llm_engine.py 90 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import os
5
import copy
Antoni Baum's avatar
Antoni Baum committed
6
import time
7
from collections import Counter as collectionsCounter
8
from collections import deque
9
from contextlib import contextmanager
10
from dataclasses import dataclass
11
from functools import partial
12
from typing import (TYPE_CHECKING, Any, Callable, ClassVar, Deque, Dict,
13
                    Iterable, List, Literal, Mapping, NamedTuple, Optional)
14
from typing import Sequence as GenericSequence
15
from typing import Set, Type, Union, cast
16

17
import torch
18
from typing_extensions import TypeVar
19

20
import vllm.envs as envs
21
22
23
from vllm.config import (DecodingConfig, LoRAConfig, ModelConfig,
                         ObservabilityConfig, ParallelConfig, SchedulerConfig,
                         VllmConfig)
24
from vllm.core.scheduler import ScheduledSequenceGroup, SchedulerOutputs
Woosuk Kwon's avatar
Woosuk Kwon committed
25
from vllm.engine.arg_utils import EngineArgs
26
from vllm.engine.metrics_types import StatLoggerBase, Stats
27
28
29
30
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
31
32
from vllm.entrypoints.openai.logits_processors import (
    get_logits_processors as get_openai_logits_processors)
33
from vllm.executor.executor_base import ExecutorBase
34
from vllm.inputs import ProcessorInputs, PromptType, SingletonInputs
35
from vllm.inputs.parse import split_enc_dec_inputs
36
from vllm.inputs.preprocess import InputPreprocessor
Woosuk Kwon's avatar
Woosuk Kwon committed
37
from vllm.logger import init_logger
38
from vllm.logits_process import get_bad_words_logits_processors
39
from vllm.lora.request import LoRARequest
40
41
from vllm.model_executor.guided_decoding import (
    get_local_guided_decoding_logits_processor)
42
from vllm.model_executor.layers.sampler import SamplerOutput
43
from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalRegistry
44
from vllm.multimodal.processing import EncDecMultiModalProcessor
45
from vllm.outputs import (PoolingRequestOutput, RequestOutput,
46
47
                          RequestOutputFactory)
from vllm.pooling_params import PoolingParams
48
from vllm.sampling_params import RequestOutputKind, SamplingParams
49
50
51
from vllm.sequence import (ExecuteModelRequest, ParallelSampleSequenceGroup,
                           PoolingSequenceGroupOutput, Sequence, SequenceGroup,
                           SequenceGroupBase, SequenceGroupMetadata,
zhuwenwen's avatar
zhuwenwen committed
52
                           SequenceGroupOutput, SequenceStatus, CompletionSequenceGroupOutput, VLLM_INVALID_TOKEN_ID)
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
58
# DEBUG add cpm tokenizer
from vllm.transformers_utils.tokenizers import CPM9GTokenizer
59
from vllm.transformers_utils.tokenizer_group import (
60
    TokenizerGroup, init_tokenizer_from_configs)
yhu422's avatar
yhu422 committed
61
62
from vllm.usage.usage_lib import (UsageContext, is_usage_stats_enabled,
                                  usage_message)
63
from vllm.utils import Counter, Device, resolve_obj_by_qualname, weak_bind
64
from vllm.version import __version__ as VLLM_VERSION
65
from vllm.worker.model_runner_base import InputProcessingError
66
from vllm.profiler.prof import profile
67
68

logger = init_logger(__name__)
69
_LOCAL_LOGGING_INTERVAL_SEC = 5
Woosuk Kwon's avatar
Woosuk Kwon committed
70

71
_O = TypeVar("_O", RequestOutput, PoolingRequestOutput)
72
_R = TypeVar("_R", default=Any)
73
74


75
76
77
78
79
@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
80
81
    allow_async_output_proc: bool = False
    last_output: Optional[SamplerOutput] = None
82
83


84
85
86
87
88
89
class OutputData(NamedTuple):
    outputs: List[SamplerOutput]
    seq_group_metadata_list: List[SequenceGroupMetadata]
    scheduler_outputs: SchedulerOutputs
    is_async: bool
    is_last_step: bool
90
91
92
93
94
95
    # 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]
96
97
98
    skip: List[int]


99
class SchedulerContext:
100

101
    def __init__(self, multi_step_stream_outputs: bool = False):
102
103
        self.output_queue: Deque[OutputData] = deque()
        self.request_outputs: List[Union[RequestOutput,
104
                                         PoolingRequestOutput]] = []
105
106
107
108
        self.seq_group_metadata_list: Optional[
            List[SequenceGroupMetadata]] = None
        self.scheduler_outputs: Optional[SchedulerOutputs] = None

109
110
        self.multi_step_stream_outputs: bool = multi_step_stream_outputs

111
112
113
    def append_output(self, outputs: List[SamplerOutput],
                      seq_group_metadata_list: List[SequenceGroupMetadata],
                      scheduler_outputs: SchedulerOutputs, is_async: bool,
114
115
                      is_last_step: bool,
                      is_first_step_output: Optional[bool]):
116
117
118
119
120
121
        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,
122
                       is_first_step_output=is_first_step_output,
123
                       skip=[]))
124
125


126
class LLMEngine:
Zhuohan Li's avatar
Zhuohan Li committed
127
    """An LLM engine that receives requests and generates texts.
128

Woosuk Kwon's avatar
Woosuk Kwon committed
129
    This is the main class for the vLLM engine. It receives requests
130
131
132
133
134
135
    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.

136
137
138
    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.
139

140
    The config arguments are derived from [`EngineArgs`][vllm.EngineArgs].
141
142

    Args:
143
        vllm_config: The configuration for initializing and running vLLM.
144
145
        executor_class: The model executor class for managing distributed
            execution.
146
        log_stats: Whether to log statistics.
147
        usage_context: Specified entry point, used for usage info collection.
148
    """
149

150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
    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)}")

175
        return cast(_O, output)
176
177
178

    @classmethod
    def validate_outputs(
zhuwenwen's avatar
zhuwenwen committed
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
        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

197
198
        return outputs_

199
    tokenizer: Optional[TokenizerGroup]
200

201
202
    def __init__(
        self,
203
        vllm_config: VllmConfig,
204
        executor_class: Type[ExecutorBase],
205
        log_stats: bool,
yhu422's avatar
yhu422 committed
206
        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
207
        stat_loggers: Optional[Dict[str, StatLoggerBase]] = None,
208
        mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
209
        use_cached_outputs: bool = False,
210
    ) -> None:
211
212
213
214
215
216
        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.")
217

218
        self.vllm_config = vllm_config
219
220
221
222
223
224
225
226
227
        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
228
        )
229
        self.observability_config = vllm_config.observability_config or ObservabilityConfig(  # noqa
230
231
        )

232
        logger.info(
233
            "Initializing a V0 LLM engine (v%s) with config: %s, "
234
            "use_cached_outputs=%s, ",
235
            VLLM_VERSION,
236
            vllm_config,
237
            use_cached_outputs,
238
        )
239

240
        self.log_stats = log_stats
241
        self.use_cached_outputs = use_cached_outputs
242

243
        if self.model_config.skip_tokenizer_init:
244
            self.tokenizer = None
245
            self.detokenizer = None
246
            tokenizer_group = None
247
248
249
250
        elif self.model_config.tokenizer_mode == "cpm":
            self.tokenizer = CPM9GTokenizer(self.model_config.model, trust_remote_code=True)
            self.detokenizer = Detokenizer(self.tokenizer, self.model_config.tokenizer_mode) 
            tokenizer_group = self.get_tokenizer_group()
251
        else:
252
253
254
            self.tokenizer = self._init_tokenizer()
            self.detokenizer = Detokenizer(self.tokenizer)
            tokenizer_group = self.get_tokenizer_group()
255
256
257
258
259
260
261

        # 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)
262

263
        self.seq_counter = Counter()
264
265
        self.generation_config_fields = (
            self.model_config.try_get_generation_config())
266

267
        self.input_preprocessor = InputPreprocessor(self.model_config,
268
269
                                                    self.tokenizer,
                                                    mm_registry)
270

271
        self.model_executor = executor_class(vllm_config=vllm_config)
272

273
        if self.model_config.runner_type != "pooling":
274
            self._initialize_kv_caches()
275

yhu422's avatar
yhu422 committed
276
277
        # If usage stat is enabled, collect relevant info.
        if is_usage_stats_enabled():
278
279
            from vllm.model_executor.model_loader import (
                get_architecture_class_name)
yhu422's avatar
yhu422 committed
280
            usage_message.report_usage(
281
                get_architecture_class_name(self.model_config),
yhu422's avatar
yhu422 committed
282
283
284
285
                usage_context,
                extra_kvs={
                    # Common configuration
                    "dtype":
286
                    str(self.model_config.dtype),
yhu422's avatar
yhu422 committed
287
                    "tensor_parallel_size":
288
                    self.parallel_config.tensor_parallel_size,
yhu422's avatar
yhu422 committed
289
                    "block_size":
290
                    self.cache_config.block_size,
yhu422's avatar
yhu422 committed
291
                    "gpu_memory_utilization":
292
                    self.cache_config.gpu_memory_utilization,
yhu422's avatar
yhu422 committed
293
294
295

                    # Quantization
                    "quantization":
296
                    self.model_config.quantization,
yhu422's avatar
yhu422 committed
297
                    "kv_cache_dtype":
298
                    str(self.cache_config.cache_dtype),
yhu422's avatar
yhu422 committed
299
300
301

                    # Feature flags
                    "enable_lora":
302
                    bool(self.lora_config),
yhu422's avatar
yhu422 committed
303
                    "enable_prefix_caching":
304
                    self.cache_config.enable_prefix_caching,
yhu422's avatar
yhu422 committed
305
                    "enforce_eager":
306
                    self.model_config.enforce_eager,
yhu422's avatar
yhu422 committed
307
                    "disable_custom_all_reduce":
308
                    self.parallel_config.disable_custom_all_reduce,
yhu422's avatar
yhu422 committed
309
                })
zhuwenwen's avatar
zhuwenwen committed
310

311
312
313
314
315
316
        self.cached_scheduler_outputs = [
            SchedulerOutputState()
            for _ in range(self.parallel_config.pipeline_parallel_size)
        ]

        self.scheduler_contexts = [
317
318
            SchedulerContext(multi_step_stream_outputs=self.scheduler_config.
                             multi_step_stream_outputs)
319
320
321
            for _ in range(self.parallel_config.pipeline_parallel_size)
        ]

322
        if self.model_config.use_async_output_proc:
323
            process_model_outputs = weak_bind(self._process_model_outputs)
zhuwenwen's avatar
zhuwenwen committed
324
325

            self.async_callbacks = [
326
327
                partial(process_model_outputs,
                        ctx=self.scheduler_contexts[v_id])
zhuwenwen's avatar
zhuwenwen committed
328
329
                for v_id in range(self.parallel_config.pipeline_parallel_size)
            ]
330
331
        else:
            self.async_callbacks = []
332
333
334

        # Currently used by AsyncLLMEngine to ensure quick append
        # of request outputs to asyncio queues
335
        self.process_request_outputs_callback: Optional[Callable] = None
336

337
        # Create the scheduler.
338
339
        # NOTE: the cache_config here have been updated with the numbers of
        # GPU and CPU blocks, which are profiled in the distributed executor.
340
341
342
343
344
        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
345
        self.scheduler = [
346
            Scheduler(
347
348
                self.scheduler_config, self.cache_config, self.lora_config,
                self.parallel_config.pipeline_parallel_size,
349
                self.async_callbacks[v_id]
350
351
                if self.model_config.use_async_output_proc else None)
            for v_id in range(self.parallel_config.pipeline_parallel_size)
352
        ]
zhuwenwen's avatar
zhuwenwen committed
353

354
355
        # Metric Logging.
        if self.log_stats:
356
357
358
            if stat_loggers is not None:
                self.stat_loggers = stat_loggers
            else:
359
360
361
362
363
364
365
                # 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)

366
367
368
                self.stat_loggers = {
                    "logging":
                    LoggingStatLogger(
369
370
                        local_interval=_LOCAL_LOGGING_INTERVAL_SEC,
                        vllm_config=vllm_config),
371
372
373
                    "prometheus":
                    PrometheusStatLogger(
                        local_interval=_LOCAL_LOGGING_INTERVAL_SEC,
374
375
                        labels=dict(
                            model_name=self.model_config.served_model_name),
376
                        vllm_config=vllm_config),
377
378
379
                }
                self.stat_loggers["prometheus"].info("cache_config",
                                                     self.cache_config)
380

381
382
383
384
385
386
        self.tracer = None
        if self.observability_config.otlp_traces_endpoint:
            self.tracer = init_tracer(
                "vllm.llm_engine",
                self.observability_config.otlp_traces_endpoint)

387
388
389
390
391
392
393
394
        # 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,
395
                get_tokenizer_for_seq,
396
397
                stop_checker=StopChecker(self.scheduler_config.max_model_len,
                                         get_tokenizer_for_seq),
398
            ))
399
400
        
        self.tree_decoding = os.environ.get('VLLM_TREE_DECODING') == '1'
401

402
403
        self.seq_id_to_seq_group: Dict[str, SequenceGroupBase] = {}

404
405
406
        # 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
407
        profile.StartTracer()
408

409
410
411
        # Don't keep the dummy data in memory
        self.reset_mm_cache()

412
413
414
415
416
417
    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.
        """
418
        start = time.time()
419
420
421
422
423
        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
424
425
426
427
            logger.info(
                "Overriding num_gpu_blocks=%d with "
                "num_gpu_blocks_override=%d", num_gpu_blocks,
                num_gpu_blocks_override)
428
429
430
431
432
433
            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)
434
435
436
        elapsed = time.time() - start
        logger.info(("init engine (profile, create kv cache, "
                     "warmup model) took %.2f seconds"), elapsed)
437

438
    @classmethod
439
    def _get_executor_cls(cls,
440
                          engine_config: VllmConfig) -> Type[ExecutorBase]:
441
        # distributed_executor_backend must be set in VllmConfig.__post_init__
442
443
        distributed_executor_backend = (
            engine_config.parallel_config.distributed_executor_backend)
444
        # Initialize the cluster and specify the executor class.
445
446
447
448
449
450
        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
451
452
453
454
455
456
457
458
459
460
461
462
463
        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.
464
465
            from vllm.executor.uniproc_executor import UniProcExecutor
            executor_class = UniProcExecutor
466
467
468
469
470
471
472
473
        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}")
474
475
        return executor_class

476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
    @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,
        )

492
493
494
495
496
497
498
499
500
    @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.
501
502
503
504
505
506
507
508
509
        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
510
            usage_context=usage_context,
511
            stat_loggers=stat_loggers,
512
            disable_log_stats=engine_args.disable_log_stats,
yhu422's avatar
yhu422 committed
513
        )
514

515
516
517
518
519
    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!")

520
521
522
523
524
525
    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()

526
527
    def get_tokenizer_group(self) -> TokenizerGroup:
        if self.tokenizer is None:
528
529
            raise ValueError("Unable to get tokenizer because "
                             "skip_tokenizer_init is True")
530

531
        return self.tokenizer
532

533
    def get_tokenizer(
534
535
536
        self,
        lora_request: Optional[LoRARequest] = None,
    ) -> AnyTokenizer:
537
538
539
540
        if self.model_config.tokenizer_mode == "cpm":
            return self.tokenizer
        else:
            return self.get_tokenizer_group().get_lora_tokenizer(lora_request)
541

542
    def _init_tokenizer(self) -> TokenizerGroup:
543
544
545
        return init_tokenizer_from_configs(
            model_config=self.model_config,
            scheduler_config=self.scheduler_config,
546
            lora_config=self.lora_config)
547

548
549
    def _verify_args(self) -> None:
        self.model_config.verify_with_parallel_config(self.parallel_config)
550
        self.cache_config.verify_with_parallel_config(self.parallel_config)
551
552
553
554
        if self.lora_config:
            self.lora_config.verify_with_model_config(self.model_config)
            self.lora_config.verify_with_scheduler_config(
                self.scheduler_config)
555

556
557
558
    def _add_processed_request(
        self,
        request_id: str,
559
        processed_inputs: ProcessorInputs,
560
561
562
        params: Union[SamplingParams, PoolingParams],
        arrival_time: float,
        lora_request: Optional[LoRARequest],
563
        trace_headers: Optional[Mapping[str, str]] = None,
564
        priority: int = 0,
565
    ) -> Optional[SequenceGroup]:
566
567
568
        """Add a processed request to the engine's request pool.
        return the created sequence group.
        """
569
570
571
572
573
574
575
576
577
578
579
580
581
        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,
                priority=priority,
            )
            return None

582
        self._validate_model_inputs(processed_inputs, lora_request)
583
584
585
        # Create the sequences.
        block_size = self.cache_config.block_size
        seq_id = next(self.seq_counter)
586
587
588
589
590
        #DEBUG @TODO change tokenizer false
        if self.model_config.tokenizer_mode == "cpm":
            eos_token_id = self.tokenizer.eos_id
        else:
            eos_token_id = self.input_preprocessor.get_eos_token_id(lora_request)
591

592
        encoder_inputs, decoder_inputs = split_enc_dec_inputs(processed_inputs)
593
594

        seq = Sequence(seq_id, decoder_inputs, block_size, eos_token_id,
595
                       lora_request)
596

597
        encoder_seq = (None if encoder_inputs is None else Sequence(
598
            seq_id, encoder_inputs, block_size, eos_token_id, lora_request))
599

600
601
602
603
604
605
606
607
        # 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,
608
                trace_headers=trace_headers,
609
610
                encoder_seq=encoder_seq,
                priority=priority)
611
612
613
614
615
616
617
        elif isinstance(params, PoolingParams):
            seq_group = self._create_sequence_group_with_pooling(
                request_id,
                seq,
                params,
                arrival_time=arrival_time,
                lora_request=lora_request,
618
619
                encoder_seq=encoder_seq,
                priority=priority)
620
621
622
623
        else:
            raise ValueError(
                "Either SamplingParams or PoolingParams must be provided.")

624
625
626
627
628
629
630
631
        # 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)

632
633
        return seq_group

634
635
    def stop_remote_worker_execution_loop(self) -> None:
        self.model_executor.stop_remote_worker_execution_loop()
636

637
638
639
    def add_request(
        self,
        request_id: str,
640
        prompt: PromptType,
641
        params: Union[SamplingParams, PoolingParams],
642
        arrival_time: Optional[float] = None,
643
        lora_request: Optional[LoRARequest] = None,
644
        tokenization_kwargs: Optional[dict[str, Any]] = None,
645
        trace_headers: Optional[Mapping[str, str]] = None,
646
        priority: int = 0,
647
    ) -> None:
Zhuohan Li's avatar
Zhuohan Li committed
648
        """Add a request to the engine's request pool.
649
650

        The request is added to the request pool and will be processed by the
Zhuohan Li's avatar
Zhuohan Li committed
651
        scheduler as `engine.step()` is called. The exact scheduling policy is
652
653
654
655
        determined by the scheduler.

        Args:
            request_id: The unique ID of the request.
656
657
            prompt: The prompt to the LLM. See
                [PromptType][vllm.inputs.PromptType]
658
659
                for more details about the format of each input.
            params: Parameters for sampling or pooling.
660
661
                [SamplingParams][vllm.SamplingParams] for text generation.
                [PoolingParams][vllm.PoolingParams] for pooling.
662
            arrival_time: The arrival time of the request. If None, we use
663
                the current monotonic time.
664
            lora_request: The LoRA request to add.
665
            trace_headers: OpenTelemetry trace headers.
666
667
            priority: The priority of the request.
                Only applicable with priority scheduling.
668
669
670
671

        Details:
            - Set arrival_time to the current time if it is None.
            - Set prompt_token_ids to the encoded prompt if it is None.
672
673
674
675
676
            - 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.
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692

        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
            >>> ...
693
        """
694
695
696
697
        if not isinstance(request_id, str):
            raise TypeError(
                f"request_id must be a string, got {type(request_id)}")

698
699
700
        if lora_request is not None and not self.lora_config:
            raise ValueError(f"Got lora_request {lora_request} but LoRA is "
                             "not enabled!")
701

702
        if priority != 0 and not self.scheduler_config.policy == "priority":
703
704
705
            raise ValueError(f"Got priority {priority} but "
                             "Priority scheduling is not enabled.")

706
707
708
709
710
711
712
        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")

713
        if arrival_time is None:
714
            arrival_time = time.time()
715

716
717
718
719
720
        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
721

722
723
724
725
        #DEBUG anrongqiao
        if self.model_config.tokenizer_mode == "cpm":
            lora_request = None
            
726
        processed_inputs = self.input_preprocessor.preprocess(
727
            prompt,
728
            tokenization_kwargs=tokenization_kwargs,
729
            lora_request=lora_request,
730
        )
731

732
733
734
735
736
737
        self._add_processed_request(
            request_id=request_id,
            processed_inputs=processed_inputs,
            params=params,
            arrival_time=arrival_time,
            lora_request=lora_request,
738
            trace_headers=trace_headers,
739
            priority=priority,
740
        )
741
742
743
744
745
746

    def _create_sequence_group_with_sampling(
        self,
        request_id: str,
        seq: Sequence,
        sampling_params: SamplingParams,
747
748
        arrival_time: float,
        lora_request: Optional[LoRARequest],
749
        trace_headers: Optional[Mapping[str, str]] = None,
750
        encoder_seq: Optional[Sequence] = None,
751
        priority: int = 0,
752
753
754
755
756
757
758
759
760
761
    ) -> 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.")

762
763
764
        sampling_params = self._build_logits_processors(
            sampling_params, lora_request)

765
766
767
        # Defensive copy of SamplingParams, which are used by the sampler,
        # this doesn't deep-copy LogitsProcessor objects
        sampling_params = sampling_params.clone()
768

769
        sampling_params.update_from_generation_config(
770
            self.generation_config_fields, seq.eos_token_id)
771

772
        # Create the sequence group.
773
774
775
776
        draft_size = 1
        if self.vllm_config.speculative_config is not None:
            draft_size = \
                self.vllm_config.speculative_config.num_speculative_tokens + 1
777
778
779
780
781
782
783
784
785
        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,
                                  encoder_seq=encoder_seq,
                                  priority=priority,
                                  draft_size=draft_size)
786

787
788
789
790
791
792
793
        return seq_group

    def _create_sequence_group_with_pooling(
        self,
        request_id: str,
        seq: Sequence,
        pooling_params: PoolingParams,
794
795
        arrival_time: float,
        lora_request: Optional[LoRARequest],
796
        encoder_seq: Optional[Sequence] = None,
797
        priority: int = 0,
798
799
800
801
802
    ) -> 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.
803
804
805
806
807
808
809
        seq_group = SequenceGroup(request_id=request_id,
                                  seqs=[seq],
                                  arrival_time=arrival_time,
                                  lora_request=lora_request,
                                  pooling_params=pooling_params,
                                  encoder_seq=encoder_seq,
                                  priority=priority)
810
        return seq_group
811

Antoni Baum's avatar
Antoni Baum committed
812
813
    def abort_request(self, request_id: Union[str, Iterable[str]]) -> None:
        """Aborts a request(s) with the given ID.
814
815

        Args:
Antoni Baum's avatar
Antoni Baum committed
816
            request_id: The ID(s) of the request to abort.
817
818

        Details:
819
            - Refer to [vllm.core.scheduler.Scheduler.abort_seq_group][].
820
821
822
823
824
825

        Example:
            >>> # initialize engine and add a request with request_id
            >>> request_id = str(0)
            >>> # abort the request
            >>> engine.abort_request(request_id)
826
        """
827
        for scheduler in self.scheduler:
828
829
            scheduler.abort_seq_group(
                request_id, seq_id_to_seq_group=self.seq_id_to_seq_group)
830

831
832
833
834
    def get_vllm_config(self) -> VllmConfig:
        """Gets the vllm configuration."""
        return self.vllm_config

835
836
837
838
    def get_model_config(self) -> ModelConfig:
        """Gets the model configuration."""
        return self.model_config

839
840
841
842
    def get_parallel_config(self) -> ParallelConfig:
        """Gets the parallel configuration."""
        return self.parallel_config

843
844
845
846
    def get_decoding_config(self) -> DecodingConfig:
        """Gets the decoding configuration."""
        return self.decoding_config

847
848
849
850
851
852
853
854
    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

855
    def get_num_unfinished_requests(self) -> int:
856
        """Gets the number of unfinished requests."""
857
858
        return sum(scheduler.get_num_unfinished_seq_groups()
                   for scheduler in self.scheduler)
859

860
    def has_unfinished_requests(self) -> bool:
861
        """Returns True if there are unfinished requests."""
862
863
864
865
866
867
868
869
870
        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()
871

872
873
874
875
    def reset_mm_cache(self) -> bool:
        """Reset the multi-modal cache."""
        return self.input_preprocessor.mm_registry.reset_processor_cache()

876
    def reset_prefix_cache(self, device: Optional[Device] = None) -> bool:
877
878
879
880
        """Reset prefix cache for all devices."""

        success = True
        for scheduler in self.scheduler:
881
            success = success and scheduler.reset_prefix_cache(device)
882
883
        return success

884
    @staticmethod
885
886
    def _process_sequence_group_outputs(
        seq_group: SequenceGroup,
887
        outputs: List[PoolingSequenceGroupOutput],
888
    ) -> None:
889
        seq_group.pooled_data = outputs[0].data
890
891
892
893
894
895

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

        return

896
897
898
899
900
901
902
903
    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.

904
        seq_group: SequenceGroup to update the num_computed_tokens for.
905
        seq_group_meta: Metadata of the given SequenceGroup.
906
        is_first_step_output: Optional[bool] -
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
            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)

935
936
937
938
939
    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.
940

941
942
        ctx: The virtual engine context to work on
        request_id: If provided, then only this request is going to be processed
943
        """
944

945
        now = time.time()
946

947
        if len(ctx.output_queue) == 0:
948
949
            return None

950
        # Get pending async postprocessor
951
952
953
954
        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,
955
             is_last_step, is_first_step_output, skip) = ctx.output_queue[0]
956
957
        else:
            (outputs, seq_group_metadata_list, scheduler_outputs, is_async,
958
959
             is_last_step, is_first_step_output,
             skip) = ctx.output_queue.popleft()
960
961
962
963
964

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

965
        has_multiple_outputs: bool = len(outputs) > 1
966
        outputs_by_sequence_group: List[List[SequenceGroupOutput]]
967
968
969
970
971
        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].
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
            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
989
990
991
            # We have outputs for multiple steps submitted in a single burst,
            # so invalidate is_first_step_output.
            is_first_step_output = None
992
        elif len(outputs) == 1:
993
            outputs_by_sequence_group = outputs
994
995
        else:
            return None
996

997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
        # 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

1014
        finished_before: List[int] = []
1015
        finished_now: List[int] = []
1016
        empty_seq_indices: List[int] = []
1017
1018
1019
1020
1021
        for i in indices:
            if i in skip:
                continue

            seq_group_meta = seq_group_metadata_list[i]
1022
            scheduled_seq_group = scheduler_outputs.scheduled_seq_groups[i]
1023

1024
            seq_group: SequenceGroup = scheduled_seq_group.seq_group
1025
1026
1027
1028
1029

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

1030
            output: List[SequenceGroupOutput]
1031
            if has_multiple_outputs:
1032
1033
1034
1035
                output = outputs_by_sequence_group[i]
            else:
                output = [outputs_by_sequence_group[0][i]]

1036
            # tree style speculative decoding may generate empty output in first step
1037
1038
1039
1040
1041
1042
1043
1044
1045
            if self.tree_decoding and outputs and isinstance(output[0], CompletionSequenceGroupOutput):
                samples = [o.samples[0] for o in output]
                valid_samples = [
                    sample for sample in samples
                    if sample.output_token != VLLM_INVALID_TOKEN_ID
                ]
                if len(valid_samples) == 0:
                    empty_seq_indices.append(i)
                    continue
1046

1047
            if not is_async:
1048
1049
1050
1051
1052
1053
                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(
1054
                        seq_group_meta.token_chunk_size or 0)
1055
1056
1057

            if outputs:
                for o in outputs:
1058
1059
1060
1061
                    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 += (
1062
                                o.model_forward_time or 0)
1063
1064
1065
1066
1067
                        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 += (
1068
                                o.model_execute_time or 0)
1069
1070
1071
                        else:
                            seq_group.metrics.model_execute_time = (
                                o.model_execute_time)
1072

1073
            if self.model_config.runner_type == "pooling":
1074
                self._process_sequence_group_outputs(seq_group, output)
1075
1076
1077
1078
1079
            else:
                self.output_processor.process_prompt_logprob(seq_group, output)
                if seq_group_meta.do_sample:
                    self.output_processor.process_outputs(
                        seq_group, output, is_async)
1080

1081
1082
            if seq_group.is_finished():
                finished_now.append(i)
1083

1084
1085
1086
        # Generate outputs for the requests that finished this iteration
        for i in finished_now:
            scheduled_seq_group = scheduler_outputs.scheduled_seq_groups[i]
1087

1088
            seq_group = scheduled_seq_group.seq_group
1089
            seq_group.maybe_set_first_token_time(now)
1090
1091
            if not seq_group.is_prefill():
                seq_group.set_last_token_time(now)
1092
            request_output = RequestOutputFactory.create(
1093
1094
1095
                seq_group,
                self.seq_id_to_seq_group,
                use_cache=self.use_cached_outputs)
1096
1097
            if request_output:
                ctx.request_outputs.append(request_output)
1098

1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
        # 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

1111
1112
1113
1114
1115
        # Free currently finished requests
        if finished_now:
            for scheduler in self.scheduler:
                scheduler.free_finished_seq_groups()

1116
1117
        # For multi-step without streaming, don't create outputs each iteration
        if not is_last_step and not ctx.multi_step_stream_outputs:
1118
1119
1120
1121
            # 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)
1122
                ctx.request_outputs.clear()
1123
1124
1125
            return

        # Create the outputs
1126
        for i in indices:
1127
            if i in skip or i in finished_before or i in finished_now or i in empty_seq_indices:
1128
1129
                continue  # Avoids double processing

1130
1131
            scheduled_seq_group = scheduler_outputs.scheduled_seq_groups[i]

1132
            seq_group = scheduled_seq_group.seq_group
1133
            seq_group.maybe_set_first_token_time(now)
1134
1135
            if not seq_group.is_prefill():
                seq_group.set_last_token_time(now)
1136
            request_output = RequestOutputFactory.create(
1137
1138
1139
                seq_group,
                self.seq_id_to_seq_group,
                use_cache=self.use_cached_outputs)
1140
            if request_output:
1141
                ctx.request_outputs.append(request_output)
1142

1143
1144
1145
1146
1147
1148
1149
1150
        # 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

1151
        for seq_group in scheduler_outputs.ignored_seq_groups:
1152
1153
1154
1155
1156
            params = seq_group.sampling_params
            if params is not None and params.output_kind == (
                    RequestOutputKind.DELTA) and not seq_group.is_finished():
                continue

1157
            request_output = RequestOutputFactory.create(
1158
1159
1160
1161
                seq_group,
                self.seq_id_to_seq_group,
                use_cache=self.use_cached_outputs,
            )
1162
1163
            if request_output:
                ctx.request_outputs.append(request_output)
1164

1165
1166
1167
1168
        # 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)
1169
            ctx.request_outputs.clear()
1170

1171
1172
1173
1174
        # 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:
1175
            # Log stats.
1176
1177
            self.do_log_stats(scheduler_outputs, outputs, finished_before,
                              skip)
1178
1179

            # Tracing
1180
            self.do_tracing(scheduler_outputs, finished_before)
1181
1182
1183
1184

        return None

    def _advance_to_next_step(
1185
            self, output: SamplerOutput,
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
            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

1199
1200
1201
1202
1203
            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)
1204
            else:
1205
1206
1207
1208
                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)
1209
1210
1211
1212

            if seq_group_metadata.do_sample:
                assert len(sequence_group_outputs.samples) == 1, (
                    "Async output processor expects a single sample"
1213
                    " (i.e sampling_params.n == 1)")
1214
1215
1216
1217
                sample = sequence_group_outputs.samples[0]

                assert len(seq_group.seqs) == 1
                seq = seq_group.seqs[0]
1218

1219
1220
1221
                if self.scheduler_config.is_multi_step:
                    is_prefill_append = seq.data.get_num_uncomputed_tokens(
                    ) == 0
1222
1223
                    seq.append_token_id(sample.output_token, sample.logprobs,
                                        sample.output_embed)
1224
1225
1226
                    if not is_prefill_append:
                        seq_group.update_num_computed_tokens(1)
                else:
1227
1228
                    seq.append_token_id(sample.output_token, sample.logprobs,
                                        sample.output_embed)
1229

1230
    def step(self) -> List[Union[RequestOutput, PoolingRequestOutput]]:
Antoni Baum's avatar
Antoni Baum committed
1231
1232
        """Performs one decoding iteration and returns newly generated results.

1233
1234
1235
1236
        <figure markdown="span">
        ![Overview of the step function](https://i.imgur.com/sv2HssD.png)
        <figcaption>Overview of the step function</figcaption>
        </figure>
1237
1238

        Details:
1239
1240
        - Step 1: Schedules the sequences to be executed in the next
            iteration and the token blocks to be swapped in/out/copy.
1241

1242
1243
1244
1245
            - 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.
1246

1247
1248
        - Step 2: Calls the distributed executor to execute the model.
        - Step 3: Processes the model output. This mainly includes:
1249

1250
1251
1252
1253
            - 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.
1254

1255
        - Finally, it creates and returns the newly generated results.
1256
1257

        Example:
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
        ```
        # 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
1281
        """
1282
1283
1284
1285
        if self.parallel_config.pipeline_parallel_size > 1:
            raise NotImplementedError(
                "Pipeline parallelism is only supported through AsyncLLMEngine "
                "as performance will be severely degraded otherwise.")
1286

1287
        # For llm_engine, there is no pipeline parallel support, so the engine
1288
        # used is always 0.
1289
1290
        virtual_engine = 0

1291
1292
        # These are cached outputs from previous iterations. None if on first
        # iteration
1293
        cached_outputs = self.cached_scheduler_outputs[virtual_engine]
1294
1295
        seq_group_metadata_list = cached_outputs.seq_group_metadata_list
        scheduler_outputs = cached_outputs.scheduler_outputs
1296
        allow_async_output_proc = cached_outputs.allow_async_output_proc
1297

1298
1299
        ctx = self.scheduler_contexts[virtual_engine]

1300
1301
1302
        # Clear outputs for each new scheduler iteration
        ctx.request_outputs.clear()

1303
1304
1305
        # 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.
1306
1307
1308
1309
1310
        # 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:
1311
            # Schedule iteration
1312
            (seq_group_metadata_list, scheduler_outputs,
1313
1314
             allow_async_output_proc
             ) = self.scheduler[virtual_engine].schedule()
1315

1316
1317
            ctx.seq_group_metadata_list = seq_group_metadata_list
            ctx.scheduler_outputs = scheduler_outputs
1318

1319
1320
            finished_requests_ids = self.scheduler[
                virtual_engine].get_and_reset_finished_requests_ids()
1321
1322
1323
1324
1325
            # 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]
1326

1327
1328
            # Maybe switch from async mode to sync mode
            if not allow_async_output_proc and len(ctx.output_queue) > 0:
1329
                self._process_model_outputs(ctx=ctx)
1330

1331
1332
1333
1334
1335
            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(
1336
                    virtual_engine, seq_group_metadata_list, scheduler_outputs,
1337
                    allow_async_output_proc)
1338
1339
        else:
            finished_requests_ids = list()
1340
1341
1342

        assert seq_group_metadata_list is not None
        assert scheduler_outputs is not None
Antoni Baum's avatar
Antoni Baum committed
1343

1344
        if not scheduler_outputs.is_empty():
1345
1346
1347
1348
1349
1350

            # 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 = \
1351
                self._get_last_sampled_token_ids(virtual_engine)
1352

1353
            execute_model_req = ExecuteModelRequest(
1354
1355
1356
1357
                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,
1358
1359
                num_lookahead_slots=scheduler_outputs.num_lookahead_slots,
                running_queue_size=scheduler_outputs.running_queue_size,
1360
1361
1362
1363
1364
                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)

1365
            if allow_async_output_proc:
1366
1367
                execute_model_req.async_callback = self.async_callbacks[
                    virtual_engine]
1368

1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
            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
1386

1387
            # We need to do this here so that last step's sampled_token_ids can
1388
1389
            # be passed to the next iteration for PP.
            if self.scheduler_config.is_multi_step:
1390
                self._update_cached_scheduler_output(virtual_engine, outputs)
1391
        else:
1392
1393
            # Nothing scheduled => If there is pending async postprocessor,
            # then finish it here.
1394
1395
            if len(ctx.output_queue) > 0:
                self._process_model_outputs(ctx=ctx)
1396
            # No outputs in this case
1397
            outputs = []
Antoni Baum's avatar
Antoni Baum committed
1398

1399
1400
1401
1402
1403
1404
        # 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):
1405
            # clear the cache if we have finished all the steps.
1406
1407
1408
            if self.scheduler_config.is_multi_step:
                self.cached_scheduler_outputs[0] = SchedulerOutputState()

1409
1410
1411
1412
1413
1414
            # 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

1415
            # Add results to the output_queue
1416
1417
1418
1419
            ctx.append_output(outputs=outputs,
                              seq_group_metadata_list=seq_group_metadata_list,
                              scheduler_outputs=scheduler_outputs,
                              is_async=allow_async_output_proc,
1420
1421
                              is_last_step=True,
                              is_first_step_output=is_first_step_output)
1422
1423
1424

            if outputs and allow_async_output_proc:
                assert len(outputs) == 1, (
1425
                    "Async postprocessor expects only a single output set")
1426

1427
                self._advance_to_next_step(
1428
                    outputs[0], seq_group_metadata_list,
1429
                    scheduler_outputs.scheduled_seq_groups)
1430

1431
            # Check if need to run the usual non-async path
1432
            if not allow_async_output_proc:
1433
                self._process_model_outputs(ctx=ctx)
1434

1435
                # Log stats.
1436
                self.do_log_stats(scheduler_outputs, outputs)
1437

1438
1439
1440
                # Tracing
                self.do_tracing(scheduler_outputs)
        else:
1441
            # Multi-step case
1442
            return ctx.request_outputs
1443

1444
        if not self.has_unfinished_requests():
1445
1446
            # Drain async postprocessor (if exists)
            if len(ctx.output_queue) > 0:
1447
                self._process_model_outputs(ctx=ctx)
1448
            assert len(ctx.output_queue) == 0
1449

1450
1451
1452
1453
1454
            # 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.
1455
            logger.debug("Stopping remote worker execution loop.")
1456
1457
            self.model_executor.stop_remote_worker_execution_loop()

1458
        return ctx.request_outputs
Antoni Baum's avatar
Antoni Baum committed
1459

1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
    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)

1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
    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:]
        ]):
1507
1508
            raise AssertionError("All running sequence groups should "
                                 "have the same remaining steps.")
1509
1510
1511
1512
1513
1514

        return ref_remaining_steps > 0

    def _cache_scheduler_outputs_for_multi_step(
            self, virtual_engine: int,
            seq_group_metadata_list: Optional[List[SequenceGroupMetadata]],
1515
1516
1517
1518
1519
1520
1521
1522
            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
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546

    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
Antoni Baum's avatar
Antoni Baum committed
1547

1548
    def add_logger(self, logger_name: str, logger: StatLoggerBase) -> None:
1549
1550
1551
1552
        if not self.log_stats:
            raise RuntimeError(
                "Stat logging is disabled. Set `disable_log_stats=False` "
                "argument to enable.")
1553
1554
1555
1556
1557
        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:
1558
1559
1560
1561
        if not self.log_stats:
            raise RuntimeError(
                "Stat logging is disabled. Set `disable_log_stats=False` "
                "argument to enable.")
1562
1563
1564
1565
        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]

1566
1567
1568
    def do_log_stats(self,
                     scheduler_outputs: Optional[SchedulerOutputs] = None,
                     model_output: Optional[List[SamplerOutput]] = None,
1569
1570
                     finished_before: Optional[List[int]] = None,
                     skip: Optional[List[int]] = None) -> None:
1571
1572
        """Forced log when no requests active."""
        if self.log_stats:
1573
            stats = self._get_stats(scheduler_outputs, model_output,
1574
                                    finished_before, skip)
1575
            for logger in self.stat_loggers.values():
1576
                logger.log(stats)
1577

1578
1579
1580
    def _get_stats(self,
                   scheduler_outputs: Optional[SchedulerOutputs],
                   model_output: Optional[List[SamplerOutput]] = None,
1581
1582
                   finished_before: Optional[List[int]] = None,
                   skip: Optional[List[int]] = None) -> Stats:
1583
1584
1585
1586
1587
1588
1589
        """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.
1590
1591
1592
1593
            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.
1594
        """
1595
        now = time.time()
Woosuk Kwon's avatar
Woosuk Kwon committed
1596

1597
1598
        # System State
        #   Scheduler State
1599
1600
1601
1602
1603
1604
        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)
1605
1606

        # KV Cache Usage in %
1607
        num_total_gpu = self.cache_config.num_gpu_blocks
1608
        gpu_cache_usage_sys = 0.
1609
        if num_total_gpu:  # Guard against both None and 0
1610
1611
1612
            num_free_gpu = sum(
                scheduler.block_manager.get_num_free_gpu_blocks()
                for scheduler in self.scheduler)
1613
            gpu_cache_usage_sys = 1.0 - (num_free_gpu / num_total_gpu)
Woosuk Kwon's avatar
Woosuk Kwon committed
1614

1615
        num_total_cpu = self.cache_config.num_cpu_blocks
1616
        cpu_cache_usage_sys = 0.
1617
        if num_total_cpu:  # Guard against both None and 0
1618
1619
1620
            num_free_cpu = sum(
                scheduler.block_manager.get_num_free_cpu_blocks()
                for scheduler in self.scheduler)
1621
1622
            cpu_cache_usage_sys = 1.0 - (num_free_cpu / num_total_cpu)

1623
1624
1625
1626
1627
1628
1629
        # 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)

1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
        # 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,
            )

1644
1645
1646
        # Iteration stats
        num_prompt_tokens_iter = 0
        num_generation_tokens_iter = 0
harrywu's avatar
harrywu committed
1647
        num_tokens_iter = 0
1648
1649
        time_to_first_tokens_iter: List[float] = []
        time_per_output_tokens_iter: List[float] = []
1650
1651
        num_preemption_iter = (0 if scheduler_outputs is None else
                               scheduler_outputs.preempted)
1652
1653
1654
1655

        # Request stats
        #   Latency
        time_e2e_requests: List[float] = []
harrywu's avatar
harrywu committed
1656
1657
1658
1659
        time_queue_requests: List[float] = []
        time_inference_requests: List[float] = []
        time_prefill_requests: List[float] = []
        time_decode_requests: List[float] = []
1660
1661
1662
1663
        #   Metadata
        num_prompt_tokens_requests: List[int] = []
        num_generation_tokens_requests: List[int] = []
        n_requests: List[int] = []
harrywu's avatar
harrywu committed
1664
        max_num_generation_tokens_requests: List[int] = []
1665
        max_tokens_requests: List[int] = []
1666
1667
        finished_reason_requests: List[str] = []

1668
        # LoRA requests
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
        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)

1687
1688
        # NOTE: This loop assumes prefill seq_groups are before
        # decode seq_groups in scheduled_seq_groups.
1689
        if scheduler_outputs is not None:
1690
1691
1692
1693
            # 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

1694
            num_generation_tokens_from_prefill_groups = 0
1695
1696
1697
1698
            # 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.
1699
1700
1701

            for idx, scheduled_seq_group in enumerate(
                    scheduler_outputs.scheduled_seq_groups):
1702
1703
1704
1705
1706
                # Skip double logging when using async output proc
                if finished_before and idx in finished_before:
                    actual_num_batched_tokens -= 1
                    continue

1707
1708
1709
1710
                # Currently, skip == preempted sequences, so we need to skip
                # their log stats
                if skip and idx in skip:
                    continue
1711

1712
                group_was_prefill = idx < scheduler_outputs.num_prefill_groups
1713
                seq_group = scheduled_seq_group.seq_group
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725

                # 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():
1726
                        latency = seq_group.get_last_token_latency()
1727
1728
1729
1730
1731
1732
1733
                        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.
1734
                    latency = seq_group.get_last_token_latency()
1735
                    time_per_output_tokens_iter.append(latency)
1736
1737
1738
1739
1740
1741
1742
1743
1744
                    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
1745
1746
1747
1748
1749
1750

                # 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.
1751
                if seq_group.is_finished():
1752
                    # Latency timings
1753
1754
                    time_e2e_requests.append(now -
                                             seq_group.metrics.arrival_time)
harrywu's avatar
harrywu committed
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
                    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)
1767
1768
1769
1770
1771
1772
1773
                    # 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
1774
1775
1776
                    max_num_generation_tokens_requests.append(
                        max(seq.get_output_len()
                            for seq in seq_group.get_seqs()))
1777
1778
                    if seq_group.sampling_params is not None:
                        n_requests.append(seq_group.sampling_params.n)
1779
1780
                        max_tokens_requests.append(
                            seq_group.sampling_params.max_tokens)
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
                    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 = (
1793
                actual_num_batched_tokens - num_prompt_tokens_iter +
1794
                num_generation_tokens_from_prefill_groups)
harrywu's avatar
harrywu committed
1795
1796
            num_tokens_iter = (num_generation_tokens_iter +
                               num_prompt_tokens_iter)
1797

1798
1799
        return Stats(
            now=now,
1800
1801
1802
1803
1804
1805
1806
1807
            # 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,
1808
1809
1810
            #   Prefix Cache Hit Rate
            cpu_prefix_cache_hit_rate=cpu_prefix_cache_hit_rate,
            gpu_prefix_cache_hit_rate=gpu_prefix_cache_hit_rate,
1811
1812
1813
1814

            # Iteration stats
            num_prompt_tokens_iter=num_prompt_tokens_iter,
            num_generation_tokens_iter=num_generation_tokens_iter,
harrywu's avatar
harrywu committed
1815
            num_tokens_iter=num_tokens_iter,
1816
1817
            time_to_first_tokens_iter=time_to_first_tokens_iter,
            time_per_output_tokens_iter=time_per_output_tokens_iter,
1818
            num_preemption_iter=num_preemption_iter,
1819
1820
1821
1822

            # Request stats
            #   Latency
            time_e2e_requests=time_e2e_requests,
harrywu's avatar
harrywu committed
1823
1824
1825
1826
            time_queue_requests=time_queue_requests,
            time_inference_requests=time_inference_requests,
            time_prefill_requests=time_prefill_requests,
            time_decode_requests=time_decode_requests,
1827
1828
1829
            #   Metadata
            num_prompt_tokens_requests=num_prompt_tokens_requests,
            num_generation_tokens_requests=num_generation_tokens_requests,
harrywu's avatar
harrywu committed
1830
1831
            max_num_generation_tokens_requests=
            max_num_generation_tokens_requests,
1832
            n_requests=n_requests,
1833
            max_tokens_requests=max_tokens_requests,
1834
            finished_reason_requests=finished_reason_requests,
1835
1836
1837
            max_lora=str(max_lora_stat),
            waiting_lora_adapters=list(waiting_lora_adapters.keys()),
            running_lora_adapters=list(running_lora_adapters.keys()))
1838

1839
    def add_lora(self, lora_request: LoRARequest) -> bool:
1840
        return self.model_executor.add_lora(lora_request)
1841
1842

    def remove_lora(self, lora_id: int) -> bool:
1843
        return self.model_executor.remove_lora(lora_id)
1844

1845
    def list_loras(self) -> Set[int]:
1846
        return self.model_executor.list_loras()
1847

1848
1849
1850
    def pin_lora(self, lora_id: int) -> bool:
        return self.model_executor.pin_lora(lora_id)

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
        model_config = self.model_config
zhuwenwen's avatar
zhuwenwen committed
1968
1969
1970
1971
1972
1973
1974
1975
        if self.tokenizer is None:
            tokenizer = None
        elif self.model_config.tokenizer_mode != "cpm":
            tokenizer = self.tokenizer.get_lora_tokenizer(lora_request)
        else:
            tokenizer = self.tokenizer
        # tokenizer = (None if self.tokenizer is None else
        #              self.tokenizer.get_lora_tokenizer(lora_request))
1976

1977
        prompt_ids = prompt_inputs.get("prompt_token_ids", [])
1978
1979
1980
        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
1981
            elif prompt_inputs["type"] == "embeds":
1982
                pass
1983
1984
1985
            else:
                raise ValueError(f"The {prompt_type} prompt cannot be empty")

1986
1987
1988
1989
1990
1991
        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")

1992
        max_prompt_len = self.model_config.max_model_len
1993
        if len(prompt_ids) > max_prompt_len:
1994
            if prompt_type == "encoder" and model_config.is_multimodal_model:
1995
1996
                mm_registry = self.input_preprocessor.mm_registry
                mm_processor = mm_registry.create_processor(
1997
1998
1999
                    model_config,
                    tokenizer=tokenizer or object(),  # Dummy if no tokenizer
                )
2000
                assert isinstance(mm_processor, EncDecMultiModalProcessor)
2001

2002
2003
2004
                if mm_processor.pad_dummy_encoder_prompt:
                    return  # Skip encoder length check for Whisper

2005
            if model_config.is_multimodal_model:
2006
                suggestion = (
2007
2008
2009
2010
                    "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.")
2011
2012
2013
2014
2015
2016
2017
2018
2019
            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}")
2020
2021
2022
2023

            # 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
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033

    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 = []
2034

2035
2036
2037
2038
2039
        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
2040
2041
2042
2043
2044
2045
2046

            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 \
2047
                self.decoding_config.backend
2048

2049
            if self.decoding_config.reasoning_backend:
2050
2051
                logger.debug("Building with reasoning backend %s",
                             self.decoding_config.reasoning_backend)
2052

2053
            processor = get_local_guided_decoding_logits_processor(
2054
2055
                guided_params=guided_decoding,
                tokenizer=tokenizer,
2056
2057
2058
                model_config=self.model_config,
                reasoning_backend=self.decoding_config.reasoning_backend,
            )
2059
2060
2061
2062
2063
2064
2065
2066
2067
            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)

2068
            processors = get_openai_logits_processors(
2069
2070
2071
2072
2073
2074
2075
2076
2077
                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

2078
2079
2080
2081
2082
2083
        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)

2084
2085
2086
2087
2088
2089
2090
        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
2091

2092
2093
2094
2095
2096
2097
2098
2099
    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)

2100

2101
2102
2103
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