model",
            max_concurrent_workers=self.parallel_config.
            max_parallel_loading_workers,
        )
311

312
313
    def _verify_args(self) -> None:
        self.model_config.verify_with_parallel_config(self.parallel_config)
314
        self.cache_config.verify_with_parallel_config(self.parallel_config)
315
316
317
318
        if self.lora_config:
            self.lora_config.verify_with_model_config(self.model_config)
            self.lora_config.verify_with_scheduler_config(
                self.scheduler_config)
319
320

    def _init_cache(self) -> None:
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
        """Profiles the memory usage and initializes the KV cache.

        The engine will first conduct a profiling of the existing memory usage.
        Then, it calculate the maximum possible number of GPU and CPU blocks
        that can be allocated with the remaining free memory.
        More details can be found in the
        :meth:`~vllm.worker.worker.Worker.profile_num_available_blocks` method
        from class :class:`~vllm.worker.Worker`.

        Afterwards, as there may be multiple workers,
        we take the minimum number of blocks across all workers
        to ensure this can be applied to all of them.

        Finally, the engine will initialize the KV cache
        with the calculated number of blocks.

        .. tip::
            You may limit the usage of GPU memory
            by adjusting the `gpu_memory_utilization` parameters.
        """
341
342
343
344
345
        # Get the maximum number of blocks that can be allocated on GPU and CPU.
        num_blocks = self._run_workers(
            "profile_num_available_blocks",
            block_size=self.cache_config.block_size,
            gpu_memory_utilization=self.cache_config.gpu_memory_utilization,
346
            cpu_swap_space=self.cache_config.swap_space_bytes,
347
            cache_dtype=self.cache_config.cache_dtype,
348
349
350
351
352
353
354
355
        )

        # Since we use a shared centralized controller, we take the minimum
        # number of blocks across all workers to make sure all the memory
        # operators can be applied to all workers.
        num_gpu_blocks = min(b[0] for b in num_blocks)
        num_cpu_blocks = min(b[1] for b in num_blocks)
        # FIXME(woosuk): Change to debug log.
356
357
        logger.info(f"# GPU blocks: {num_gpu_blocks}, "
                    f"# CPU blocks: {num_cpu_blocks}")
358

359
        if num_gpu_blocks <= 0:
360
361
362
            raise ValueError("No available memory for the cache blocks. "
                             "Try increasing `gpu_memory_utilization` when "
                             "initializing the engine.")
363
364
365
366
367
368
369
370
        max_seq_len = self.cache_config.block_size * num_gpu_blocks
        if self.model_config.max_model_len > max_seq_len:
            raise ValueError(
                f"The model's max seq len ({self.model_config.max_model_len}) "
                "is larger than the maximum number of tokens that can be "
                f"stored in KV cache ({max_seq_len}). Try increasing "
                "`gpu_memory_utilization` or decreasing `max_model_len` when "
                "initializing the engine.")
371

372
373
374
375
376
        self.cache_config.num_gpu_blocks = num_gpu_blocks
        self.cache_config.num_cpu_blocks = num_cpu_blocks

        # Initialize the cache.
        self._run_workers("init_cache_engine", cache_config=self.cache_config)
377
378
379
        # Warm up the model. This includes capturing the model into CUDA graph
        # if enforce_eager is False.
        self._run_workers("warm_up_model")
380

381
    @classmethod
Zhuohan Li's avatar
Zhuohan Li committed
382
383
384
385
386
    def from_engine_args(cls, engine_args: EngineArgs) -> "LLMEngine":
        """Creates an LLM engine from the engine arguments."""
        # Create the engine configs.
        engine_configs = engine_args.create_engine_configs()
        parallel_config = engine_configs[2]
387
        # Initialize the cluster.
388
        placement_group = initialize_cluster(parallel_config)
Zhuohan Li's avatar
Zhuohan Li committed
389
        # Create the LLM engine.
390
        engine = cls(*engine_configs,
391
                     placement_group,
Zhuohan Li's avatar
Zhuohan Li committed
392
393
                     log_stats=not engine_args.disable_log_stats)
        return engine
394

395
396
397
398
399
400
401
402
403
404
405
406
407
408
    def encode_request(
        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 = self.tokenizer.encode(request_id=request_id,
                                                     prompt=prompt,
                                                     lora_request=lora_request)
        return prompt_token_ids

409
410
411
    def add_request(
        self,
        request_id: str,
Woosuk Kwon's avatar
Woosuk Kwon committed
412
        prompt: Optional[str],
413
414
415
        sampling_params: SamplingParams,
        prompt_token_ids: Optional[List[int]] = None,
        arrival_time: Optional[float] = None,
416
        lora_request: Optional[LoRARequest] = None,
417
        prefix_pos: Optional[int] = None,
418
    ) -> None:
Zhuohan Li's avatar
Zhuohan Li committed
419
        """Add a request to the engine's request pool.
420
421

        The request is added to the request pool and will be processed by the
Zhuohan Li's avatar
Zhuohan Li committed
422
        scheduler as `engine.step()` is called. The exact scheduling policy is
423
424
425
426
427
428
429
430
431
432
        determined by the scheduler.

        Args:
            request_id: The unique ID of the request.
            prompt: The prompt string. Can be None if prompt_token_ids is
                provided.
            sampling_params: The sampling parameters for text generation.
            prompt_token_ids: The token IDs of the prompt. If None, we
                use the tokenizer to convert the prompts to token IDs.
            arrival_time: The arrival time of the request. If None, we use
433
                the current monotonic time.
434
435
436
437
438
            prefix_pos: If not None, we use the given position as the prefix
                position for each prompt. We will cache the prefix's KV
                cache and reuse it for the next request with the same prefix.
                This is an experimental feature, and may be replaced with
                automatic prefix caching in the future.
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462

        Details:
            - Set arrival_time to the current time if it is None.
            - Set prompt_token_ids to the encoded prompt if it is None.
            - Create `best_of` number of :class:`~vllm.Sequence` objects.
            - Create a :class:`~vllm.SequenceGroup` object
              from the list of :class:`~vllm.Sequence`.
            - Add the :class:`~vllm.SequenceGroup` object to the scheduler.

        Example:
            >>> # initialize engine
            >>> engine = LLMEngine.from_engine_args(engine_args)
            >>> # set request arguments
            >>> example_prompt = "Who is the president of the United States?"
            >>> sampling_params = SamplingParams(temperature=0.0)
            >>> request_id = 0
            >>>
            >>> # add the request to the engine
            >>> engine.add_request(
            >>>    str(request_id),
            >>>    example_prompt,
            >>>    SamplingParams(temperature=0.0))
            >>> # continue the request processing
            >>> ...
463
        """
464
465
466
        if lora_request is not None and not self.lora_config:
            raise ValueError(f"Got lora_request {lora_request} but LoRA is "
                             "not enabled!")
467
        if arrival_time is None:
468
            arrival_time = time.monotonic()
469
470
471
472
473
        prompt_token_ids = self.encode_request(
            request_id=request_id,
            prompt=prompt,
            prompt_token_ids=prompt_token_ids,
            lora_request=lora_request)
474
475
476

        # Create the sequences.
        block_size = self.cache_config.block_size
477
        seq_id = next(self.seq_counter)
478
479
        seq = Sequence(seq_id, prompt, prompt_token_ids, block_size,
                       lora_request)
480

481
482
        # Check whether the input specifies prefix
        prefix = self.scheduler.prefix_pool.add_or_get_prefix(
483
484
            prompt_token_ids[:prefix_pos], lora_request.lora_int_id
            if lora_request else 0) if prefix_pos is not None else None
485

486
487
488
        # Defensive copy of SamplingParams, which are used by the sampler
        sampling_params = copy.deepcopy(sampling_params)

489
        # Create the sequence group.
490
        seq_group = SequenceGroup(request_id, [seq], sampling_params,
491
                                  arrival_time, lora_request, prefix)
492
493
494
495

        # Add the sequence group to the scheduler.
        self.scheduler.add_seq_group(seq_group)

Antoni Baum's avatar
Antoni Baum committed
496
497
    def abort_request(self, request_id: Union[str, Iterable[str]]) -> None:
        """Aborts a request(s) with the given ID.
498
499

        Args:
Antoni Baum's avatar
Antoni Baum committed
500
            request_id: The ID(s) of the request to abort.
501
502
503
504
505
506
507
508
509
510
511

        Details:
            - Refer to the
              :meth:`~vllm.core.scheduler.Scheduler.abort_seq_group`
              from class :class:`~vllm.core.scheduler.Scheduler`.

        Example:
            >>> # initialize engine and add a request with request_id
            >>> request_id = str(0)
            >>> # abort the request
            >>> engine.abort_request(request_id)
512
        """
513
514
        self.scheduler.abort_seq_group(request_id)

515
516
517
518
    def get_model_config(self) -> ModelConfig:
        """Gets the model configuration."""
        return self.model_config

519
    def get_num_unfinished_requests(self) -> int:
520
        """Gets the number of unfinished requests."""
521
522
        return self.scheduler.get_num_unfinished_seq_groups()

523
    def has_unfinished_requests(self) -> bool:
524
        """Returns True if there are unfinished requests."""
525
526
        return self.scheduler.has_unfinished_seqs()

527
528
529
530
531
532
533
534
535
536
537
538
539
540
    def _check_beam_search_early_stopping(
        self,
        early_stopping: Union[bool, str],
        sampling_params: SamplingParams,
        best_running_seq: Sequence,
        current_worst_seq: Sequence,
    ) -> bool:
        assert sampling_params.use_beam_search
        length_penalty = sampling_params.length_penalty
        if early_stopping is True:
            return True

        current_worst_score = (current_worst_seq.get_beam_search_score(
            length_penalty=length_penalty,
541
542
            eos_token_id=self.get_tokenizer_for_seq(
                current_worst_seq).eos_token_id))
543
544
545
        if early_stopping is False:
            highest_attainable_score = (best_running_seq.get_beam_search_score(
                length_penalty=length_penalty,
546
547
                eos_token_id=self.get_tokenizer_for_seq(
                    best_running_seq).eos_token_id))
548
549
550
551
552
553
554
555
556
557
558
559
560
        else:
            assert early_stopping == "never"
            if length_penalty > 0.0:
                # If length_penalty > 0.0, beam search will prefer longer
                # sequences. The highest attainable score calculation is
                # based on the longest possible sequence length in this case.
                max_possible_length = max(
                    best_running_seq.get_prompt_len() +
                    sampling_params.max_tokens,
                    self.scheduler_config.max_model_len)
                highest_attainable_score = (
                    best_running_seq.get_beam_search_score(
                        length_penalty=length_penalty,
561
562
                        eos_token_id=self.get_tokenizer_for_seq(
                            best_running_seq).eos_token_id,
563
564
565
566
567
568
569
570
                        seq_len=max_possible_length))
            else:
                # Otherwise, beam search will prefer shorter sequences. The
                # highest attainable score calculation is based on the current
                # sequence length.
                highest_attainable_score = (
                    best_running_seq.get_beam_search_score(
                        length_penalty=length_penalty,
571
572
                        eos_token_id=self.get_tokenizer_for_seq(
                            best_running_seq).eos_token_id))
573
574
        return current_worst_score >= highest_attainable_score

575
    def _process_sequence_group_outputs(self, seq_group: SequenceGroup,
Zhuohan Li's avatar
Zhuohan Li committed
576
                                        outputs: SequenceGroupOutput) -> None:
577

578
579
580
581
582
583
584
        # Process prompt logprobs
        prompt_logprobs = outputs.prompt_logprobs
        if prompt_logprobs is not None:
            seq_group.prompt_logprobs = prompt_logprobs

        # Process samples
        samples = outputs.samples
585
586
587
588
589
590
591
592
593
594
595
596
597
        parent_seqs = seq_group.get_seqs(status=SequenceStatus.RUNNING)
        existing_finished_seqs = seq_group.get_finished_seqs()
        parent_child_dict = {
            parent_seq.seq_id: []
            for parent_seq in parent_seqs
        }
        for sample in samples:
            parent_child_dict[sample.parent_seq_id].append(sample)
        # List of (child, parent)
        child_seqs: List[Tuple[Sequence, Sequence]] = []

        # Process the child samples for each parent sequence
        for parent in parent_seqs:
Zhuohan Li's avatar
Zhuohan Li committed
598
            child_samples: List[SequenceOutput] = parent_child_dict[
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
                parent.seq_id]
            if len(child_samples) == 0:
                # This parent sequence has no children samples. Remove
                # the parent sequence from the sequence group since it will
                # not be used in the future iterations.
                parent.status = SequenceStatus.FINISHED_ABORTED
                seq_group.remove(parent.seq_id)
                self.scheduler.free_seq(parent)
                continue
            # Fork the parent sequence if there are multiple child samples.
            for child_sample in child_samples[:-1]:
                new_child_seq_id = next(self.seq_counter)
                child = parent.fork(new_child_seq_id)
                child.append_token_id(child_sample.output_token,
                                      child_sample.logprobs)
                child_seqs.append((child, parent))
            # Continue the parent sequence for the last child sample.
            # We reuse the parent sequence here to reduce redundant memory
            # copies, especially when using non-beam search sampling methods.
            last_child_sample = child_samples[-1]
            parent.append_token_id(last_child_sample.output_token,
                                   last_child_sample.logprobs)
            child_seqs.append((parent, parent))

        for seq, _ in child_seqs:
624
            self._decode_sequence(seq, seq_group.sampling_params)
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
            self._check_stop(seq, seq_group.sampling_params)

        # Non-beam search case
        if not seq_group.sampling_params.use_beam_search:
            # For newly created child sequences, add them to the sequence group
            # and fork them in block manager if they are not finished.
            for seq, parent in child_seqs:
                if seq is not parent:
                    seq_group.add(seq)
                    if not seq.is_finished():
                        self.scheduler.fork_seq(parent, seq)

            # Free the finished and selected parent sequences' memory in block
            # manager. Keep them in the sequence group as candidate output.
            # NOTE: we need to fork the new sequences before freeing the
            # old sequences.
            for seq, parent in child_seqs:
                if seq is parent and seq.is_finished():
                    self.scheduler.free_seq(seq)
            return

        # Beam search case
        # Select the child sequences to keep in the sequence group.
        selected_child_seqs = []
        unselected_child_seqs = []
        beam_width = seq_group.sampling_params.best_of
        length_penalty = seq_group.sampling_params.length_penalty

        # Select the newly finished sequences with the highest scores
        # to replace existing finished sequences.
        # Tuple of (seq, parent, is_new)
        existing_finished_seqs = [(seq, None, False)
                                  for seq in existing_finished_seqs]
        new_finished_seqs = [(seq, parent, True) for seq, parent in child_seqs
                             if seq.is_finished()]
        all_finished_seqs = existing_finished_seqs + new_finished_seqs
        # Sort the finished sequences by their scores.
        all_finished_seqs.sort(key=lambda x: x[0].get_beam_search_score(
            length_penalty=length_penalty,
664
            eos_token_id=self.get_tokenizer_for_seq(x[0]).eos_token_id),
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
                               reverse=True)
        for seq, parent, is_new in all_finished_seqs[:beam_width]:
            if is_new:
                # A newly generated child sequence finishes and has a high
                # score, so we will add it into the sequence group.
                selected_child_seqs.append((seq, parent))
        for seq, parent, is_new in all_finished_seqs[beam_width:]:
            if is_new:
                # A newly generated child sequence finishes but has a low
                # score, so we will not add it into the sequence group.
                # Additionally, if this sequence is a continuation of a
                # parent sequence, we will need remove the parent sequence
                # from the sequence group.
                unselected_child_seqs.append((seq, parent))
            else:
                # An existing finished sequence has a low score, so we will
                # remove it from the sequence group.
                seq_group.remove(seq.seq_id)

        # select the top beam_width sequences from the running
        # sequences for the next iteration to continue the beam
        # search.
        running_child_seqs = [(seq, parent) for seq, parent in child_seqs
                              if not seq.is_finished()]
        # Sort the running sequences by their scores.
        running_child_seqs.sort(key=lambda x: x[0].get_beam_search_score(
            length_penalty=length_penalty,
692
            eos_token_id=self.get_tokenizer_for_seq(x[0]).eos_token_id),
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
                                reverse=True)

        # Check if we can stop the beam search.
        if len(running_child_seqs) == 0:
            # No running sequences, stop the beam search.
            stop_beam_search = True
        elif len(all_finished_seqs) < beam_width:
            # Not enough finished sequences, continue the beam search.
            stop_beam_search = False
        else:
            # Check the early stopping criteria
            best_running_seq = running_child_seqs[0][0]
            current_worst_seq = all_finished_seqs[beam_width - 1][0]
            stop_beam_search = self._check_beam_search_early_stopping(
                seq_group.sampling_params.early_stopping,
                seq_group.sampling_params, best_running_seq, current_worst_seq)

        if stop_beam_search:
            # Stop the beam search and remove all the running sequences from
            # the sequence group.
            unselected_child_seqs.extend(running_child_seqs)
        else:
            # Continue the beam search and select the top beam_width sequences
            # to continue the beam search.
            selected_child_seqs.extend(running_child_seqs[:beam_width])
            # The remaining running sequences will not be used in the next
            # iteration. Again, if these sequences are continuations of
            # parent sequences, we will need to remove the parent sequences
            # from the sequence group.
            unselected_child_seqs.extend(running_child_seqs[beam_width:])

        # For newly created child sequences, add them to the sequence group
        # and fork them in block manager if they are not finished.
        for seq, parent in selected_child_seqs:
            if seq is not parent:
                seq_group.add(seq)
                if not seq.is_finished():
                    self.scheduler.fork_seq(parent, seq)

        # Free the finished and selected parent sequences' memory in block
        # manager. Keep them in the sequence group as candidate output.
        for seq, parent in selected_child_seqs:
            if seq is parent and seq.is_finished():
                self.scheduler.free_seq(seq)

        # Remove the unselected parent sequences from the sequence group and
        # free their memory in block manager.
        for seq, parent in unselected_child_seqs:
            if seq is parent:
                # Remove the parent sequence if it is not selected for next
                # iteration
                seq_group.remove(seq.seq_id)
                self.scheduler.free_seq(seq)

    def _process_model_outputs(
            self, output: SamplerOutput,
Antoni Baum's avatar
Antoni Baum committed
749
            scheduler_outputs: SchedulerOutputs) -> List[RequestOutput]:
750
        now = time.time()
751
752
        # Update the scheduled sequence groups with the model outputs.
        scheduled_seq_groups = scheduler_outputs.scheduled_seq_groups
753
754
        for seq_group, outputs in zip(scheduled_seq_groups, output):
            self._process_sequence_group_outputs(seq_group, outputs)
755
756
757

        # Free the finished sequence groups.
        self.scheduler.free_finished_seq_groups()
758
759
760

        # Create the outputs.
        request_outputs: List[RequestOutput] = []
761
        for seq_group in scheduled_seq_groups:
762
            seq_group.maybe_set_first_token_time(now)
763
764
765
            request_output = RequestOutput.from_seq_group(seq_group)
            request_outputs.append(request_output)
        for seq_group in scheduler_outputs.ignored_seq_groups:
766
            request_output = RequestOutput.from_seq_group(seq_group)
767
            request_outputs.append(request_output)
Woosuk Kwon's avatar
Woosuk Kwon committed
768

769
770
771
772
773
774
        # Update prefix state, now all the uncomputed prefixes are computed.
        for seq_group in scheduled_seq_groups:
            if (seq_group.prefix is not None and seq_group.prefix.allocated
                    and not seq_group.prefix.computed):
                seq_group.prefix.computed = True

775
        # Log stats.
Woosuk Kwon's avatar
Woosuk Kwon committed
776
        if self.log_stats:
777
778
            self.stat_logger.log(self._get_stats(scheduler_outputs))

779
780
        return request_outputs

Antoni Baum's avatar
Antoni Baum committed
781
782
783
    def step(self) -> List[RequestOutput]:
        """Performs one decoding iteration and returns newly generated results.

784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
        .. figure:: https://i.imgur.com/sv2HssD.png
            :alt: Overview of the step function
            :align: center

            Overview of the step function.

        Details:
            - Step 1: Schedules the sequences to be executed in the next
              iteration and the token blocks to be swapped in/out/copy.

                - Depending on the scheduling policy,
                  sequences may be `preempted/reordered`.
                - A Sequence Group (SG) refer to a group of sequences
                  that are generated from the same prompt.

            - Step 2: Calls the workers to execute the model.
            - Step 3: Processes the model output. This mainly includes:

                - Decodes the relevant outputs.
                - Updates the scheduled sequence groups with model outputs
                  based on its `sampling parameters` (`use_beam_search` or not).
                - Frees the finished sequence groups.

            - Finally, it creates and returns the newly generated results.

        Example:
            >>> # Please see the example/ folder for more detailed examples.
            >>>
            >>> # initialize engine and request arguments
            >>> engine = LLMEngine.from_engine_args(engine_args)
            >>> example_inputs = [(0, "What is LLM?",
            >>>    SamplingParams(temperature=0.0))]
            >>>
            >>> # Start the engine with an event loop
            >>> while True:
            >>>     if example_inputs:
            >>>         req_id, prompt, sampling_params = example_inputs.pop(0)
            >>>         engine.add_request(str(req_id), prompt, sampling_params)
            >>>
            >>>     # continue the request processing
            >>>     request_outputs = engine.step()
            >>>     for request_output in request_outputs:
            >>>         if request_output.finished:
            >>>             # return or show the request output
            >>>
            >>>     if not (engine.has_unfinished_requests() or example_inputs):
            >>>         break
Antoni Baum's avatar
Antoni Baum committed
831
        """
832
        seq_group_metadata_list, scheduler_outputs = self.scheduler.schedule()
Antoni Baum's avatar
Antoni Baum committed
833

834
835
836
837
838
839
840
841
842
        if not scheduler_outputs.is_empty():
            # Execute the model.
            all_outputs = self._run_workers(
                "execute_model",
                driver_kwargs={
                    "seq_group_metadata_list": seq_group_metadata_list,
                    "blocks_to_swap_in": scheduler_outputs.blocks_to_swap_in,
                    "blocks_to_swap_out": scheduler_outputs.blocks_to_swap_out,
                    "blocks_to_copy": scheduler_outputs.blocks_to_copy,
843
844
                },
                use_ray_compiled_dag=USE_RAY_COMPILED_DAG)
845
846
847
848
849

            # Only the driver worker returns the sampling results.
            output = all_outputs[0]
        else:
            output = []
Antoni Baum's avatar
Antoni Baum committed
850

851
        return self._process_model_outputs(output, scheduler_outputs)
Antoni Baum's avatar
Antoni Baum committed
852

853
    def do_log_stats(self) -> None:
854
855
856
        """Forced log when no requests active."""
        if self.log_stats:
            self.stat_logger.log(self._get_stats(scheduler_outputs=None))
857

858
859
860
    def _get_stats(self,
                   scheduler_outputs: Optional[SchedulerOutputs]) -> Stats:
        """Get Stats to be Logged to Prometheus."""
861
        now = time.monotonic()
Woosuk Kwon's avatar
Woosuk Kwon committed
862

863
864
865
866
        # KV Cache Usage in %.
        num_total_gpu = self.cache_config.num_gpu_blocks
        num_free_gpu = self.scheduler.block_manager.get_num_free_gpu_blocks()
        gpu_cache_usage = 1.0 - (num_free_gpu / num_total_gpu)
Woosuk Kwon's avatar
Woosuk Kwon committed
867

868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
        num_total_cpu = self.cache_config.num_cpu_blocks
        cpu_cache_usage = 0.
        if num_total_cpu > 0:
            num_free_cpu = self.scheduler.block_manager.get_num_free_cpu_blocks(
            )
            cpu_cache_usage = 1.0 - (num_free_cpu / num_total_cpu)

        # Scheduler State
        num_running = len(self.scheduler.running)
        num_swapped = len(self.scheduler.swapped)
        num_waiting = len(self.scheduler.waiting)

        # Iteration stats if we have scheduler output.
        num_prompt_tokens = 0
        num_generation_tokens = 0
        time_to_first_tokens = []
        time_per_output_tokens = []
        time_e2e_requests = []
        if scheduler_outputs is not None:
            prompt_run = scheduler_outputs.prompt_run

            # Number of Tokens.
            if prompt_run:
891
892
893
                num_prompt_tokens = sum(
                    len(seq_group.prompt_token_ids)
                    for seq_group in scheduler_outputs.scheduled_seq_groups)
894
895
896
                num_generation_tokens = sum(
                    seq_group.num_seqs()
                    for seq_group in scheduler_outputs.scheduled_seq_groups)
897
898
899
900
901
902
            else:
                num_generation_tokens = scheduler_outputs.num_batched_tokens

            # Latency Timings.
            time_last_iters = []
            for seq_group in scheduler_outputs.scheduled_seq_groups:
903
                # Time since last token. (n.b. updates seq_group.metrics.last_token_time)
904
905
906
                time_last_iters.append(seq_group.get_last_latency(now))
                # Time since arrival for all finished requests.
                if seq_group.is_finished():
907
908
                    time_e2e_requests.append(now -
                                             seq_group.metrics.arrival_time)
909
910
911
912
913
914
915
916
917

            time_to_first_tokens = time_last_iters if prompt_run else []
            time_per_output_tokens = [] if prompt_run else time_last_iters

        return Stats(
            now=now,
            num_running=num_running,
            num_swapped=num_swapped,
            num_waiting=num_waiting,
918
919
            gpu_cache_usage=gpu_cache_usage,
            cpu_cache_usage=cpu_cache_usage,
920
921
922
923
924
            num_prompt_tokens=num_prompt_tokens,
            num_generation_tokens=num_generation_tokens,
            time_to_first_tokens=time_to_first_tokens,
            time_per_output_tokens=time_per_output_tokens,
            time_e2e_requests=time_e2e_requests,
925
926
        )

927
    def _decode_sequence(self, seq: Sequence, prms: SamplingParams) -> None:
928
        """Decodes the new token for a sequence."""
929
930
        (new_tokens, new_output_text, prefix_offset,
         read_offset) = detokenize_incrementally(
931
             self.get_tokenizer_for_seq(seq),
932
933
934
935
             all_input_ids=seq.get_token_ids(),
             prev_tokens=seq.tokens,
             prefix_offset=seq.prefix_offset,
             read_offset=seq.read_offset,
936
937
             skip_special_tokens=prms.skip_special_tokens,
             spaces_between_special_tokens=prms.spaces_between_special_tokens,
938
939
940
941
942
943
944
945
         )
        if seq.tokens is None:
            seq.tokens = new_tokens
        else:
            seq.tokens.extend(new_tokens)
        seq.prefix_offset = prefix_offset
        seq.read_offset = read_offset
        seq.output_text += new_output_text
946
947
948

    def _check_stop(self, seq: Sequence,
                    sampling_params: SamplingParams) -> None:
949
        """Stop the finished sequences."""
950
951
        for stop_str in sampling_params.stop:
            if seq.output_text.endswith(stop_str):
952
                self._finalize_sequence(seq, sampling_params, stop_str)
953
954
                seq.status = SequenceStatus.FINISHED_STOPPED
                return
955
        if seq.get_last_token_id() in sampling_params.stop_token_ids:
956
957
958
            stop_str = self.get_tokenizer_for_seq(seq).convert_ids_to_tokens(
                seq.get_last_token_id())
            self._finalize_sequence(seq, sampling_params, stop_str)
959
960
            seq.status = SequenceStatus.FINISHED_STOPPED
            return
961
962
963
964
965
966
967
968
969
970
971
972

        # Check if the sequence has reached max_model_len.
        if seq.get_len() > self.scheduler_config.max_model_len:
            seq.status = SequenceStatus.FINISHED_LENGTH_CAPPED
            return

        # Check if the sequence has reached max_tokens.
        if seq.get_output_len() == sampling_params.max_tokens:
            seq.status = SequenceStatus.FINISHED_LENGTH_CAPPED
            return

        # Check if the sequence has generated the EOS token.
973
974
        if ((not sampling_params.ignore_eos) and seq.get_last_token_id()
                == self.get_tokenizer_for_seq(seq).eos_token_id):
975
976
            seq.status = SequenceStatus.FINISHED_STOPPED
            return
977

978
979
980
981
982
983
984
985
    def _finalize_sequence(self, seq: Sequence,
                           sampling_params: SamplingParams,
                           stop_string: str) -> None:
        if not sampling_params.include_stop_str_in_output and stop_string:
            # Truncate the output text so that the stop string is
            # not included in the output.
            seq.output_text = seq.output_text[:-len(stop_string)]

986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
    def add_lora(self, lora_request: LoRARequest) -> bool:
        assert lora_request.lora_int_id > 0, "lora_id must be greater than 0."
        return self._run_workers(
            "add_lora",
            lora_request=lora_request,
        )

    def remove_lora(self, lora_id: int) -> bool:
        assert lora_id > 0, "lora_id must be greater than 0."
        return self._run_workers(
            "remove_lora",
            lora_id=lora_id,
        )

    def list_loras(self) -> List[int]:
        return self._run_workers("list_loras")

1003
1004
1005
1006
    def _run_workers(
        self,
        method: str,
        *args,
1007
1008
        driver_args: Optional[List[Any]] = None,
        driver_kwargs: Optional[Dict[str, Any]] = None,
1009
        max_concurrent_workers: Optional[int] = None,
1010
        use_ray_compiled_dag: bool = False,
1011
1012
1013
        **kwargs,
    ) -> Any:
        """Runs the given method on all workers."""
1014

1015
        if max_concurrent_workers:
1016
1017
1018
            raise NotImplementedError(
                "max_concurrent_workers is not supported yet.")

1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
        if use_ray_compiled_dag:
            # Right now, compiled DAG can only accept a single
            # input. TODO(sang): Fix it.
            output_channels = self.forward_dag.execute(1)
        else:
            # Start the ray workers first.
            ray_worker_outputs = [
                worker.execute_method.remote(method, *args, **kwargs)
                for worker in self.workers
            ]
1029
1030
1031
1032
1033

        if driver_args is None:
            driver_args = args
        if driver_kwargs is None:
            driver_kwargs = kwargs
1034

1035
1036
1037
        # Start the driver worker after all the ray workers.
        driver_worker_output = getattr(self.driver_worker,
                                       method)(*driver_args, **driver_kwargs)
1038

1039
1040
        # Get the results of the ray workers.
        if self.workers:
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
            if use_ray_compiled_dag:
                try:
                    ray_worker_outputs = [
                        pickle.loads(chan.begin_read())
                        for chan in output_channels
                    ]
                finally:
                    # Has to call end_read in order to reuse the DAG.
                    for chan in output_channels:
                        chan.end_read()
            else:
                ray_worker_outputs = ray.get(ray_worker_outputs)
1053

1054
        return [driver_worker_output] + ray_worker_outputs
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074

    def _compiled_ray_dag(self):
        import pkg_resources
        required_version = "2.9"
        current_version = pkg_resources.get_distribution("ray").version
        if current_version < required_version:
            raise ValueError(f"Ray version {required_version} or greater is "
                             f"required, but found {current_version}")

        from ray.dag import MultiOutputNode, InputNode
        assert self.parallel_config.worker_use_ray

        # Right now, compiled DAG requires at least 1 arg. We send
        # a dummy value for now. It will be fixed soon.
        with InputNode() as input_data:
            forward_dag = MultiOutputNode([
                worker.execute_model_compiled_dag_remote.bind(input_data)
                for worker in self.workers
            ])
        return forward_dag.experimental_compile()