llm_engine.py 89.8 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import copy
Antoni Baum's avatar
Antoni Baum committed
5
import time
6
from collections import Counter as collectionsCounter
7
from collections import deque
8
from contextlib import contextmanager
9
from dataclasses import dataclass
10
from functools import partial
11
from typing import (TYPE_CHECKING, Any, Callable, ClassVar, Deque, Dict,
12
                    Iterable, List, Literal, Mapping, NamedTuple, Optional)
13
from typing import Sequence as GenericSequence
14
from typing import Set, Type, Union, cast
15

16
import torch
17
from typing_extensions import TypeVar
18

19
import vllm.envs as envs
20
21
22
from vllm.config import (DecodingConfig, LoRAConfig, ModelConfig,
                         ObservabilityConfig, ParallelConfig, SchedulerConfig,
                         VllmConfig)
23
from vllm.core.scheduler import ScheduledSequenceGroup, SchedulerOutputs
Woosuk Kwon's avatar
Woosuk Kwon committed
24
from vllm.engine.arg_utils import EngineArgs
25
from vllm.engine.metrics_types import StatLoggerBase, Stats
26
27
28
29
from vllm.engine.output_processor.interfaces import (
    SequenceGroupOutputProcessor)
from vllm.engine.output_processor.stop_checker import StopChecker
from vllm.engine.output_processor.util import create_output_by_sequence_group
30
31
from vllm.entrypoints.openai.logits_processors import (
    get_logits_processors as get_openai_logits_processors)
32
from vllm.executor.executor_base import ExecutorBase
33
from vllm.inputs import ProcessorInputs, PromptType, SingletonInputs
34
from vllm.inputs.parse import split_enc_dec_inputs
35
from vllm.inputs.preprocess import InputPreprocessor
Woosuk Kwon's avatar
Woosuk Kwon committed
36
from vllm.logger import init_logger
37
from vllm.logits_process import get_bad_words_logits_processors
38
from vllm.lora.request import LoRARequest
39
40
from vllm.model_executor.guided_decoding import (
    get_local_guided_decoding_logits_processor)
41
from vllm.model_executor.layers.sampler import SamplerOutput
42
from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalRegistry
43
from vllm.multimodal.processing import EncDecMultiModalProcessor
44
from vllm.outputs import (PoolingRequestOutput, RequestOutput,
45
46
                          RequestOutputFactory)
from vllm.pooling_params import PoolingParams
47
from vllm.prompt_adapter.request import PromptAdapterRequest
48
from vllm.sampling_params import RequestOutputKind, SamplingParams
49
50
51
52
from vllm.sequence import (ExecuteModelRequest, ParallelSampleSequenceGroup,
                           PoolingSequenceGroupOutput, Sequence, SequenceGroup,
                           SequenceGroupBase, SequenceGroupMetadata,
                           SequenceGroupOutput, SequenceStatus)
53
54
from vllm.tracing import (SpanAttributes, SpanKind, extract_trace_context,
                          init_tracer)
55
from vllm.transformers_utils.detokenizer import Detokenizer
56
from vllm.transformers_utils.tokenizer import AnyTokenizer
57
from vllm.transformers_utils.tokenizer_group import (
58
    TokenizerGroup, init_tokenizer_from_configs)
yhu422's avatar
yhu422 committed
59
60
from vllm.usage.usage_lib import (UsageContext, is_usage_stats_enabled,
                                  usage_message)
61
from vllm.utils import Counter, Device, resolve_obj_by_qualname, weak_bind
62
from vllm.version import __version__ as VLLM_VERSION
63
from vllm.worker.model_runner_base import InputProcessingError
64
65

logger = init_logger(__name__)
66
_LOCAL_LOGGING_INTERVAL_SEC = 5
Woosuk Kwon's avatar
Woosuk Kwon committed
67

68
_O = TypeVar("_O", RequestOutput, PoolingRequestOutput)
69
_R = TypeVar("_R", default=Any)
70
71


72
73
74
75
76
@dataclass
class SchedulerOutputState:
    """Caches the scheduler outputs for a virtual engine. Used for Multi-Step"""
    seq_group_metadata_list: Optional[List[SequenceGroupMetadata]] = None
    scheduler_outputs: Optional[SchedulerOutputs] = None
77
78
    allow_async_output_proc: bool = False
    last_output: Optional[SamplerOutput] = None
79
80


81
82
83
84
85
86
class OutputData(NamedTuple):
    outputs: List[SamplerOutput]
    seq_group_metadata_list: List[SequenceGroupMetadata]
    scheduler_outputs: SchedulerOutputs
    is_async: bool
    is_last_step: bool
87
88
89
90
91
92
    # Indicates if this output is from the first step of the
    # multi-step. When multi-step is disabled, this is always
    # set to True.
    # is_first_step_output is invalid when `outputs` has
    # outputs from multiple steps.
    is_first_step_output: Optional[bool]
93
94
95
    skip: List[int]


96
class SchedulerContext:
97

98
    def __init__(self, multi_step_stream_outputs: bool = False):
99
100
        self.output_queue: Deque[OutputData] = deque()
        self.request_outputs: List[Union[RequestOutput,
101
                                         PoolingRequestOutput]] = []
102
103
104
105
        self.seq_group_metadata_list: Optional[
            List[SequenceGroupMetadata]] = None
        self.scheduler_outputs: Optional[SchedulerOutputs] = None

106
107
        self.multi_step_stream_outputs: bool = multi_step_stream_outputs

108
109
110
    def append_output(self, outputs: List[SamplerOutput],
                      seq_group_metadata_list: List[SequenceGroupMetadata],
                      scheduler_outputs: SchedulerOutputs, is_async: bool,
111
112
                      is_last_step: bool,
                      is_first_step_output: Optional[bool]):
113
114
115
116
117
118
        self.output_queue.append(
            OutputData(outputs=outputs,
                       seq_group_metadata_list=seq_group_metadata_list,
                       scheduler_outputs=scheduler_outputs,
                       is_async=is_async,
                       is_last_step=is_last_step,
119
                       is_first_step_output=is_first_step_output,
120
                       skip=[]))
121
122


123
class LLMEngine:
Zhuohan Li's avatar
Zhuohan Li committed
124
    """An LLM engine that receives requests and generates texts.
125

Woosuk Kwon's avatar
Woosuk Kwon committed
126
    This is the main class for the vLLM engine. It receives requests
127
128
129
130
131
132
    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.

133
134
135
    The [`LLM`][vllm.LLM] class wraps this class for offline batched inference
    and the [`AsyncLLMEngine`][vllm.engine.async_llm_engine.AsyncLLMEngine]
    class wraps this class for online serving.
136

137
    The config arguments are derived from [`EngineArgs`][vllm.EngineArgs].
138
139

    Args:
140
        vllm_config: The configuration for initializing and running vLLM.
141
142
        executor_class: The model executor class for managing distributed
            execution.
143
        log_stats: Whether to log statistics.
144
        usage_context: Specified entry point, used for usage info collection.
145
    """
146

147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
    DO_VALIDATE_OUTPUT: ClassVar[bool] = False
    """A flag to toggle whether to validate the type of request output."""

    @classmethod
    @contextmanager
    def enable_output_validation(cls):
        cls.DO_VALIDATE_OUTPUT = True

        yield

        cls.DO_VALIDATE_OUTPUT = False

    @classmethod
    def validate_output(
        cls,
        output: object,
        output_type: Type[_O],
    ) -> _O:
        do_validate = cls.DO_VALIDATE_OUTPUT

        if ((TYPE_CHECKING or do_validate)
                and not isinstance(output, output_type)):
            raise TypeError(f"Expected output of type {output_type}, "
                            f"but found type {type(output)}")

172
        return cast(_O, output)
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195

    @classmethod
    def validate_outputs(
        cls,
        outputs: GenericSequence[object],
        output_type: Type[_O],
    ) -> List[_O]:
        do_validate = cls.DO_VALIDATE_OUTPUT

        outputs_: List[_O]
        if TYPE_CHECKING or do_validate:
            outputs_ = []
            for output in outputs:
                if not isinstance(output, output_type):
                    raise TypeError(f"Expected output of type {output_type}, "
                                    f"but found type {type(output)}")

                outputs_.append(output)
        else:
            outputs_ = outputs

        return outputs_

196
    tokenizer: Optional[TokenizerGroup]
197

198
199
    def __init__(
        self,
200
        vllm_config: VllmConfig,
201
        executor_class: Type[ExecutorBase],
202
        log_stats: bool,
yhu422's avatar
yhu422 committed
203
        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
204
        stat_loggers: Optional[Dict[str, StatLoggerBase]] = None,
205
        mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
206
        use_cached_outputs: bool = False,
207
    ) -> None:
208
209
210
211
212
213
        if envs.VLLM_USE_V1:
            raise ValueError(
                "Using V0 LLMEngine, but envs.VLLM_USE_V1=True. "
                "This should not happen. As a workaround, try using "
                "LLMEngine.from_vllm_config(...) or explicitly set "
                "VLLM_USE_V1=0 or 1 and report this issue on Github.")
214

215
        self.vllm_config = vllm_config
216
217
218
219
220
221
222
223
224
        self.model_config = vllm_config.model_config
        self.cache_config = vllm_config.cache_config
        self.lora_config = vllm_config.lora_config
        self.parallel_config = vllm_config.parallel_config
        self.scheduler_config = vllm_config.scheduler_config
        self.device_config = vllm_config.device_config
        self.speculative_config = vllm_config.speculative_config  # noqa
        self.load_config = vllm_config.load_config
        self.decoding_config = vllm_config.decoding_config or DecodingConfig(  # noqa
225
        )
226
227
        self.prompt_adapter_config = vllm_config.prompt_adapter_config  # noqa
        self.observability_config = vllm_config.observability_config or ObservabilityConfig(  # noqa
228
229
        )

230
        logger.info(
231
            "Initializing a V0 LLM engine (v%s) with config: %s, "
232
            "use_cached_outputs=%s, ",
233
            VLLM_VERSION,
234
            vllm_config,
235
            use_cached_outputs,
236
        )
237

238
        self.log_stats = log_stats
239
        self.use_cached_outputs = use_cached_outputs
240

241
        if not self.model_config.skip_tokenizer_init:
242
            self.tokenizer = self._init_tokenizer()
243
            self.detokenizer = Detokenizer(self.tokenizer)
244
            tokenizer_group = self.get_tokenizer_group()
245
246
        else:
            self.tokenizer = None
247
            self.detokenizer = None
248
249
250
251
252
253
254
255
            tokenizer_group = None

        # Ensure that the function doesn't contain a reference to self,
        # to avoid engine GC issues
        def get_tokenizer_for_seq(sequence: Sequence) -> AnyTokenizer:
            assert tokenizer_group, ("tokenizer_group cannot be None, "
                                     "make sure skip_tokenizer_init is False")
            return tokenizer_group.get_lora_tokenizer(sequence.lora_request)
256

257
        self.seq_counter = Counter()
258
259
        self.generation_config_fields = (
            self.model_config.try_get_generation_config())
260

261
        self.input_preprocessor = InputPreprocessor(self.model_config,
262
263
                                                    self.tokenizer,
                                                    mm_registry)
264

265
        self.model_executor = executor_class(vllm_config=vllm_config)
266

267
        if self.model_config.runner_type != "pooling":
268
            self._initialize_kv_caches()
269

yhu422's avatar
yhu422 committed
270
271
        # If usage stat is enabled, collect relevant info.
        if is_usage_stats_enabled():
272
273
            from vllm.model_executor.model_loader import (
                get_architecture_class_name)
yhu422's avatar
yhu422 committed
274
            usage_message.report_usage(
275
                get_architecture_class_name(self.model_config),
yhu422's avatar
yhu422 committed
276
277
278
279
                usage_context,
                extra_kvs={
                    # Common configuration
                    "dtype":
280
                    str(self.model_config.dtype),
yhu422's avatar
yhu422 committed
281
                    "tensor_parallel_size":
282
                    self.parallel_config.tensor_parallel_size,
yhu422's avatar
yhu422 committed
283
                    "block_size":
284
                    self.cache_config.block_size,
yhu422's avatar
yhu422 committed
285
                    "gpu_memory_utilization":
286
                    self.cache_config.gpu_memory_utilization,
yhu422's avatar
yhu422 committed
287
288
289

                    # Quantization
                    "quantization":
290
                    self.model_config.quantization,
yhu422's avatar
yhu422 committed
291
                    "kv_cache_dtype":
292
                    str(self.cache_config.cache_dtype),
yhu422's avatar
yhu422 committed
293
294
295

                    # Feature flags
                    "enable_lora":
296
                    bool(self.lora_config),
297
                    "enable_prompt_adapter":
298
                    bool(self.prompt_adapter_config),
yhu422's avatar
yhu422 committed
299
                    "enable_prefix_caching":
300
                    self.cache_config.enable_prefix_caching,
yhu422's avatar
yhu422 committed
301
                    "enforce_eager":
302
                    self.model_config.enforce_eager,
yhu422's avatar
yhu422 committed
303
                    "disable_custom_all_reduce":
304
                    self.parallel_config.disable_custom_all_reduce,
yhu422's avatar
yhu422 committed
305
306
                })

307
308
309
310
311
312
        self.cached_scheduler_outputs = [
            SchedulerOutputState()
            for _ in range(self.parallel_config.pipeline_parallel_size)
        ]

        self.scheduler_contexts = [
313
314
            SchedulerContext(multi_step_stream_outputs=self.scheduler_config.
                             multi_step_stream_outputs)
315
316
317
            for _ in range(self.parallel_config.pipeline_parallel_size)
        ]

318
        if self.model_config.use_async_output_proc:
319
320
321
322
323
324
325
326
327
            process_model_outputs = weak_bind(self._process_model_outputs)

            self.async_callbacks = [
                partial(process_model_outputs,
                        ctx=self.scheduler_contexts[v_id])
                for v_id in range(self.parallel_config.pipeline_parallel_size)
            ]
        else:
            self.async_callbacks = []
328
329
330

        # Currently used by AsyncLLMEngine to ensure quick append
        # of request outputs to asyncio queues
331
        self.process_request_outputs_callback: Optional[Callable] = None
332

333
        # Create the scheduler.
334
335
        # NOTE: the cache_config here have been updated with the numbers of
        # GPU and CPU blocks, which are profiled in the distributed executor.
336
337
338
339
340
        if isinstance(self.vllm_config.scheduler_config.scheduler_cls, str):
            Scheduler = resolve_obj_by_qualname(
                self.vllm_config.scheduler_config.scheduler_cls)
        else:
            Scheduler = self.vllm_config.scheduler_config.scheduler_cls
341
        self.scheduler = [
342
            Scheduler(
343
344
                self.scheduler_config, self.cache_config, self.lora_config,
                self.parallel_config.pipeline_parallel_size,
345
                self.async_callbacks[v_id]
346
347
                if self.model_config.use_async_output_proc else None)
            for v_id in range(self.parallel_config.pipeline_parallel_size)
348
        ]
Woosuk Kwon's avatar
Woosuk Kwon committed
349

350
351
        # Metric Logging.
        if self.log_stats:
352
353
354
            if stat_loggers is not None:
                self.stat_loggers = stat_loggers
            else:
355
356
357
358
359
360
361
                # Lazy import for prometheus multiprocessing.
                # We need to set PROMETHEUS_MULTIPROC_DIR environment variable
                # before prometheus_client is imported.
                # See https://prometheus.github.io/client_python/multiprocess/
                from vllm.engine.metrics import (LoggingStatLogger,
                                                 PrometheusStatLogger)

362
363
364
                self.stat_loggers = {
                    "logging":
                    LoggingStatLogger(
365
366
                        local_interval=_LOCAL_LOGGING_INTERVAL_SEC,
                        vllm_config=vllm_config),
367
368
369
                    "prometheus":
                    PrometheusStatLogger(
                        local_interval=_LOCAL_LOGGING_INTERVAL_SEC,
370
371
                        labels=dict(
                            model_name=self.model_config.served_model_name),
372
                        vllm_config=vllm_config),
373
374
375
                }
                self.stat_loggers["prometheus"].info("cache_config",
                                                     self.cache_config)
376

377
378
379
380
381
382
        self.tracer = None
        if self.observability_config.otlp_traces_endpoint:
            self.tracer = init_tracer(
                "vllm.llm_engine",
                self.observability_config.otlp_traces_endpoint)

383
384
385
386
387
388
389
390
        # Create sequence output processor, e.g. for beam search or
        # speculative decoding.
        self.output_processor = (
            SequenceGroupOutputProcessor.create_output_processor(
                self.scheduler_config,
                self.detokenizer,
                self.scheduler,
                self.seq_counter,
391
                get_tokenizer_for_seq,
392
393
                stop_checker=StopChecker(self.scheduler_config.max_model_len,
                                         get_tokenizer_for_seq),
394
395
            ))

396
397
        self.seq_id_to_seq_group: Dict[str, SequenceGroupBase] = {}

398
399
400
401
        # Flag to set when an input fails to process and the engine should run
        # the next step without re-scheduling.
        self._skip_scheduling_next_step = False

402
403
404
        # Don't keep the dummy data in memory
        self.reset_mm_cache()

405
406
407
408
409
410
    def _initialize_kv_caches(self) -> None:
        """Initialize the KV cache in the worker(s).

        The workers will determine the number of blocks in both the GPU cache
        and the swap CPU cache.
        """
411
        start = time.time()
412
413
414
415
416
        num_gpu_blocks, num_cpu_blocks = (
            self.model_executor.determine_num_available_blocks())

        if self.cache_config.num_gpu_blocks_override is not None:
            num_gpu_blocks_override = self.cache_config.num_gpu_blocks_override
417
418
419
420
            logger.info(
                "Overriding num_gpu_blocks=%d with "
                "num_gpu_blocks_override=%d", num_gpu_blocks,
                num_gpu_blocks_override)
421
422
423
424
425
426
            num_gpu_blocks = num_gpu_blocks_override

        self.cache_config.num_gpu_blocks = num_gpu_blocks
        self.cache_config.num_cpu_blocks = num_cpu_blocks

        self.model_executor.initialize_cache(num_gpu_blocks, num_cpu_blocks)
427
428
429
        elapsed = time.time() - start
        logger.info(("init engine (profile, create kv cache, "
                     "warmup model) took %.2f seconds"), elapsed)
430

431
    @classmethod
432
    def _get_executor_cls(cls,
433
                          engine_config: VllmConfig) -> Type[ExecutorBase]:
434
        # distributed_executor_backend must be set in VllmConfig.__post_init__
435
436
        distributed_executor_backend = (
            engine_config.parallel_config.distributed_executor_backend)
437
        # Initialize the cluster and specify the executor class.
438
439
440
441
442
443
        if isinstance(distributed_executor_backend, type):
            if not issubclass(distributed_executor_backend, ExecutorBase):
                raise TypeError(
                    "distributed_executor_backend must be a subclass of "
                    f"ExecutorBase. Got {distributed_executor_backend}.")
            executor_class = distributed_executor_backend
444
445
446
447
448
449
450
451
452
453
454
455
456
        elif distributed_executor_backend == "ray":
            from vllm.executor.ray_distributed_executor import (
                RayDistributedExecutor)
            executor_class = RayDistributedExecutor
        elif distributed_executor_backend == "mp":
            from vllm.executor.mp_distributed_executor import (
                MultiprocessingDistributedExecutor)
            assert not envs.VLLM_USE_RAY_SPMD_WORKER, (
                "multiprocessing distributed executor backend does not "
                "support VLLM_USE_RAY_SPMD_WORKER=1")
            executor_class = MultiprocessingDistributedExecutor
        elif distributed_executor_backend == "uni":
            # JAX-style, single-process, multi-device executor.
457
458
            from vllm.executor.uniproc_executor import UniProcExecutor
            executor_class = UniProcExecutor
459
460
461
462
463
464
465
466
        elif distributed_executor_backend == "external_launcher":
            # executor with external launcher
            from vllm.executor.uniproc_executor import (  # noqa
                ExecutorWithExternalLauncher)
            executor_class = ExecutorWithExternalLauncher
        else:
            raise ValueError("unrecognized distributed_executor_backend: "
                             f"{distributed_executor_backend}")
467
468
        return executor_class

469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
    @classmethod
    def from_vllm_config(
        cls,
        vllm_config: VllmConfig,
        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
        stat_loggers: Optional[Dict[str, StatLoggerBase]] = None,
        disable_log_stats: bool = False,
    ) -> "LLMEngine":
        return cls(
            vllm_config=vllm_config,
            executor_class=cls._get_executor_cls(vllm_config),
            log_stats=(not disable_log_stats),
            usage_context=usage_context,
            stat_loggers=stat_loggers,
        )

485
486
487
488
489
490
491
492
493
    @classmethod
    def from_engine_args(
        cls,
        engine_args: EngineArgs,
        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
        stat_loggers: Optional[Dict[str, StatLoggerBase]] = None,
    ) -> "LLMEngine":
        """Creates an LLM engine from the engine arguments."""
        # Create the engine configs.
494
495
496
497
498
499
500
501
502
        vllm_config = engine_args.create_engine_config(usage_context)

        engine_cls = cls
        if envs.VLLM_USE_V1:
            from vllm.v1.engine.llm_engine import LLMEngine as V1LLMEngine
            engine_cls = V1LLMEngine

        return engine_cls.from_vllm_config(
            vllm_config=vllm_config,
yhu422's avatar
yhu422 committed
503
            usage_context=usage_context,
504
            stat_loggers=stat_loggers,
505
            disable_log_stats=engine_args.disable_log_stats,
yhu422's avatar
yhu422 committed
506
        )
507

508
509
510
511
512
    def __reduce__(self):
        # This is to ensure that the LLMEngine is not referenced in
        # the closure used to initialize Ray worker actors
        raise RuntimeError("LLMEngine should not be pickled!")

513
514
515
516
517
518
    def __del__(self):
        # Shutdown model executor when engine is garbage collected
        # Use getattr since __init__ can fail before the field is set
        if model_executor := getattr(self, "model_executor", None):
            model_executor.shutdown()

519
520
    def get_tokenizer_group(self) -> TokenizerGroup:
        if self.tokenizer is None:
521
522
            raise ValueError("Unable to get tokenizer because "
                             "skip_tokenizer_init is True")
523

524
        return self.tokenizer
525

526
    def get_tokenizer(
527
528
529
        self,
        lora_request: Optional[LoRARequest] = None,
    ) -> AnyTokenizer:
530
        return self.get_tokenizer_group().get_lora_tokenizer(lora_request)
531

532
    def _init_tokenizer(self) -> TokenizerGroup:
533
534
535
        return init_tokenizer_from_configs(
            model_config=self.model_config,
            scheduler_config=self.scheduler_config,
536
            lora_config=self.lora_config)
537

538
539
    def _verify_args(self) -> None:
        self.model_config.verify_with_parallel_config(self.parallel_config)
540
        self.cache_config.verify_with_parallel_config(self.parallel_config)
541
542
543
544
        if self.lora_config:
            self.lora_config.verify_with_model_config(self.model_config)
            self.lora_config.verify_with_scheduler_config(
                self.scheduler_config)
545
546
547
        if self.prompt_adapter_config:
            self.prompt_adapter_config.verify_with_model_config(
                self.model_config)
548

549
550
551
    def _add_processed_request(
        self,
        request_id: str,
552
        processed_inputs: ProcessorInputs,
553
554
555
        params: Union[SamplingParams, PoolingParams],
        arrival_time: float,
        lora_request: Optional[LoRARequest],
556
        prompt_adapter_request: Optional[PromptAdapterRequest],
557
        trace_headers: Optional[Mapping[str, str]] = None,
558
        priority: int = 0,
559
    ) -> Optional[SequenceGroup]:
560
561
562
        """Add a processed request to the engine's request pool.
        return the created sequence group.
        """
563
564
565
566
567
568
569
570
571
572
573
574
575
576
        if isinstance(params, SamplingParams) and params.n > 1:
            ParallelSampleSequenceGroup.add_request(
                request_id,
                self,
                params,
                processed_inputs=processed_inputs,
                arrival_time=arrival_time,
                lora_request=lora_request,
                trace_headers=trace_headers,
                prompt_adapter_request=prompt_adapter_request,
                priority=priority,
            )
            return None

577
        self._validate_model_inputs(processed_inputs, lora_request)
578
579
580
        # Create the sequences.
        block_size = self.cache_config.block_size
        seq_id = next(self.seq_counter)
581
        eos_token_id = self.input_preprocessor.get_eos_token_id(lora_request)
582

583
        encoder_inputs, decoder_inputs = split_enc_dec_inputs(processed_inputs)
584
585

        seq = Sequence(seq_id, decoder_inputs, block_size, eos_token_id,
586
                       lora_request, prompt_adapter_request)
587

588
589
590
        encoder_seq = (None if encoder_inputs is None else Sequence(
            seq_id, encoder_inputs, block_size, eos_token_id, lora_request,
            prompt_adapter_request))
591

592
593
594
595
596
597
598
599
        # Create a SequenceGroup based on SamplingParams or PoolingParams
        if isinstance(params, SamplingParams):
            seq_group = self._create_sequence_group_with_sampling(
                request_id,
                seq,
                params,
                arrival_time=arrival_time,
                lora_request=lora_request,
600
                trace_headers=trace_headers,
601
                prompt_adapter_request=prompt_adapter_request,
602
603
                encoder_seq=encoder_seq,
                priority=priority)
604
605
606
607
608
609
610
        elif isinstance(params, PoolingParams):
            seq_group = self._create_sequence_group_with_pooling(
                request_id,
                seq,
                params,
                arrival_time=arrival_time,
                lora_request=lora_request,
611
                prompt_adapter_request=prompt_adapter_request,
612
613
                encoder_seq=encoder_seq,
                priority=priority)
614
615
616
617
        else:
            raise ValueError(
                "Either SamplingParams or PoolingParams must be provided.")

618
619
620
621
622
623
624
625
        # Add the sequence group to the scheduler with least unfinished seqs.
        costs = [
            scheduler.get_num_unfinished_seq_groups()
            for scheduler in self.scheduler
        ]
        min_cost_scheduler = self.scheduler[costs.index(min(costs))]
        min_cost_scheduler.add_seq_group(seq_group)

626
627
        return seq_group

628
629
    def stop_remote_worker_execution_loop(self) -> None:
        self.model_executor.stop_remote_worker_execution_loop()
630

631
632
633
    def add_request(
        self,
        request_id: str,
634
        prompt: PromptType,
635
        params: Union[SamplingParams, PoolingParams],
636
        arrival_time: Optional[float] = None,
637
        lora_request: Optional[LoRARequest] = None,
638
        tokenization_kwargs: Optional[dict[str, Any]] = None,
639
        trace_headers: Optional[Mapping[str, str]] = None,
640
        prompt_adapter_request: Optional[PromptAdapterRequest] = None,
641
        priority: int = 0,
642
    ) -> None:
Zhuohan Li's avatar
Zhuohan Li committed
643
        """Add a request to the engine's request pool.
644
645

        The request is added to the request pool and will be processed by the
Zhuohan Li's avatar
Zhuohan Li committed
646
        scheduler as `engine.step()` is called. The exact scheduling policy is
647
648
649
650
        determined by the scheduler.

        Args:
            request_id: The unique ID of the request.
651
652
            prompt: The prompt to the LLM. See
                [PromptType][vllm.inputs.PromptType]
653
654
                for more details about the format of each input.
            params: Parameters for sampling or pooling.
655
656
                [SamplingParams][vllm.SamplingParams] for text generation.
                [PoolingParams][vllm.PoolingParams] for pooling.
657
            arrival_time: The arrival time of the request. If None, we use
658
                the current monotonic time.
659
            lora_request: The LoRA request to add.
660
            trace_headers: OpenTelemetry trace headers.
661
            prompt_adapter_request: The prompt adapter request to add.
662
663
            priority: The priority of the request.
                Only applicable with priority scheduling.
664
665
666
667

        Details:
            - Set arrival_time to the current time if it is None.
            - Set prompt_token_ids to the encoded prompt if it is None.
668
669
670
671
672
            - Create `n` number of [Sequence][vllm.Sequence] objects.
            - Create a [SequenceGroup][vllm.SequenceGroup] object
              from the list of [Sequence][vllm.Sequence].
            - Add the [SequenceGroup][vllm.SequenceGroup] object to the
              scheduler.
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688

        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
            >>> ...
689
        """
690
691
692
        if lora_request is not None and not self.lora_config:
            raise ValueError(f"Got lora_request {lora_request} but LoRA is "
                             "not enabled!")
693

694
        if priority != 0 and not self.scheduler_config.policy == "priority":
695
696
697
            raise ValueError(f"Got priority {priority} but "
                             "Priority scheduling is not enabled.")

698
699
700
701
702
703
704
        if isinstance(params, SamplingParams) \
            and (params.guided_decoding or params.logits_processors) \
            and self.scheduler_config.num_scheduler_steps > 1:
            raise ValueError(
                "Guided decoding and logits processors are not supported "
                "in multi-step decoding")

705
        if arrival_time is None:
706
            arrival_time = time.time()
707

708
709
710
711
712
713
        if (isinstance(prompt, dict)
                and prompt.get("prompt_embeds", None) is not None
                and not prompt.get("prompt_token_ids", None)):
            seq_len = prompt["prompt_embeds"].shape[0]
            prompt["prompt_token_ids"] = [0] * seq_len

714
        processed_inputs = self.input_preprocessor.preprocess(
715
            prompt,
716
            tokenization_kwargs=tokenization_kwargs,
717
            lora_request=lora_request,
718
719
            prompt_adapter_request=prompt_adapter_request,
        )
720

721
        self._add_processed_request(
722
723
724
725
726
            request_id=request_id,
            processed_inputs=processed_inputs,
            params=params,
            arrival_time=arrival_time,
            lora_request=lora_request,
727
            prompt_adapter_request=prompt_adapter_request,
728
            trace_headers=trace_headers,
729
            priority=priority,
730
        )
731
732
733
734
735
736

    def _create_sequence_group_with_sampling(
        self,
        request_id: str,
        seq: Sequence,
        sampling_params: SamplingParams,
737
738
        arrival_time: float,
        lora_request: Optional[LoRARequest],
739
        trace_headers: Optional[Mapping[str, str]] = None,
740
        prompt_adapter_request: Optional[PromptAdapterRequest] = None,
741
        encoder_seq: Optional[Sequence] = None,
742
        priority: int = 0,
743
744
745
746
747
748
749
750
751
752
    ) -> SequenceGroup:
        """Creates a SequenceGroup with SamplingParams."""
        max_logprobs = self.get_model_config().max_logprobs
        if (sampling_params.logprobs
                and sampling_params.logprobs > max_logprobs) or (
                    sampling_params.prompt_logprobs
                    and sampling_params.prompt_logprobs > max_logprobs):
            raise ValueError(f"Cannot request more than "
                             f"{max_logprobs} logprobs.")

753
754
755
        sampling_params = self._build_logits_processors(
            sampling_params, lora_request)

756
757
758
        # Defensive copy of SamplingParams, which are used by the sampler,
        # this doesn't deep-copy LogitsProcessor objects
        sampling_params = sampling_params.clone()
759

760
        sampling_params.update_from_generation_config(
761
            self.generation_config_fields, seq.eos_token_id)
762

763
        # Create the sequence group.
764
765
766
767
        draft_size = 1
        if self.vllm_config.speculative_config is not None:
            draft_size = \
                self.vllm_config.speculative_config.num_speculative_tokens + 1
768
769
770
771
772
773
774
        seq_group = SequenceGroup(
            request_id=request_id,
            seqs=[seq],
            arrival_time=arrival_time,
            sampling_params=sampling_params,
            lora_request=lora_request,
            trace_headers=trace_headers,
775
            prompt_adapter_request=prompt_adapter_request,
776
            encoder_seq=encoder_seq,
777
778
            priority=priority,
            draft_size=draft_size)
779

780
781
782
783
784
785
786
        return seq_group

    def _create_sequence_group_with_pooling(
        self,
        request_id: str,
        seq: Sequence,
        pooling_params: PoolingParams,
787
788
        arrival_time: float,
        lora_request: Optional[LoRARequest],
789
        prompt_adapter_request: Optional[PromptAdapterRequest],
790
        encoder_seq: Optional[Sequence] = None,
791
        priority: int = 0,
792
793
794
795
796
    ) -> SequenceGroup:
        """Creates a SequenceGroup with PoolingParams."""
        # Defensive copy of PoolingParams, which are used by the pooler
        pooling_params = pooling_params.clone()
        # Create the sequence group.
797
798
799
800
801
802
        seq_group = SequenceGroup(
            request_id=request_id,
            seqs=[seq],
            arrival_time=arrival_time,
            lora_request=lora_request,
            pooling_params=pooling_params,
803
            prompt_adapter_request=prompt_adapter_request,
804
805
            encoder_seq=encoder_seq,
            priority=priority)
806
        return seq_group
807

Antoni Baum's avatar
Antoni Baum committed
808
809
    def abort_request(self, request_id: Union[str, Iterable[str]]) -> None:
        """Aborts a request(s) with the given ID.
810
811

        Args:
Antoni Baum's avatar
Antoni Baum committed
812
            request_id: The ID(s) of the request to abort.
813
814

        Details:
815
            - Refer to [vllm.core.scheduler.Scheduler.abort_seq_group][].
816
817
818
819
820
821

        Example:
            >>> # initialize engine and add a request with request_id
            >>> request_id = str(0)
            >>> # abort the request
            >>> engine.abort_request(request_id)
822
        """
823
        for scheduler in self.scheduler:
824
825
            scheduler.abort_seq_group(
                request_id, seq_id_to_seq_group=self.seq_id_to_seq_group)
826

827
828
829
830
    def get_vllm_config(self) -> VllmConfig:
        """Gets the vllm configuration."""
        return self.vllm_config

831
832
833
834
    def get_model_config(self) -> ModelConfig:
        """Gets the model configuration."""
        return self.model_config

835
836
837
838
    def get_parallel_config(self) -> ParallelConfig:
        """Gets the parallel configuration."""
        return self.parallel_config

839
840
841
842
    def get_decoding_config(self) -> DecodingConfig:
        """Gets the decoding configuration."""
        return self.decoding_config

843
844
845
846
847
848
849
850
    def get_scheduler_config(self) -> SchedulerConfig:
        """Gets the scheduler configuration."""
        return self.scheduler_config

    def get_lora_config(self) -> LoRAConfig:
        """Gets the LoRA configuration."""
        return self.lora_config

851
    def get_num_unfinished_requests(self) -> int:
852
        """Gets the number of unfinished requests."""
853
854
        return sum(scheduler.get_num_unfinished_seq_groups()
                   for scheduler in self.scheduler)
855

856
    def has_unfinished_requests(self) -> bool:
857
        """Returns True if there are unfinished requests."""
858
859
860
861
862
863
864
865
866
        return any(scheduler.has_unfinished_seqs()
                   for scheduler in self.scheduler)

    def has_unfinished_requests_for_virtual_engine(
            self, virtual_engine: int) -> bool:
        """
        Returns True if there are unfinished requests for the virtual engine.
        """
        return self.scheduler[virtual_engine].has_unfinished_seqs()
867

868
869
870
871
    def reset_mm_cache(self) -> bool:
        """Reset the multi-modal cache."""
        return self.input_preprocessor.mm_registry.reset_processor_cache()

872
    def reset_prefix_cache(self, device: Optional[Device] = None) -> bool:
873
874
875
876
        """Reset prefix cache for all devices."""

        success = True
        for scheduler in self.scheduler:
877
            success = success and scheduler.reset_prefix_cache(device)
878
879
        return success

880
    @staticmethod
881
882
    def _process_sequence_group_outputs(
        seq_group: SequenceGroup,
883
        outputs: List[PoolingSequenceGroupOutput],
884
    ) -> None:
885
        seq_group.pooled_data = outputs[0].data
886
887
888
889
890
891

        for seq in seq_group.get_seqs():
            seq.status = SequenceStatus.FINISHED_STOPPED

        return

892
893
894
895
896
897
898
899
    def _update_num_computed_tokens_for_multi_step_prefill(
            self, seq_group: SequenceGroup,
            seq_group_meta: SequenceGroupMetadata,
            is_first_step_output: Optional[bool]):
        """
        This function updates num_computed_tokens for prompt sequences
        when Multi-Step is enabled.

900
        seq_group: SequenceGroup to update the num_computed_tokens for.
901
        seq_group_meta: Metadata of the given SequenceGroup.
902
        is_first_step_output: Optional[bool] -
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
            When available, is_first_step_output indicates if the appended
            output token is the output of the first-step in multi-step.
            A value of None indicates that outputs from all steps in
            in multi-step are submitted in a single burst.
        """

        assert self.scheduler_config.is_multi_step

        if not seq_group_meta.is_prompt:
            # num_computed_token updates for multi-step decodes happen after
            # the tokens are appended to the sequence.
            return

        do_update: bool = False
        if self.scheduler_config.chunked_prefill_enabled:
            # In multi-step + chunked-prefill case, the prompt sequences
            # that are scheduled are fully processed in the first step.
            do_update = is_first_step_output is None or is_first_step_output
        else:
            # Normal multi-step decoding case. In this case prompt-sequences
            # are actually single-stepped. Always update in this case.
            assert seq_group.state.num_steps == 1
            do_update = True

        if do_update:
            seq_group.update_num_computed_tokens(
                seq_group_meta.token_chunk_size)

931
932
933
934
935
    def _process_model_outputs(self,
                               ctx: SchedulerContext,
                               request_id: Optional[str] = None) -> None:
        """Apply the model output to the sequences in the scheduled seq groups
        and return responses.
936

937
938
        ctx: The virtual engine context to work on
        request_id: If provided, then only this request is going to be processed
939
        """
940

941
        now = time.time()
942

943
        if len(ctx.output_queue) == 0:
944
945
            return None

946
        # Get pending async postprocessor
947
948
949
950
        if request_id:
            # When we process only one request, no pop is required
            # (since later we will process all of the rest)
            (outputs, seq_group_metadata_list, scheduler_outputs, is_async,
951
             is_last_step, is_first_step_output, skip) = ctx.output_queue[0]
952
953
        else:
            (outputs, seq_group_metadata_list, scheduler_outputs, is_async,
954
955
             is_last_step, is_first_step_output,
             skip) = ctx.output_queue.popleft()
956
957
958
959
960

        # Sanity check
        assert len(seq_group_metadata_list) == len(
            scheduler_outputs.scheduled_seq_groups)

961
        has_multiple_outputs: bool = len(outputs) > 1
962
        outputs_by_sequence_group: List[List[SequenceGroupOutput]]
963
964
965
966
967
        if has_multiple_outputs:
            assert self.scheduler_config.is_multi_step or \
                     self.speculative_config
            # Organize outputs by [step][sequence group] instead of
            # [sequence group][step].
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
            if self.scheduler_config.is_multi_step:
                outputs_by_sequence_group = create_output_by_sequence_group(
                    outputs, len(seq_group_metadata_list))
            elif self.speculative_config:
                # Decodes are multi-steps while prefills are not, outputting at
                # most 1 token. Separate them so that we can trigger chunk
                # processing without having to pad or copy over prompts K times
                # to match decodes structure (costly with prompt_logprobs).
                num_prefills = sum(sg.is_prompt
                                   for sg in seq_group_metadata_list)
                prefills, decodes = outputs[:num_prefills], outputs[
                    num_prefills:]
                outputs_by_sequence_group = create_output_by_sequence_group(
                    decodes,
                    num_seq_groups=len(seq_group_metadata_list) - num_prefills)
                outputs_by_sequence_group = [p.outputs for p in prefills
                                             ] + outputs_by_sequence_group
985
986
987
            # We have outputs for multiple steps submitted in a single burst,
            # so invalidate is_first_step_output.
            is_first_step_output = None
988
989
990
        else:
            outputs_by_sequence_group = outputs

991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
        # Determine the requests we need to operate on
        if request_id:
            indices = []
            for i, seq_group_meta in enumerate(seq_group_metadata_list):
                if seq_group_meta.request_id == request_id:
                    assert i not in skip  # Cannot be called twice
                    indices.append(i)
                    break

            # If the request_id was not found, then it means that
            # this is a new request that has no pending async
            # postprocessor
            if not indices:
                return
        else:
            indices = range(len(seq_group_metadata_list))  # type: ignore

1008
        finished_before: List[int] = []
1009
        finished_now: List[int] = []
1010
1011
1012
1013
1014
        for i in indices:
            if i in skip:
                continue

            seq_group_meta = seq_group_metadata_list[i]
1015
            scheduled_seq_group = scheduler_outputs.scheduled_seq_groups[i]
1016

1017
            seq_group: SequenceGroup = scheduled_seq_group.seq_group
1018
1019
1020
1021
1022

            if seq_group.is_finished():
                finished_before.append(i)
                continue

1023
            output: List[SequenceGroupOutput]
1024
            if has_multiple_outputs:
1025
1026
1027
1028
                output = outputs_by_sequence_group[i]
            else:
                output = [outputs_by_sequence_group[0][i]]

1029
1030
1031
1032
1033
1034
1035
            if not is_async:
                if self.scheduler_config.is_multi_step:
                    # Updates happen only if the sequence is prefill
                    self._update_num_computed_tokens_for_multi_step_prefill(
                        seq_group, seq_group_meta, is_first_step_output)
                else:
                    seq_group.update_num_computed_tokens(
1036
                        seq_group_meta.token_chunk_size or 0)
1037
1038
1039

            if outputs:
                for o in outputs:
1040
1041
1042
1043
                    if (isinstance(o, SamplerOutput)
                            and seq_group.metrics is not None):
                        if seq_group.metrics.model_forward_time is not None:
                            seq_group.metrics.model_forward_time += (
1044
                                o.model_forward_time or 0)
1045
1046
1047
1048
1049
                        else:
                            seq_group.metrics.model_forward_time = (
                                o.model_forward_time)
                        if seq_group.metrics.model_execute_time is not None:
                            seq_group.metrics.model_execute_time += (
1050
                                o.model_execute_time or 0)
1051
1052
1053
                        else:
                            seq_group.metrics.model_execute_time = (
                                o.model_execute_time)
1054

1055
            if self.model_config.runner_type == "pooling":
1056
                self._process_sequence_group_outputs(seq_group, output)
1057
1058
1059
            else:
                self.output_processor.process_prompt_logprob(seq_group, output)
                if seq_group_meta.do_sample:
1060
                    self.output_processor.process_outputs(
1061
                        seq_group, output, is_async)
1062

1063
1064
            if seq_group.is_finished():
                finished_now.append(i)
1065

1066
1067
1068
        # Generate outputs for the requests that finished this iteration
        for i in finished_now:
            scheduled_seq_group = scheduler_outputs.scheduled_seq_groups[i]
1069

1070
1071
            seq_group = scheduled_seq_group.seq_group
            seq_group.maybe_set_first_token_time(now)
1072
1073
            if not seq_group.is_prefill():
                seq_group.set_last_token_time(now)
1074
            request_output = RequestOutputFactory.create(
1075
1076
1077
                seq_group,
                self.seq_id_to_seq_group,
                use_cache=self.use_cached_outputs)
1078
1079
            if request_output:
                ctx.request_outputs.append(request_output)
1080

1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
        # When we process a single request, we skip it for the next time,
        # and invoke the request output callback (if there was final output)
        if request_id:
            assert len(indices) == 1
            skip.append(indices[0])

            if (finished_now
                    and self.process_request_outputs_callback is not None):
                self.process_request_outputs_callback(ctx.request_outputs)
                ctx.request_outputs.clear()
            return

1093
1094
1095
1096
1097
        # Free currently finished requests
        if finished_now:
            for scheduler in self.scheduler:
                scheduler.free_finished_seq_groups()

1098
1099
        # For multi-step without streaming, don't create outputs each iteration
        if not is_last_step and not ctx.multi_step_stream_outputs:
1100
1101
1102
1103
            # Immediately process request outputs here (if callback is given)
            if (finished_now
                    and self.process_request_outputs_callback is not None):
                self.process_request_outputs_callback(ctx.request_outputs)
1104
                ctx.request_outputs.clear()
1105
1106
1107
            return

        # Create the outputs
1108
1109
        for i in indices:
            if i in skip or i in finished_before or i in finished_now:
1110
1111
                continue  # Avoids double processing

1112
1113
            scheduled_seq_group = scheduler_outputs.scheduled_seq_groups[i]

1114
            seq_group = scheduled_seq_group.seq_group
1115
            seq_group.maybe_set_first_token_time(now)
1116
1117
            if not seq_group.is_prefill():
                seq_group.set_last_token_time(now)
1118
            request_output = RequestOutputFactory.create(
1119
1120
1121
                seq_group,
                self.seq_id_to_seq_group,
                use_cache=self.use_cached_outputs)
1122
            if request_output:
1123
                ctx.request_outputs.append(request_output)
1124

1125
1126
1127
1128
1129
1130
1131
1132
        # For multi-step with streaming, create outputs each iteration
        if not is_last_step and ctx.multi_step_stream_outputs:
            # Immediately process request outputs here (if callback is given)
            if self.process_request_outputs_callback is not None:
                self.process_request_outputs_callback(ctx.request_outputs)
                ctx.request_outputs.clear()
            return

1133
        for seq_group in scheduler_outputs.ignored_seq_groups:
1134
1135
1136
1137
1138
            params = seq_group.sampling_params
            if params is not None and params.output_kind == (
                    RequestOutputKind.DELTA) and not seq_group.is_finished():
                continue

1139
            request_output = RequestOutputFactory.create(
1140
1141
1142
1143
                seq_group,
                self.seq_id_to_seq_group,
                use_cache=self.use_cached_outputs,
            )
1144
1145
            if request_output:
                ctx.request_outputs.append(request_output)
1146

1147
1148
1149
1150
        # Immediately process request outputs here (if callback is given)
        if (ctx.request_outputs
                and self.process_request_outputs_callback is not None):
            self.process_request_outputs_callback(ctx.request_outputs)
1151
            ctx.request_outputs.clear()
1152

1153
1154
1155
1156
        # For async case, we need to record the stats here.
        # For non-async case, the stats are done in the
        # LLMEngine/AsyncLLMEngine directly
        if is_async:
1157
            # Log stats.
1158
1159
            self.do_log_stats(scheduler_outputs, outputs, finished_before,
                              skip)
1160
1161

            # Tracing
1162
            self.do_tracing(scheduler_outputs, finished_before)
1163
1164
1165
1166

        return None

    def _advance_to_next_step(
1167
            self, output: SamplerOutput,
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
            seq_group_metadata_list: List[SequenceGroupMetadata],
            scheduled_seq_groups: List[ScheduledSequenceGroup]) -> None:
        """Given model output from a single run, append the tokens to the
        sequences. This is normally done inside output processor, but it is
        required if the worker is to perform async forward pass to next step.
        """
        for seq_group_metadata, sequence_group_outputs, scheduled_seq_group in \
            zip(seq_group_metadata_list, output, scheduled_seq_groups):
            seq_group = scheduled_seq_group.seq_group

            if seq_group.is_finished():
                continue

1181
1182
1183
1184
1185
            if self.scheduler_config.is_multi_step:
                # Updates happen only if the sequence is prefill
                self._update_num_computed_tokens_for_multi_step_prefill(
                    seq_group, seq_group_metadata,
                    seq_group.state.num_steps == 1)
1186
            else:
1187
1188
1189
1190
                token_chunk_size = (seq_group_metadata.token_chunk_size
                                    if seq_group_metadata.token_chunk_size
                                    is not None else 0)
                seq_group.update_num_computed_tokens(token_chunk_size)
1191

1192
1193
1194
            if seq_group_metadata.do_sample:
                assert len(sequence_group_outputs.samples) == 1, (
                    "Async output processor expects a single sample"
1195
                    " (i.e sampling_params.n == 1)")
1196
1197
1198
1199
                sample = sequence_group_outputs.samples[0]

                assert len(seq_group.seqs) == 1
                seq = seq_group.seqs[0]
1200
1201
1202
1203

                if self.scheduler_config.is_multi_step:
                    is_prefill_append = seq.data.get_num_uncomputed_tokens(
                    ) == 0
1204
1205
                    seq.append_token_id(sample.output_token, sample.logprobs,
                                        sample.output_embed)
1206
1207
1208
                    if not is_prefill_append:
                        seq_group.update_num_computed_tokens(1)
                else:
1209
1210
                    seq.append_token_id(sample.output_token, sample.logprobs,
                                        sample.output_embed)
1211

1212
    def step(self) -> List[Union[RequestOutput, PoolingRequestOutput]]:
Antoni Baum's avatar
Antoni Baum committed
1213
1214
        """Performs one decoding iteration and returns newly generated results.

1215
1216
1217
1218
        <figure markdown="span">
        ![Overview of the step function](https://i.imgur.com/sv2HssD.png)
        <figcaption>Overview of the step function</figcaption>
        </figure>
1219
1220

        Details:
1221
1222
        - Step 1: Schedules the sequences to be executed in the next
            iteration and the token blocks to be swapped in/out/copy.
1223

1224
1225
1226
1227
            - 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.
1228

1229
1230
        - Step 2: Calls the distributed executor to execute the model.
        - Step 3: Processes the model output. This mainly includes:
1231

1232
1233
1234
1235
            - 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.
1236

1237
        - Finally, it creates and returns the newly generated results.
1238
1239

        Example:
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
        ```
        # 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
1263
        """
1264
1265
1266
1267
        if self.parallel_config.pipeline_parallel_size > 1:
            raise NotImplementedError(
                "Pipeline parallelism is only supported through AsyncLLMEngine "
                "as performance will be severely degraded otherwise.")
1268

1269
        # For llm_engine, there is no pipeline parallel support, so the engine
1270
        # used is always 0.
1271
1272
        virtual_engine = 0

1273
1274
        # These are cached outputs from previous iterations. None if on first
        # iteration
1275
        cached_outputs = self.cached_scheduler_outputs[virtual_engine]
1276
1277
        seq_group_metadata_list = cached_outputs.seq_group_metadata_list
        scheduler_outputs = cached_outputs.scheduler_outputs
1278
        allow_async_output_proc = cached_outputs.allow_async_output_proc
1279

1280
1281
        ctx = self.scheduler_contexts[virtual_engine]

1282
1283
1284
        # Clear outputs for each new scheduler iteration
        ctx.request_outputs.clear()

1285
1286
1287
        # Skip the scheduler if there are any remaining steps in the seq groups.
        # This ensures that the scheduler is only called again when the current
        # batch has completed.
1288
1289
1290
1291
1292
        # The scheduler is also skipped if a single request caused the last
        # engine step to fail, and the previous schedule needs to be rerun.
        if not self._has_remaining_steps(
                seq_group_metadata_list
        ) and not self._skip_scheduling_next_step:
1293
            # Schedule iteration
1294
            (seq_group_metadata_list, scheduler_outputs,
1295
1296
             allow_async_output_proc
             ) = self.scheduler[virtual_engine].schedule()
1297

1298
1299
            ctx.seq_group_metadata_list = seq_group_metadata_list
            ctx.scheduler_outputs = scheduler_outputs
1300

1301
1302
            finished_requests_ids = self.scheduler[
                virtual_engine].get_and_reset_finished_requests_ids()
1303
1304
1305
1306
1307
            # When n>1, elements in self.seq_id_to_seq_group should be deleted
            # here, otherwise memory leaks.
            for finished_request_id in finished_requests_ids:
                if finished_request_id in self.seq_id_to_seq_group:
                    del self.seq_id_to_seq_group[finished_request_id]
1308

1309
1310
            # Maybe switch from async mode to sync mode
            if not allow_async_output_proc and len(ctx.output_queue) > 0:
1311
                self._process_model_outputs(ctx=ctx)
1312

1313
1314
1315
1316
1317
            if (self.scheduler_config.is_multi_step
                    and scheduler_outputs.num_lookahead_slots > 0):
                # cache the scheduler outputs for the next iteration if we have
                # lookahead slots
                self._cache_scheduler_outputs_for_multi_step(
1318
                    virtual_engine, seq_group_metadata_list, scheduler_outputs,
1319
                    allow_async_output_proc)
1320
1321
        else:
            finished_requests_ids = list()
1322
1323
1324

        assert seq_group_metadata_list is not None
        assert scheduler_outputs is not None
Antoni Baum's avatar
Antoni Baum committed
1325

1326
        if not scheduler_outputs.is_empty():
1327
1328
1329
1330
1331
1332

            # Check if we have a cached last_output from the previous iteration.
            # For supporting PP this is probably the best way to pass the
            # sampled_token_ids, as a separate broadcast over all the PP stages
            # will cause one virtual engine's microbatch to block the pipeline.
            last_sampled_token_ids = \
1333
                self._get_last_sampled_token_ids(virtual_engine)
1334

1335
            execute_model_req = ExecuteModelRequest(
1336
1337
1338
1339
                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,
1340
1341
                num_lookahead_slots=scheduler_outputs.num_lookahead_slots,
                running_queue_size=scheduler_outputs.running_queue_size,
1342
1343
1344
1345
1346
                finished_requests_ids=finished_requests_ids,
                # We use ExecuteModelRequest to pass the last sampled_token_ids
                # to each of the non-last PP stages for in-place prepare_input.
                last_sampled_token_ids=last_sampled_token_ids)

1347
            if allow_async_output_proc:
1348
1349
                execute_model_req.async_callback = self.async_callbacks[
                    virtual_engine]
1350

1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
            try:
                outputs = self.model_executor.execute_model(
                    execute_model_req=execute_model_req)
                self._skip_scheduling_next_step = False
            except InputProcessingError as e:
                # The input for this request cannot be processed, so we must
                # abort it. If there are remaining requests in the batch that
                # have been scheduled, they will be retried on the next step.
                invalid_request_id = e.request_id
                self._abort_and_cache_schedule(
                    request_id=invalid_request_id,
                    virtual_engine=virtual_engine,
                    seq_group_metadata_list=seq_group_metadata_list,
                    scheduler_outputs=scheduler_outputs,
                    allow_async_output_proc=allow_async_output_proc)
                # Raise so the caller is notified that this request failed
                raise
1368

1369
            # We need to do this here so that last step's sampled_token_ids can
1370
1371
            # be passed to the next iteration for PP.
            if self.scheduler_config.is_multi_step:
1372
                self._update_cached_scheduler_output(virtual_engine, outputs)
1373
        else:
1374
1375
            # Nothing scheduled => If there is pending async postprocessor,
            # then finish it here.
1376
1377
            if len(ctx.output_queue) > 0:
                self._process_model_outputs(ctx=ctx)
1378
            # No outputs in this case
1379
            outputs = []
Antoni Baum's avatar
Antoni Baum committed
1380

1381
1382
1383
1384
1385
1386
        # Finish the current step for all the sequence groups.
        if self.scheduler_config.is_multi_step:
            for seq_group in seq_group_metadata_list:
                seq_group.finish_step()

        if not self._has_remaining_steps(seq_group_metadata_list):
1387
            # clear the cache if we have finished all the steps.
1388
1389
1390
            if self.scheduler_config.is_multi_step:
                self.cached_scheduler_outputs[0] = SchedulerOutputState()

1391
1392
1393
1394
1395
1396
            # is_first_step_output is True only when the num_steps of all
            # the sequences are 1. When the num_steps > 1,
            # multi_step_model_runner does the first-step output append.
            is_first_step_output: bool = False if not seq_group_metadata_list \
                else seq_group_metadata_list[0].state.num_steps == 1

1397
            # Add results to the output_queue
1398
1399
1400
1401
            ctx.append_output(outputs=outputs,
                              seq_group_metadata_list=seq_group_metadata_list,
                              scheduler_outputs=scheduler_outputs,
                              is_async=allow_async_output_proc,
1402
1403
                              is_last_step=True,
                              is_first_step_output=is_first_step_output)
1404
1405
1406

            if outputs and allow_async_output_proc:
                assert len(outputs) == 1, (
1407
                    "Async postprocessor expects only a single output set")
1408

1409
                self._advance_to_next_step(
1410
                    outputs[0], seq_group_metadata_list,
1411
                    scheduler_outputs.scheduled_seq_groups)
1412

1413
            # Check if need to run the usual non-async path
1414
            if not allow_async_output_proc:
1415
                self._process_model_outputs(ctx=ctx)
1416

1417
                # Log stats.
1418
                self.do_log_stats(scheduler_outputs, outputs)
1419

1420
1421
1422
                # Tracing
                self.do_tracing(scheduler_outputs)
        else:
1423
            # Multi-step case
1424
            return ctx.request_outputs
1425

1426
        if not self.has_unfinished_requests():
1427
1428
            # Drain async postprocessor (if exists)
            if len(ctx.output_queue) > 0:
1429
                self._process_model_outputs(ctx=ctx)
1430
            assert len(ctx.output_queue) == 0
1431

1432
1433
1434
1435
1436
            # Stop the execute model loop in parallel workers until there are
            # more requests to process. This avoids waiting indefinitely in
            # torch.distributed ops which may otherwise timeout, and unblocks
            # the RPC thread in the workers so that they can process any other
            # queued control plane messages, such as add/remove lora adapters.
1437
            logger.debug("Stopping remote worker execution loop.")
1438
1439
            self.model_executor.stop_remote_worker_execution_loop()

1440
        return ctx.request_outputs
Antoni Baum's avatar
Antoni Baum committed
1441

1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
    def _abort_and_cache_schedule(
            self, request_id: str, virtual_engine: int,
            seq_group_metadata_list: List[SequenceGroupMetadata],
            scheduler_outputs: SchedulerOutputs,
            allow_async_output_proc: bool) -> None:
        """Aborts a single request, and caches the scheduler outputs minus that
        request. This allows the next step to continue processing the remaining
        requests without having to re-run the scheduler."""

        # Abort the request and remove its sequence group from the current
        # schedule
        self.abort_request(request_id)
        for i, metadata in enumerate(seq_group_metadata_list):
            if metadata.request_id == request_id:
                del seq_group_metadata_list[i]
                break
        for i, group in enumerate(scheduler_outputs.scheduled_seq_groups):
            if group.seq_group.request_id == request_id:
                del scheduler_outputs.scheduled_seq_groups[i]
                break

        # If there are still other sequence groups left in the schedule, cache
        # them and flag the engine to reuse the schedule.
        if len(seq_group_metadata_list) > 0:
            self._skip_scheduling_next_step = True
            # Reuse multi-step caching logic
            self._cache_scheduler_outputs_for_multi_step(
                virtual_engine=virtual_engine,
                scheduler_outputs=scheduler_outputs,
                seq_group_metadata_list=seq_group_metadata_list,
                allow_async_output_proc=allow_async_output_proc)

1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
    def _has_remaining_steps(
        self, seq_group_metadata_list: Optional[List[SequenceGroupMetadata]]
    ) -> bool:
        if (not self.scheduler_config.is_multi_step
                or not seq_group_metadata_list):
            return False

        # TODO(will) this is a sanity check for nowto make sure that all the
        # seqs are on the same steps. Eventually we will want to do some sort of
        # dynamic scheduling when doing multi-step decoding.
        ref_remaining_steps = seq_group_metadata_list[0].state.remaining_steps
        if any([
                seq_group.state.remaining_steps != ref_remaining_steps
                for seq_group in seq_group_metadata_list[1:]
        ]):
1489
1490
            raise AssertionError("All running sequence groups should "
                                 "have the same remaining steps.")
1491
1492
1493
1494
1495
1496

        return ref_remaining_steps > 0

    def _cache_scheduler_outputs_for_multi_step(
            self, virtual_engine: int,
            seq_group_metadata_list: Optional[List[SequenceGroupMetadata]],
1497
1498
1499
1500
1501
1502
1503
1504
            scheduler_outputs: SchedulerOutputs,
            allow_async_output_proc: bool) -> None:
        co = self.cached_scheduler_outputs[virtual_engine]

        co.seq_group_metadata_list = seq_group_metadata_list
        co.scheduler_outputs = scheduler_outputs
        co.allow_async_output_proc = allow_async_output_proc
        co.last_output = None
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529

    def _update_cached_scheduler_output(
            self, virtual_engine: int,
            output: List[Optional[SamplerOutput]]) -> None:
        if (self.parallel_config.pipeline_parallel_size > 1 and len(output) > 0
                and output[0] is not None):
            last_output = output[-1]
            assert last_output is not None
            assert last_output.sampled_token_ids_cpu is not None
            assert last_output.sampled_token_ids is None
            assert last_output.sampled_token_probs is None
            self.cached_scheduler_outputs[
                virtual_engine].last_output = last_output

    def _get_last_sampled_token_ids(
            self, virtual_engine: int) -> Optional[torch.Tensor]:
        cached_last_output = self.cached_scheduler_outputs[
            virtual_engine].last_output
        if (self.scheduler_config.is_multi_step
                and self.parallel_config.pipeline_parallel_size > 1
                and cached_last_output is not None
                and cached_last_output.sampled_token_ids_cpu is not None):
            return cached_last_output.sampled_token_ids_cpu
        return None

1530
    def add_logger(self, logger_name: str, logger: StatLoggerBase) -> None:
1531
1532
1533
1534
        if not self.log_stats:
            raise RuntimeError(
                "Stat logging is disabled. Set `disable_log_stats=False` "
                "argument to enable.")
1535
1536
1537
1538
1539
        if logger_name in self.stat_loggers:
            raise KeyError(f"Logger with name {logger_name} already exists.")
        self.stat_loggers[logger_name] = logger

    def remove_logger(self, logger_name: str) -> None:
1540
1541
1542
1543
        if not self.log_stats:
            raise RuntimeError(
                "Stat logging is disabled. Set `disable_log_stats=False` "
                "argument to enable.")
1544
1545
1546
1547
        if logger_name not in self.stat_loggers:
            raise KeyError(f"Logger with name {logger_name} does not exist.")
        del self.stat_loggers[logger_name]

1548
1549
1550
    def do_log_stats(self,
                     scheduler_outputs: Optional[SchedulerOutputs] = None,
                     model_output: Optional[List[SamplerOutput]] = None,
1551
1552
                     finished_before: Optional[List[int]] = None,
                     skip: Optional[List[int]] = None) -> None:
1553
1554
        """Forced log when no requests active."""
        if self.log_stats:
1555
            stats = self._get_stats(scheduler_outputs, model_output,
1556
                                    finished_before, skip)
1557
            for logger in self.stat_loggers.values():
1558
                logger.log(stats)
1559

1560
1561
1562
    def _get_stats(self,
                   scheduler_outputs: Optional[SchedulerOutputs],
                   model_output: Optional[List[SamplerOutput]] = None,
1563
1564
                   finished_before: Optional[List[int]] = None,
                   skip: Optional[List[int]] = None) -> Stats:
1565
1566
1567
1568
1569
1570
1571
        """Get Stats to be Logged to Prometheus.

        Args:
            scheduler_outputs: Optional, used to populate metrics related to
                the scheduled batch,
            model_output: Optional, used to emit speculative decoding metrics
                which are created by the workers.
1572
1573
1574
1575
            finished_before: Optional, indices of sequences that were finished
                before. These sequences will be ignored.
            skip: Optional, indices of sequences that were preempted. These
                sequences will be ignored.
1576
        """
1577
        now = time.time()
Woosuk Kwon's avatar
Woosuk Kwon committed
1578

1579
1580
        # System State
        #   Scheduler State
1581
1582
1583
1584
1585
1586
        num_running_sys = sum(
            len(scheduler.running) for scheduler in self.scheduler)
        num_swapped_sys = sum(
            len(scheduler.swapped) for scheduler in self.scheduler)
        num_waiting_sys = sum(
            len(scheduler.waiting) for scheduler in self.scheduler)
1587
1588

        # KV Cache Usage in %
1589
        num_total_gpu = self.cache_config.num_gpu_blocks
1590
        gpu_cache_usage_sys = 0.
1591
        if num_total_gpu:  # Guard against both None and 0
1592
1593
1594
            num_free_gpu = sum(
                scheduler.block_manager.get_num_free_gpu_blocks()
                for scheduler in self.scheduler)
1595
            gpu_cache_usage_sys = 1.0 - (num_free_gpu / num_total_gpu)
Woosuk Kwon's avatar
Woosuk Kwon committed
1596

1597
        num_total_cpu = self.cache_config.num_cpu_blocks
1598
        cpu_cache_usage_sys = 0.
1599
        if num_total_cpu:  # Guard against both None and 0
1600
1601
1602
            num_free_cpu = sum(
                scheduler.block_manager.get_num_free_cpu_blocks()
                for scheduler in self.scheduler)
1603
1604
            cpu_cache_usage_sys = 1.0 - (num_free_cpu / num_total_cpu)

1605
1606
1607
1608
1609
1610
1611
        # Prefix Cache Hit Rate. Note that we always use
        # the cache hit rate of the first virtual engine.
        cpu_prefix_cache_hit_rate = self.scheduler[
            0].get_prefix_cache_hit_rate(Device.CPU)
        gpu_prefix_cache_hit_rate = self.scheduler[
            0].get_prefix_cache_hit_rate(Device.GPU)

1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
        # Exchange the uasge and cache hit stats between gpu and cpu when
        # running on cpu because the cpu_worker.py intentionally reports the
        # number of cpu blocks as gpu blocks in favor of cache management.
        if self.device_config.device_type == "cpu":
            num_total_gpu, num_total_cpu = num_total_cpu, num_total_gpu
            gpu_cache_usage_sys, cpu_cache_usage_sys = (
                cpu_cache_usage_sys,
                gpu_cache_usage_sys,
            )
            gpu_prefix_cache_hit_rate, cpu_prefix_cache_hit_rate = (
                cpu_prefix_cache_hit_rate,
                gpu_prefix_cache_hit_rate,
            )

1626
1627
1628
        # Iteration stats
        num_prompt_tokens_iter = 0
        num_generation_tokens_iter = 0
harrywu's avatar
harrywu committed
1629
        num_tokens_iter = 0
1630
1631
        time_to_first_tokens_iter: List[float] = []
        time_per_output_tokens_iter: List[float] = []
1632
1633
        num_preemption_iter = (0 if scheduler_outputs is None else
                               scheduler_outputs.preempted)
1634
1635
1636
1637

        # Request stats
        #   Latency
        time_e2e_requests: List[float] = []
harrywu's avatar
harrywu committed
1638
1639
1640
1641
        time_queue_requests: List[float] = []
        time_inference_requests: List[float] = []
        time_prefill_requests: List[float] = []
        time_decode_requests: List[float] = []
1642
1643
1644
1645
        #   Metadata
        num_prompt_tokens_requests: List[int] = []
        num_generation_tokens_requests: List[int] = []
        n_requests: List[int] = []
harrywu's avatar
harrywu committed
1646
        max_num_generation_tokens_requests: List[int] = []
1647
        max_tokens_requests: List[int] = []
1648
1649
        finished_reason_requests: List[str] = []

1650
        # LoRA requests
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
        running_lora_adapters = dict(
            collectionsCounter([
                running_request.lora_request.lora_name
                for scheduler in self.scheduler
                for running_request in scheduler.running
                if running_request.lora_request
            ]))
        waiting_lora_adapters = dict(
            collectionsCounter([
                waiting_request.lora_request.lora_name
                for scheduler in self.scheduler
                for waiting_request in scheduler.waiting
                if waiting_request.lora_request
            ]))
        max_lora_stat = "0"
        if self.lora_config:
            max_lora_stat = str(self.lora_config.max_loras)

1669
1670
        # NOTE: This loop assumes prefill seq_groups are before
        # decode seq_groups in scheduled_seq_groups.
1671
        if scheduler_outputs is not None:
1672
1673
1674
1675
            # For async postprocessor, already finished sequences need to be
            # not counted (to avoid double counting)
            actual_num_batched_tokens = scheduler_outputs.num_batched_tokens  # type: ignore

1676
            num_generation_tokens_from_prefill_groups = 0
1677
1678
1679
1680
            # NOTE: if scheduler_outputs.num_prefill_groups > 0 and
            # the len of scheduler_outputs.scheduled_seq_groups is !=
            # scheduler_outputs.num_prefill_groups, this means that
            # chunked prefills have been detected.
1681
1682
1683

            for idx, scheduled_seq_group in enumerate(
                    scheduler_outputs.scheduled_seq_groups):
1684
1685
1686
1687
                # Skip double logging when using async output proc
                if finished_before and idx in finished_before:
                    actual_num_batched_tokens -= 1
                    continue
1688
1689
1690
1691
1692

                # Currently, skip == preempted sequences, so we need to skip
                # their log stats
                if skip and idx in skip:
                    continue
1693

1694
                group_was_prefill = idx < scheduler_outputs.num_prefill_groups
1695
                seq_group = scheduled_seq_group.seq_group
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707

                # NOTE: a seq_group that completed all of its prefill tokens
                # in the last iteration will have seq_group.is_prefill() = False
                # with group_was_prefill = True
                if group_was_prefill:
                    # Number of prompt tokens.
                    num_prompt_tokens_iter += (
                        scheduled_seq_group.token_chunk_size)

                    # If the seq_group just finished the prefill state
                    # get TTFT.
                    if not seq_group.is_prefill():
1708
                        latency = seq_group.get_last_token_latency()
1709
1710
1711
1712
1713
1714
1715
                        time_to_first_tokens_iter.append(latency)

                        # One generation token per finished prefill.
                        num_generation_tokens_from_prefill_groups += (
                            seq_group.num_seqs())
                else:
                    # TPOTs.
1716
                    latency = seq_group.get_last_token_latency()
1717
                    time_per_output_tokens_iter.append(latency)
1718
1719
1720
1721
1722
1723
1724
1725
1726
                    if seq_group.state.current_step == 0:
                        # For async_output_proc, the do_log_stats()
                        # is called following init_multi_step(), which
                        # sets the current_step to zero.
                        actual_num_batched_tokens +=\
                            seq_group.state.num_steps - 1
                    else:
                        actual_num_batched_tokens +=\
                            seq_group.state.current_step - 1
1727
1728
1729
1730
1731
1732

                # Because of chunked prefill, we can have a single sequence
                # group that does multiple prompt_runs. To prevent logging
                # the same metadata more than once per request, we standardize
                # on logging request level information for finished requests,
                # which can only happen once.
1733
                if seq_group.is_finished():
1734
                    # Latency timings
1735
1736
                    time_e2e_requests.append(now -
                                             seq_group.metrics.arrival_time)
harrywu's avatar
harrywu committed
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
                    if (seq_group.metrics.first_scheduled_time is not None and
                            seq_group.metrics.first_token_time is not None):
                        time_queue_requests.append(
                            seq_group.metrics.first_scheduled_time -
                            seq_group.metrics.arrival_time)
                        time_prefill_requests.append(
                            seq_group.metrics.first_token_time -
                            seq_group.metrics.first_scheduled_time)
                        time_decode_requests.append(
                            now - seq_group.metrics.first_token_time)
                        time_inference_requests.append(
                            now - seq_group.metrics.first_scheduled_time)
1749
1750
1751
1752
1753
1754
1755
                    # Metadata
                    num_prompt_tokens_requests.append(
                        len(seq_group.prompt_token_ids))
                    num_generation_tokens_requests.extend([
                        seq.get_output_len()
                        for seq in seq_group.get_finished_seqs()
                    ])
harrywu's avatar
harrywu committed
1756
1757
1758
                    max_num_generation_tokens_requests.append(
                        max(seq.get_output_len()
                            for seq in seq_group.get_seqs()))
1759
1760
                    if seq_group.sampling_params is not None:
                        n_requests.append(seq_group.sampling_params.n)
1761
1762
                        max_tokens_requests.append(
                            seq_group.sampling_params.max_tokens)
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
                    finished_reason_requests.extend([
                        SequenceStatus.get_finished_reason(seq.status)
                        for seq in seq_group.get_finished_seqs()
                    ])

            # Number of generation tokens.
            #   num_batched_tokens equals the number of prompt_tokens plus the
            #   number of decode_tokens in a single iteration. So,
            #   num_generation_tokens = num_batched_tokens - num_prompt_tokens
            #   + num_generation_tokens_from_prefill_groups (since we generate
            #   one token on prefills on iters where the prefill finishes).
            num_generation_tokens_iter = (
1775
                actual_num_batched_tokens - num_prompt_tokens_iter +
1776
                num_generation_tokens_from_prefill_groups)
harrywu's avatar
harrywu committed
1777
1778
            num_tokens_iter = (num_generation_tokens_iter +
                               num_prompt_tokens_iter)
1779
1780
        # Spec decode, if enabled, emits specialized metrics from the worker in
        # sampler output.
1781
1782
        if model_output and isinstance(model_output[0], SamplerOutput) and (
                model_output[0].spec_decode_worker_metrics is not None):
1783
1784
1785
1786
            spec_decode_metrics = model_output[0].spec_decode_worker_metrics
        else:
            spec_decode_metrics = None

1787
1788
        return Stats(
            now=now,
1789
1790
1791
1792
1793
1794
1795
1796
            # System stats
            #   Scheduler State
            num_running_sys=num_running_sys,
            num_swapped_sys=num_swapped_sys,
            num_waiting_sys=num_waiting_sys,
            #   KV Cache Usage in %
            gpu_cache_usage_sys=gpu_cache_usage_sys,
            cpu_cache_usage_sys=cpu_cache_usage_sys,
1797
1798
1799
            #   Prefix Cache Hit Rate
            cpu_prefix_cache_hit_rate=cpu_prefix_cache_hit_rate,
            gpu_prefix_cache_hit_rate=gpu_prefix_cache_hit_rate,
1800
1801
1802
1803

            # Iteration stats
            num_prompt_tokens_iter=num_prompt_tokens_iter,
            num_generation_tokens_iter=num_generation_tokens_iter,
harrywu's avatar
harrywu committed
1804
            num_tokens_iter=num_tokens_iter,
1805
1806
            time_to_first_tokens_iter=time_to_first_tokens_iter,
            time_per_output_tokens_iter=time_per_output_tokens_iter,
1807
            spec_decode_metrics=spec_decode_metrics,
1808
            num_preemption_iter=num_preemption_iter,
1809
1810
1811
1812

            # Request stats
            #   Latency
            time_e2e_requests=time_e2e_requests,
harrywu's avatar
harrywu committed
1813
1814
1815
1816
            time_queue_requests=time_queue_requests,
            time_inference_requests=time_inference_requests,
            time_prefill_requests=time_prefill_requests,
            time_decode_requests=time_decode_requests,
1817
1818
1819
            #   Metadata
            num_prompt_tokens_requests=num_prompt_tokens_requests,
            num_generation_tokens_requests=num_generation_tokens_requests,
harrywu's avatar
harrywu committed
1820
1821
            max_num_generation_tokens_requests=
            max_num_generation_tokens_requests,
1822
            n_requests=n_requests,
1823
            max_tokens_requests=max_tokens_requests,
1824
            finished_reason_requests=finished_reason_requests,
1825
1826
1827
            max_lora=str(max_lora_stat),
            waiting_lora_adapters=list(waiting_lora_adapters.keys()),
            running_lora_adapters=list(running_lora_adapters.keys()))
1828

1829
    def add_lora(self, lora_request: LoRARequest) -> bool:
1830
        return self.model_executor.add_lora(lora_request)
1831
1832

    def remove_lora(self, lora_id: int) -> bool:
1833
        return self.model_executor.remove_lora(lora_id)
1834

1835
    def list_loras(self) -> Set[int]:
1836
        return self.model_executor.list_loras()
1837

1838
1839
1840
    def pin_lora(self, lora_id: int) -> bool:
        return self.model_executor.pin_lora(lora_id)

1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
    def add_prompt_adapter(
            self, prompt_adapter_request: PromptAdapterRequest) -> bool:
        return self.model_executor.add_prompt_adapter(prompt_adapter_request)

    def remove_prompt_adapter(self, prompt_adapter_id: int) -> bool:
        return self.model_executor.remove_prompt_adapter(prompt_adapter_id)

    def list_prompt_adapters(self) -> List[int]:
        return self.model_executor.list_prompt_adapters()

1851
1852
1853
1854
1855
1856
    def start_profile(self) -> None:
        self.model_executor.start_profile()

    def stop_profile(self) -> None:
        self.model_executor.stop_profile()

1857
1858
1859
1860
1861
    def sleep(self, level: int = 1) -> None:
        assert self.vllm_config.model_config.enable_sleep_mode, (
            "Sleep mode is not enabled in the model config")
        self.model_executor.sleep(level=level)

1862
    def wake_up(self, tags: Optional[list[str]] = None) -> None:
1863
1864
        assert self.vllm_config.model_config.enable_sleep_mode, (
            "Sleep mode is not enabled in the model config")
1865
        self.model_executor.wake_up(tags)
1866

1867
1868
1869
    def is_sleeping(self) -> bool:
        return self.model_executor.is_sleeping

1870
    def check_health(self) -> None:
1871
        self.model_executor.check_health()
1872
1873
1874
1875

    def is_tracing_enabled(self) -> bool:
        return self.tracer is not None

1876
1877
1878
    def do_tracing(self,
                   scheduler_outputs: SchedulerOutputs,
                   finished_before: Optional[List[int]] = None) -> None:
1879
1880
1881
        if self.tracer is None:
            return

1882
1883
1884
1885
1886
1887
        for idx, scheduled_seq_group in enumerate(
                scheduler_outputs.scheduled_seq_groups):
            # Skip double tracing when using async output proc
            if finished_before and idx in finished_before:
                continue

1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
            seq_group = scheduled_seq_group.seq_group
            if seq_group.is_finished():
                self.create_trace_span(seq_group)

    def create_trace_span(self, seq_group: SequenceGroup) -> None:
        if self.tracer is None or seq_group.sampling_params is None:
            return
        arrival_time_nano_seconds = int(seq_group.metrics.arrival_time * 1e9)

        trace_context = extract_trace_context(seq_group.trace_headers)

        with self.tracer.start_as_current_span(
                "llm_request",
                kind=SpanKind.SERVER,
                context=trace_context,
                start_time=arrival_time_nano_seconds) as seq_span:
            metrics = seq_group.metrics
            ttft = metrics.first_token_time - metrics.arrival_time
            e2e_time = metrics.finished_time - metrics.arrival_time
1907
            seq_span.set_attribute(SpanAttributes.GEN_AI_RESPONSE_MODEL,
1908
                                   self.model_config.model)
1909
            seq_span.set_attribute(SpanAttributes.GEN_AI_REQUEST_ID,
1910
                                   seq_group.request_id)
1911
            seq_span.set_attribute(SpanAttributes.GEN_AI_REQUEST_TEMPERATURE,
1912
                                   seq_group.sampling_params.temperature)
1913
            seq_span.set_attribute(SpanAttributes.GEN_AI_REQUEST_TOP_P,
1914
                                   seq_group.sampling_params.top_p)
1915
            seq_span.set_attribute(SpanAttributes.GEN_AI_REQUEST_MAX_TOKENS,
1916
                                   seq_group.sampling_params.max_tokens)
1917
            seq_span.set_attribute(SpanAttributes.GEN_AI_REQUEST_N,
1918
                                   seq_group.sampling_params.n)
1919
            seq_span.set_attribute(SpanAttributes.GEN_AI_USAGE_NUM_SEQUENCES,
1920
                                   seq_group.num_seqs())
1921
            seq_span.set_attribute(SpanAttributes.GEN_AI_USAGE_PROMPT_TOKENS,
1922
1923
                                   len(seq_group.prompt_token_ids))
            seq_span.set_attribute(
1924
                SpanAttributes.GEN_AI_USAGE_COMPLETION_TOKENS,
1925
1926
1927
1928
                sum([
                    seq.get_output_len()
                    for seq in seq_group.get_finished_seqs()
                ]))
1929
            seq_span.set_attribute(SpanAttributes.GEN_AI_LATENCY_TIME_IN_QUEUE,
1930
1931
                                   metrics.time_in_queue)
            seq_span.set_attribute(
1932
1933
                SpanAttributes.GEN_AI_LATENCY_TIME_TO_FIRST_TOKEN, ttft)
            seq_span.set_attribute(SpanAttributes.GEN_AI_LATENCY_E2E, e2e_time)
1934
1935
            if metrics.scheduler_time is not None:
                seq_span.set_attribute(
1936
                    SpanAttributes.GEN_AI_LATENCY_TIME_IN_SCHEDULER,
1937
1938
1939
                    metrics.scheduler_time)
            if metrics.model_forward_time is not None:
                seq_span.set_attribute(
1940
                    SpanAttributes.GEN_AI_LATENCY_TIME_IN_MODEL_FORWARD,
1941
1942
1943
                    metrics.model_forward_time / 1000.0)
            if metrics.model_execute_time is not None:
                seq_span.set_attribute(
1944
                    SpanAttributes.GEN_AI_LATENCY_TIME_IN_MODEL_EXECUTE,
1945
                    metrics.model_execute_time)
1946

1947
    def _validate_model_inputs(self, inputs: ProcessorInputs,
1948
                               lora_request: Optional[LoRARequest]):
1949
1950
        encoder_inputs, decoder_inputs = split_enc_dec_inputs(inputs)

1951
1952
1953
1954
        if encoder_inputs is not None:
            self._validate_model_input(encoder_inputs,
                                       lora_request,
                                       prompt_type="encoder")
1955

1956
1957
1958
        self._validate_model_input(decoder_inputs,
                                   lora_request,
                                   prompt_type="decoder")
1959

1960
1961
1962
1963
1964
1965
1966
    def _validate_model_input(
        self,
        prompt_inputs: SingletonInputs,
        lora_request: Optional[LoRARequest],
        *,
        prompt_type: Literal["encoder", "decoder"],
    ):
1967
1968
1969
        model_config = self.model_config
        tokenizer = (None if self.tokenizer is None else
                     self.tokenizer.get_lora_tokenizer(lora_request))
1970

1971
        prompt_ids = prompt_inputs.get("prompt_token_ids", [])
1972
1973
1974
        if not prompt_ids:
            if prompt_type == "encoder" and model_config.is_multimodal_model:
                pass  # Mllama may have empty encoder inputs for text-only data
1975
            elif prompt_inputs["type"] == "embeds":
1976
                pass
1977
1978
1979
            else:
                raise ValueError(f"The {prompt_type} prompt cannot be empty")

1980
1981
1982
1983
1984
1985
        if tokenizer is not None:
            max_input_id = max(prompt_ids, default=0)
            if max_input_id > tokenizer.max_token_id:
                raise ValueError(
                    f"Token id {max_input_id} is out of vocabulary")

1986
        max_prompt_len = self.model_config.max_model_len
1987
        if len(prompt_ids) > max_prompt_len:
1988
            if prompt_type == "encoder" and model_config.is_multimodal_model:
1989
1990
                mm_registry = self.input_preprocessor.mm_registry
                mm_processor = mm_registry.create_processor(
1991
1992
1993
                    model_config,
                    tokenizer=tokenizer or object(),  # Dummy if no tokenizer
                )
1994
                assert isinstance(mm_processor, EncDecMultiModalProcessor)
1995

1996
1997
1998
                if mm_processor.pad_dummy_encoder_prompt:
                    return  # Skip encoder length check for Whisper

1999
            if model_config.is_multimodal_model:
2000
                suggestion = (
2001
2002
2003
2004
                    "Make sure that `max_model_len` is no smaller than the "
                    "number of text tokens plus multimodal tokens. For image "
                    "inputs, the number of image tokens depends on the number "
                    "of images, and possibly their aspect ratios as well.")
2005
2006
2007
2008
2009
2010
2011
2012
2013
            else:
                suggestion = (
                    "Make sure that `max_model_len` is no smaller than the "
                    "number of text tokens.")

            raise ValueError(
                f"The {prompt_type} prompt (length {len(prompt_ids)}) is "
                f"longer than the maximum model length of {max_prompt_len}. "
                f"{suggestion}")
2014
2015
2016
2017

            # TODO: Find out how many placeholder tokens are there so we can
            # check that chunked prefill does not truncate them
            # max_batch_len = self.scheduler_config.max_num_batched_tokens
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027

    def _build_logits_processors(
            self, sampling_params: SamplingParams,
            lora_request: Optional[LoRARequest]) -> SamplingParams:
        """Constructs logits processors based on the guided_decoding,
        logits_bias, and allowed_token_ids fields in sampling_params. Deletes
        those fields and adds the constructed logits processors to the
        logits_processors field. Returns the modified sampling params."""

        logits_processors = []
2028

2029
2030
2031
2032
2033
        if sampling_params.guided_decoding is not None:
            # Defensively copy sampling params since guided decoding logits
            # processors can have different state for each request
            sampling_params = copy.copy(sampling_params)
            guided_decoding = sampling_params.guided_decoding
2034
2035
2036
2037
2038
2039
2040

            logger.debug(
                "Building guided decoding logits processor in "
                "LLMEngine. Params: %s", guided_decoding)

            tokenizer = self.get_tokenizer(lora_request=lora_request)
            guided_decoding.backend = guided_decoding.backend or \
2041
                self.decoding_config.backend
2042

2043
            if self.decoding_config.reasoning_backend:
2044
2045
                logger.debug("Building with reasoning backend %s",
                             self.decoding_config.reasoning_backend)
2046

2047
            processor = get_local_guided_decoding_logits_processor(
2048
2049
                guided_params=guided_decoding,
                tokenizer=tokenizer,
2050
2051
2052
                model_config=self.model_config,
                reasoning_backend=self.decoding_config.reasoning_backend,
            )
2053
2054
2055
2056
2057
2058
2059
2060
2061
            if processor:
                logits_processors.append(processor)

            # Unset so this doesn't get passed down to the model
            sampling_params.guided_decoding = None

        if (sampling_params.logit_bias or sampling_params.allowed_token_ids):
            tokenizer = self.get_tokenizer(lora_request=lora_request)

2062
            processors = get_openai_logits_processors(
2063
2064
2065
2066
2067
2068
2069
2070
2071
                logit_bias=sampling_params.logit_bias,
                allowed_token_ids=sampling_params.allowed_token_ids,
                tokenizer=tokenizer)
            logits_processors.extend(processors)

            # Unset so these don't get passed down to the model
            sampling_params.logit_bias = None
            sampling_params.allowed_token_ids = None

2072
2073
2074
2075
2076
2077
        if len(sampling_params.bad_words) > 0:
            tokenizer = self.get_tokenizer(lora_request)
            processors = get_bad_words_logits_processors(
                bad_words=sampling_params.bad_words, tokenizer=tokenizer)
            logits_processors.extend(processors)

2078
2079
2080
2081
2082
2083
2084
        if logits_processors:
            if sampling_params.logits_processors is None:
                sampling_params.logits_processors = logits_processors
            else:
                sampling_params.logits_processors.extend(logits_processors)

        return sampling_params
2085

2086
2087
2088
2089
2090
2091
2092
2093
    def collective_rpc(self,
                       method: Union[str, Callable[..., _R]],
                       timeout: Optional[float] = None,
                       args: tuple = (),
                       kwargs: Optional[dict[str, Any]] = None) -> list[_R]:
        return self.model_executor.collective_rpc(method, timeout, args,
                                                  kwargs)

2094

2095
2096
2097
if envs.is_set("VLLM_USE_V1") and envs.VLLM_USE_V1:
    from vllm.v1.engine.llm_engine import LLMEngine as V1LLMEngine
    LLMEngine = V1LLMEngine  # type: ignore