llm_engine.py 79.4 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4

5
import os
6
import copy
7

Antoni Baum's avatar
Antoni Baum committed
8
import time
9
from collections import Counter as collectionsCounter
10
from collections import deque
11
from contextlib import contextmanager
12
from dataclasses import dataclass
13
from functools import partial
14
from typing import (TYPE_CHECKING, Any, Callable, ClassVar, Deque, Dict,
15
                    Iterable, List, Literal, Mapping, NamedTuple, Optional)
16
from typing import Sequence as GenericSequence
17
from typing import Set, Type, Union, cast
18

19
import torch
20
from typing_extensions import TypeVar
21

22
import vllm.envs as envs
23
24
25
from vllm.config import (DecodingConfig, LoRAConfig, ModelConfig,
                         ObservabilityConfig, ParallelConfig, SchedulerConfig,
                         VllmConfig)
26
from vllm.core.scheduler import ScheduledSequenceGroup, SchedulerOutputs
Woosuk Kwon's avatar
Woosuk Kwon committed
27
from vllm.engine.arg_utils import EngineArgs
28
from vllm.engine.metrics_types import StatLoggerBase, Stats
29
30
31
from vllm.engine.output_processor.interfaces import (
    SequenceGroupOutputProcessor)
from vllm.engine.output_processor.stop_checker import StopChecker
32
33
from vllm.entrypoints.openai.logits_processors import (
    get_logits_processors as get_openai_logits_processors)
34
from vllm.executor.executor_base import ExecutorBase
35
from vllm.inputs import ProcessorInputs, PromptType, SingletonInputs
36
from vllm.inputs.parse import split_enc_dec_inputs
37
from vllm.inputs.preprocess import InputPreprocessor
Woosuk Kwon's avatar
Woosuk Kwon committed
38
from vllm.logger import init_logger
39
from vllm.logits_process import get_bad_words_logits_processors
40
from vllm.lora.request import LoRARequest
41
from vllm.model_executor.layers.sampler import SamplerOutput
42
from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalRegistry
43
from vllm.multimodal.cache import processor_only_cache_from_config
44
from vllm.multimodal.processing import EncDecMultiModalProcessor
45
from vllm.outputs import (PoolingRequestOutput, RequestOutput,
46
                          RequestOutputFactory)
47
from vllm.sampling_params import RequestOutputKind, SamplingParams
48
from vllm.sequence import (ExecuteModelRequest, ParallelSampleSequenceGroup,
49
50
                           Sequence, SequenceGroup, SequenceGroupBase,
                           SequenceGroupMetadata, SequenceGroupOutput,
51
                           SequenceStatus, CompletionSequenceGroupOutput, VLLM_INVALID_TOKEN_ID)
52
53
from vllm.tracing import (SpanAttributes, SpanKind, extract_trace_context,
                          init_tracer)
54
from vllm.transformers_utils.detokenizer import Detokenizer
55
from vllm.transformers_utils.tokenizer import AnyTokenizer
56
57
# DEBUG add cpm tokenizer
from vllm.transformers_utils.tokenizers import CPM9GTokenizer
58
from vllm.transformers_utils.tokenizer_group import (
59
    TokenizerGroup, init_tokenizer_from_configs)
yhu422's avatar
yhu422 committed
60
61
from vllm.usage.usage_lib import (UsageContext, is_usage_stats_enabled,
                                  usage_message)
62
from vllm.utils import Counter, Device, resolve_obj_by_qualname, weak_bind
63
from vllm.version import __version__ as VLLM_VERSION
64
from vllm.worker.model_runner_base import InputProcessingError
65
from vllm.profiler.prof import profile
66
67

logger = init_logger(__name__)
68
_LOCAL_LOGGING_INTERVAL_SEC = 5
Woosuk Kwon's avatar
Woosuk Kwon committed
69

70
_O = TypeVar("_O", RequestOutput, PoolingRequestOutput)
71
_R = TypeVar("_R", default=Any)
72
73


74
75
76
77
78
@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
79
80
    allow_async_output_proc: bool = False
    last_output: Optional[SamplerOutput] = None
81
82


83
84
85
86
87
88
class OutputData(NamedTuple):
    outputs: List[SamplerOutput]
    seq_group_metadata_list: List[SequenceGroupMetadata]
    scheduler_outputs: SchedulerOutputs
    is_async: bool
    is_last_step: bool
89
90
91
92
93
94
    # 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]
95
96
97
    skip: List[int]


98
class SchedulerContext:
99

100
    def __init__(self) -> None:
101
        self.output_queue: Deque[OutputData] = deque()
102
        self.request_outputs: List[RequestOutput] = []
103
104
105
106
107
108
109
        self.seq_group_metadata_list: Optional[
            List[SequenceGroupMetadata]] = None
        self.scheduler_outputs: Optional[SchedulerOutputs] = None

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


122
class LLMEngine:
Zhuohan Li's avatar
Zhuohan Li committed
123
    """An LLM engine that receives requests and generates texts.
124

Woosuk Kwon's avatar
Woosuk Kwon committed
125
    This is the main class for the vLLM engine. It receives requests
126
127
128
129
130
131
    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.

132
133
134
    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.
135

136
    The config arguments are derived from [`EngineArgs`][vllm.EngineArgs].
137
138

    Args:
139
        vllm_config: The configuration for initializing and running vLLM.
140
141
        executor_class: The model executor class for managing distributed
            execution.
142
        log_stats: Whether to log statistics.
143
        usage_context: Specified entry point, used for usage info collection.
144
    """
145

146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
    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)}")

171
        return cast(_O, output)
172
173
174

    @classmethod
    def validate_outputs(
zhuwenwen's avatar
zhuwenwen committed
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
        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

193
194
        return outputs_

195
    tokenizer: Optional[TokenizerGroup]
196

197
198
    def __init__(
        self,
199
        vllm_config: VllmConfig,
200
        executor_class: Type[ExecutorBase],
201
        log_stats: bool,
yhu422's avatar
yhu422 committed
202
        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
203
        stat_loggers: Optional[Dict[str, StatLoggerBase]] = None,
204
        mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
205
        use_cached_outputs: bool = False,
206
    ) -> None:
207
208
209
210
211
212
        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.")
213

214
        self.vllm_config = vllm_config
215
216
217
218
219
220
221
222
223
        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
224
        )
225
        self.observability_config = vllm_config.observability_config or ObservabilityConfig(  # noqa
226
227
        )

228
        logger.info(
229
            "Initializing a V0 LLM engine (v%s) with config: %s, "
230
            "use_cached_outputs=%s, ",
231
            VLLM_VERSION,
232
            vllm_config,
233
            use_cached_outputs,
234
        )
235

236
        self.log_stats = log_stats
237
        self.use_cached_outputs = use_cached_outputs
238

239
        if self.model_config.skip_tokenizer_init:
240
            self.tokenizer = None
241
            self.detokenizer = None
242
            tokenizer_group = None
243
244
245
246
        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()
247
        else:
248
249
250
            self.tokenizer = self._init_tokenizer()
            self.detokenizer = Detokenizer(self.tokenizer)
            tokenizer_group = self.get_tokenizer_group()
251
252
253
254
255
256
257

        # 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)
258

259
        self.seq_counter = Counter()
260
261
        self.generation_config_fields = (
            self.model_config.try_get_generation_config())
262

263
264
265
266
267
268
269
        self.input_preprocessor = InputPreprocessor(
            self.model_config,
            self.tokenizer,
            mm_registry,
            mm_processor_cache=processor_only_cache_from_config(
                self.model_config, mm_registry),
        )
270

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

273
        self._initialize_kv_caches()
274

yhu422's avatar
yhu422 committed
275
276
        # If usage stat is enabled, collect relevant info.
        if is_usage_stats_enabled():
277
278
            from vllm.model_executor.model_loader import (
                get_architecture_class_name)
yhu422's avatar
yhu422 committed
279
            usage_message.report_usage(
280
                get_architecture_class_name(self.model_config),
yhu422's avatar
yhu422 committed
281
282
283
284
                usage_context,
                extra_kvs={
                    # Common configuration
                    "dtype":
285
                    str(self.model_config.dtype),
yhu422's avatar
yhu422 committed
286
                    "tensor_parallel_size":
287
                    self.parallel_config.tensor_parallel_size,
yhu422's avatar
yhu422 committed
288
                    "block_size":
289
                    self.cache_config.block_size,
yhu422's avatar
yhu422 committed
290
                    "gpu_memory_utilization":
291
                    self.cache_config.gpu_memory_utilization,
yhu422's avatar
yhu422 committed
292
293
294

                    # Quantization
                    "quantization":
295
                    self.model_config.quantization,
yhu422's avatar
yhu422 committed
296
                    "kv_cache_dtype":
297
                    str(self.cache_config.cache_dtype),
yhu422's avatar
yhu422 committed
298
299
300

                    # Feature flags
                    "enable_lora":
301
                    bool(self.lora_config),
yhu422's avatar
yhu422 committed
302
                    "enable_prefix_caching":
303
                    self.cache_config.enable_prefix_caching,
yhu422's avatar
yhu422 committed
304
                    "enforce_eager":
305
                    self.model_config.enforce_eager,
yhu422's avatar
yhu422 committed
306
                    "disable_custom_all_reduce":
307
                    self.parallel_config.disable_custom_all_reduce,
yhu422's avatar
yhu422 committed
308
                })
zhuwenwen's avatar
zhuwenwen committed
309

310
311
312
313
314
315
        self.cached_scheduler_outputs = [
            SchedulerOutputState()
            for _ in range(self.parallel_config.pipeline_parallel_size)
        ]

        self.scheduler_contexts = [
316
            SchedulerContext()
317
318
319
            for _ in range(self.parallel_config.pipeline_parallel_size)
        ]

320
        if self.model_config.use_async_output_proc:
321
            process_model_outputs = weak_bind(self._process_model_outputs)
zhuwenwen's avatar
zhuwenwen committed
322
323

            self.async_callbacks = [
324
325
                partial(process_model_outputs,
                        ctx=self.scheduler_contexts[v_id])
zhuwenwen's avatar
zhuwenwen committed
326
327
                for v_id in range(self.parallel_config.pipeline_parallel_size)
            ]
328
329
        else:
            self.async_callbacks = []
330
331
332

        # Currently used by AsyncLLMEngine to ensure quick append
        # of request outputs to asyncio queues
333
        self.process_request_outputs_callback: Optional[Callable] = None
334

335
        # Create the scheduler.
336
337
        # NOTE: the cache_config here have been updated with the numbers of
        # GPU and CPU blocks, which are profiled in the distributed executor.
338
339
340
341
342
        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
343
        self.scheduler = [
344
            Scheduler(
345
346
                self.scheduler_config, self.cache_config, self.lora_config,
                self.parallel_config.pipeline_parallel_size,
347
                self.async_callbacks[v_id]
348
349
                if self.model_config.use_async_output_proc else None)
            for v_id in range(self.parallel_config.pipeline_parallel_size)
350
        ]
zhuwenwen's avatar
zhuwenwen committed
351

352
353
        # Metric Logging.
        if self.log_stats:
354
355
356
            if stat_loggers is not None:
                self.stat_loggers = stat_loggers
            else:
357
358
359
360
361
362
363
                # 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)

364
365
366
                self.stat_loggers = {
                    "logging":
                    LoggingStatLogger(
367
368
                        local_interval=_LOCAL_LOGGING_INTERVAL_SEC,
                        vllm_config=vllm_config),
369
370
371
                    "prometheus":
                    PrometheusStatLogger(
                        local_interval=_LOCAL_LOGGING_INTERVAL_SEC,
372
373
                        labels=dict(
                            model_name=self.model_config.served_model_name),
374
                        vllm_config=vllm_config),
375
376
377
                }
                self.stat_loggers["prometheus"].info("cache_config",
                                                     self.cache_config)
378

379
380
381
382
383
384
        self.tracer = None
        if self.observability_config.otlp_traces_endpoint:
            self.tracer = init_tracer(
                "vllm.llm_engine",
                self.observability_config.otlp_traces_endpoint)

385
386
387
388
389
390
391
392
        # 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,
393
                get_tokenizer_for_seq,
394
395
                stop_checker=StopChecker(self.scheduler_config.max_model_len,
                                         get_tokenizer_for_seq),
396
            ))
397
398
        
        self.tree_decoding = os.environ.get('VLLM_TREE_DECODING') == '1'
399

400
401
        self.seq_id_to_seq_group: Dict[str, SequenceGroupBase] = {}

402
403
404
        # 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
405
        profile.StartTracer()
406

407
408
409
        # Don't keep the dummy data in memory
        self.reset_mm_cache()

410
411
412
413
414
415
    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.
        """
416
        start = time.time()
417
418
419
420
421
        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
422
423
424
425
            logger.info(
                "Overriding num_gpu_blocks=%d with "
                "num_gpu_blocks_override=%d", num_gpu_blocks,
                num_gpu_blocks_override)
426
427
428
429
430
431
            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)
432
433
434
        elapsed = time.time() - start
        logger.info(("init engine (profile, create kv cache, "
                     "warmup model) took %.2f seconds"), elapsed)
435

436
    @classmethod
437
    def _get_executor_cls(cls,
438
                          engine_config: VllmConfig) -> Type[ExecutorBase]:
439
        # distributed_executor_backend must be set in VllmConfig.__post_init__
440
441
        distributed_executor_backend = (
            engine_config.parallel_config.distributed_executor_backend)
442
        # Initialize the cluster and specify the executor class.
443
444
445
446
447
448
        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
449
450
451
452
453
454
455
456
457
458
459
460
461
        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.
462
463
            from vllm.executor.uniproc_executor import UniProcExecutor
            executor_class = UniProcExecutor
464
465
466
467
468
469
470
471
        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}")
472
473
        return executor_class

474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
    @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,
        )

490
491
492
493
494
495
496
497
498
    @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.
499
500
501
502
503
504
505
506
507
        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
508
            usage_context=usage_context,
509
            stat_loggers=stat_loggers,
510
            disable_log_stats=engine_args.disable_log_stats,
yhu422's avatar
yhu422 committed
511
        )
512

513
514
515
516
517
    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!")

518
519
520
521
522
523
    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()

524
525
    def get_tokenizer_group(self) -> TokenizerGroup:
        if self.tokenizer is None:
526
527
            raise ValueError("Unable to get tokenizer because "
                             "skip_tokenizer_init is True")
528

529
        return self.tokenizer
530

531
    def get_tokenizer(
532
533
534
        self,
        lora_request: Optional[LoRARequest] = None,
    ) -> AnyTokenizer:
535
536
537
538
        if self.model_config.tokenizer_mode == "cpm":
            return self.tokenizer
        else:
            return self.get_tokenizer_group().get_lora_tokenizer(lora_request)
539

540
    def _init_tokenizer(self) -> TokenizerGroup:
541
542
543
        return init_tokenizer_from_configs(
            model_config=self.model_config,
            scheduler_config=self.scheduler_config,
544
            lora_config=self.lora_config)
545

546
547
    def _verify_args(self) -> None:
        self.model_config.verify_with_parallel_config(self.parallel_config)
548
        self.cache_config.verify_with_parallel_config(self.parallel_config)
549
550
551
552
        if self.lora_config:
            self.lora_config.verify_with_model_config(self.model_config)
            self.lora_config.verify_with_scheduler_config(
                self.scheduler_config)
553

554
555
556
    def _add_processed_request(
        self,
        request_id: str,
557
        processed_inputs: ProcessorInputs,
558
        params: SamplingParams,
559
560
        arrival_time: float,
        lora_request: Optional[LoRARequest],
561
        trace_headers: Optional[Mapping[str, str]] = None,
562
        priority: int = 0,
563
    ) -> Optional[SequenceGroup]:
564
565
566
        """Add a processed request to the engine's request pool.
        return the created sequence group.
        """
567
568
569
570
571
572
573
574
575
576
577
578
579
        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

580
        self._validate_model_inputs(processed_inputs, lora_request)
581
582
583
        # Create the sequences.
        block_size = self.cache_config.block_size
        seq_id = next(self.seq_counter)
584
585
586
587
588
        #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)
589

590
        encoder_inputs, decoder_inputs = split_enc_dec_inputs(processed_inputs)
591
592

        seq = Sequence(seq_id, decoder_inputs, block_size, eos_token_id,
593
                       lora_request)
594

595
        encoder_seq = (None if encoder_inputs is None else Sequence(
596
            seq_id, encoder_inputs, block_size, eos_token_id, lora_request))
597

598
        # Create a SequenceGroup based on SamplingParams
599
600
601
602
603
604
605
        if isinstance(params, SamplingParams):
            seq_group = self._create_sequence_group_with_sampling(
                request_id,
                seq,
                params,
                arrival_time=arrival_time,
                lora_request=lora_request,
606
                trace_headers=trace_headers,
607
608
                encoder_seq=encoder_seq,
                priority=priority)
609
        else:
610
            raise ValueError("SamplingParams must be provided.")
611

612
613
614
615
616
617
618
619
        # 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)

620
621
        return seq_group

622
623
    def stop_remote_worker_execution_loop(self) -> None:
        self.model_executor.stop_remote_worker_execution_loop()
624

625
626
627
    def add_request(
        self,
        request_id: str,
628
        prompt: PromptType,
629
        params: SamplingParams,
630
        arrival_time: Optional[float] = None,
631
        lora_request: Optional[LoRARequest] = None,
632
        tokenization_kwargs: Optional[dict[str, Any]] = None,
633
        trace_headers: Optional[Mapping[str, str]] = None,
634
        priority: int = 0,
635
    ) -> None:
Zhuohan Li's avatar
Zhuohan Li committed
636
        """Add a request to the engine's request pool.
637
638

        The request is added to the request pool and will be processed by the
Zhuohan Li's avatar
Zhuohan Li committed
639
        scheduler as `engine.step()` is called. The exact scheduling policy is
640
641
642
643
        determined by the scheduler.

        Args:
            request_id: The unique ID of the request.
644
645
            prompt: The prompt to the LLM. See
                [PromptType][vllm.inputs.PromptType]
646
                for more details about the format of each input.
647
            params: Parameters for sampling.
648
                [SamplingParams][vllm.SamplingParams] for text generation.
649
            arrival_time: The arrival time of the request. If None, we use
650
                the current monotonic time.
651
            lora_request: The LoRA request to add.
652
            trace_headers: OpenTelemetry trace headers.
653
654
            priority: The priority of the request.
                Only applicable with priority scheduling.
655
656
657
658

        Details:
            - Set arrival_time to the current time if it is None.
            - Set prompt_token_ids to the encoded prompt if it is None.
659
660
661
662
            - Create `n` number of [Sequence][vllm.sequence.Sequence] objects.
            - Create a [SequenceGroup][vllm.sequence.SequenceGroup] object
              from the list of [Sequence][vllm.sequence.Sequence].
            - Add the [SequenceGroup][vllm.sequence.SequenceGroup] object to the
663
              scheduler.
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679

        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
            >>> ...
680
        """
681
682
683
684
        if not isinstance(request_id, str):
            raise TypeError(
                f"request_id must be a string, got {type(request_id)}")

685
686
687
        if lora_request is not None and not self.lora_config:
            raise ValueError(f"Got lora_request {lora_request} but LoRA is "
                             "not enabled!")
688

689
        if priority != 0 and not self.scheduler_config.policy == "priority":
690
691
692
            raise ValueError(f"Got priority {priority} but "
                             "Priority scheduling is not enabled.")

693
        if isinstance(params, SamplingParams) \
694
            and params.logits_processors:
695
            raise ValueError(
696
                "Logits processors are not supported in multi-step decoding")
697

698
        if arrival_time is None:
699
            arrival_time = time.time()
700

701
702
703
704
705
        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
706

707
708
709
710
        #DEBUG anrongqiao
        if self.model_config.tokenizer_mode == "cpm":
            lora_request = None
            
711
        processed_inputs = self.input_preprocessor.preprocess(
712
            prompt,
713
            tokenization_kwargs=tokenization_kwargs,
714
            lora_request=lora_request,
715
        )
716

717
718
719
720
721
722
        self._add_processed_request(
            request_id=request_id,
            processed_inputs=processed_inputs,
            params=params,
            arrival_time=arrival_time,
            lora_request=lora_request,
723
            trace_headers=trace_headers,
724
            priority=priority,
725
        )
726
727
728
729
730
731

    def _create_sequence_group_with_sampling(
        self,
        request_id: str,
        seq: Sequence,
        sampling_params: SamplingParams,
732
733
        arrival_time: float,
        lora_request: Optional[LoRARequest],
734
        trace_headers: Optional[Mapping[str, str]] = None,
735
        encoder_seq: Optional[Sequence] = None,
736
        priority: int = 0,
737
738
739
740
741
742
743
744
745
746
    ) -> 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.")

747
748
749
        sampling_params = self._build_logits_processors(
            sampling_params, lora_request)

750
751
752
        # Defensive copy of SamplingParams, which are used by the sampler,
        # this doesn't deep-copy LogitsProcessor objects
        sampling_params = sampling_params.clone()
753

754
        sampling_params.update_from_generation_config(
755
            self.generation_config_fields, seq.eos_token_id)
756

757
        # Create the sequence group.
758
759
760
761
        draft_size = 1
        if self.vllm_config.speculative_config is not None:
            draft_size = \
                self.vllm_config.speculative_config.num_speculative_tokens + 1
762
763
764
765
766
767
768
769
770
        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)
771

772
773
        return seq_group

Antoni Baum's avatar
Antoni Baum committed
774
775
    def abort_request(self, request_id: Union[str, Iterable[str]]) -> None:
        """Aborts a request(s) with the given ID.
776
777

        Args:
Antoni Baum's avatar
Antoni Baum committed
778
            request_id: The ID(s) of the request to abort.
779
780

        Details:
781
            - Refer to [vllm.core.scheduler.Scheduler.abort_seq_group][].
782
783
784
785
786
787

        Example:
            >>> # initialize engine and add a request with request_id
            >>> request_id = str(0)
            >>> # abort the request
            >>> engine.abort_request(request_id)
788
        """
789
        for scheduler in self.scheduler:
790
791
            scheduler.abort_seq_group(
                request_id, seq_id_to_seq_group=self.seq_id_to_seq_group)
792

793
794
795
796
    def get_vllm_config(self) -> VllmConfig:
        """Gets the vllm configuration."""
        return self.vllm_config

797
798
799
800
    def get_model_config(self) -> ModelConfig:
        """Gets the model configuration."""
        return self.model_config

801
802
803
804
    def get_parallel_config(self) -> ParallelConfig:
        """Gets the parallel configuration."""
        return self.parallel_config

805
806
807
808
    def get_decoding_config(self) -> DecodingConfig:
        """Gets the decoding configuration."""
        return self.decoding_config

809
810
811
812
813
814
815
816
    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

817
    def get_num_unfinished_requests(self) -> int:
818
        """Gets the number of unfinished requests."""
819
820
        return sum(scheduler.get_num_unfinished_seq_groups()
                   for scheduler in self.scheduler)
821

822
    def has_unfinished_requests(self) -> bool:
823
        """Returns True if there are unfinished requests."""
824
825
826
827
828
829
830
831
832
        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()
833

834
835
    def reset_mm_cache(self) -> bool:
        """Reset the multi-modal cache."""
836
837
        self.input_preprocessor.clear_cache()
        return True
838

839
    def reset_prefix_cache(self, device: Optional[Device] = None) -> bool:
840
841
842
843
        """Reset prefix cache for all devices."""

        success = True
        for scheduler in self.scheduler:
844
            success = success and scheduler.reset_prefix_cache(device)
845
846
        return success

847
848
849
850
851
    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.
852

853
854
        ctx: The virtual engine context to work on
        request_id: If provided, then only this request is going to be processed
855
        """
856

857
        now = time.time()
858

859
        if len(ctx.output_queue) == 0:
860
861
            return None

862
        # Get pending async postprocessor
863
864
865
866
        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,
867
             is_last_step, is_first_step_output, skip) = ctx.output_queue[0]
868
869
        else:
            (outputs, seq_group_metadata_list, scheduler_outputs, is_async,
870
871
             is_last_step, is_first_step_output,
             skip) = ctx.output_queue.popleft()
872
873
874
875
876

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

877
        has_multiple_outputs: bool = len(outputs) > 1
878
        outputs_by_sequence_group: List[List[SequenceGroupOutput]]
879
880
        assert not has_multiple_outputs
        outputs_by_sequence_group = outputs
881

882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
        # 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

899
        finished_before: List[int] = []
900
        finished_now: List[int] = []
901
        empty_seq_indices: List[int] = []
902
903
904
905
906
        for i in indices:
            if i in skip:
                continue

            seq_group_meta = seq_group_metadata_list[i]
907
            scheduled_seq_group = scheduler_outputs.scheduled_seq_groups[i]
908

909
            seq_group: SequenceGroup = scheduled_seq_group.seq_group
910
911
912
913
914

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

915
            output: List[SequenceGroupOutput]
916
            if has_multiple_outputs:
917
918
919
920
                output = outputs_by_sequence_group[i]
            else:
                output = [outputs_by_sequence_group[0][i]]

921
            # tree style speculative decoding may generate empty output in first step
922
923
924
925
926
927
928
929
930
            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
931

932
            if not is_async:
933
934
                seq_group.update_num_computed_tokens(
                    seq_group_meta.token_chunk_size or 0)
935
936
937

            if outputs:
                for o in outputs:
938
939
940
941
                    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 += (
942
                                o.model_forward_time or 0)
943
944
945
946
947
                        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 += (
948
                                o.model_execute_time or 0)
949
950
951
                        else:
                            seq_group.metrics.model_execute_time = (
                                o.model_execute_time)
952

953
954
955
956
            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)
957

958
959
            if seq_group.is_finished():
                finished_now.append(i)
960

961
962
963
        # Generate outputs for the requests that finished this iteration
        for i in finished_now:
            scheduled_seq_group = scheduler_outputs.scheduled_seq_groups[i]
964

965
            seq_group = scheduled_seq_group.seq_group
966
            seq_group.maybe_set_first_token_time(now)
967
968
            if not seq_group.is_prefill():
                seq_group.set_last_token_time(now)
969
            request_output = RequestOutputFactory.create(
970
971
972
                seq_group,
                self.seq_id_to_seq_group,
                use_cache=self.use_cached_outputs)
973
974
            if request_output:
                ctx.request_outputs.append(request_output)
975

976
977
978
979
980
981
982
983
984
985
986
987
        # 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

988
989
990
991
992
993
        # Free currently finished requests
        if finished_now:
            for scheduler in self.scheduler:
                scheduler.free_finished_seq_groups()

        # Create the outputs
994
        for i in indices:
995
            if i in skip or i in finished_before or i in finished_now or i in empty_seq_indices:
996
997
                continue  # Avoids double processing

998
999
            scheduled_seq_group = scheduler_outputs.scheduled_seq_groups[i]

1000
            seq_group = scheduled_seq_group.seq_group
1001
            seq_group.maybe_set_first_token_time(now)
1002
1003
            if not seq_group.is_prefill():
                seq_group.set_last_token_time(now)
1004
            request_output = RequestOutputFactory.create(
1005
1006
1007
                seq_group,
                self.seq_id_to_seq_group,
                use_cache=self.use_cached_outputs)
1008
            if request_output:
1009
                ctx.request_outputs.append(request_output)
1010

1011
        # Create outputs only after processing the scheduler's results
1012

1013
        for seq_group in scheduler_outputs.ignored_seq_groups:
1014
1015
1016
1017
1018
            params = seq_group.sampling_params
            if params is not None and params.output_kind == (
                    RequestOutputKind.DELTA) and not seq_group.is_finished():
                continue

1019
            request_output = RequestOutputFactory.create(
1020
1021
1022
1023
                seq_group,
                self.seq_id_to_seq_group,
                use_cache=self.use_cached_outputs,
            )
1024
1025
            if request_output:
                ctx.request_outputs.append(request_output)
1026

1027
1028
1029
1030
        # 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)
1031
            ctx.request_outputs.clear()
1032

1033
1034
1035
1036
        # 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:
1037
            # Log stats.
1038
1039
            self.do_log_stats(scheduler_outputs, outputs, finished_before,
                              skip)
1040
1041

            # Tracing
1042
            self.do_tracing(scheduler_outputs, finished_before)
1043
1044
1045
1046

        return None

    def _advance_to_next_step(
1047
            self, output: SamplerOutput,
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
            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

1061
1062
1063
1064
            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)
1065
1066
1067
1068

            if seq_group_metadata.do_sample:
                assert len(sequence_group_outputs.samples) == 1, (
                    "Async output processor expects a single sample"
1069
                    " (i.e sampling_params.n == 1)")
1070
1071
1072
1073
                sample = sequence_group_outputs.samples[0]

                assert len(seq_group.seqs) == 1
                seq = seq_group.seqs[0]
1074

1075
1076
                seq.append_token_id(sample.output_token, sample.logprobs,
                                    sample.output_embed)
1077

1078
    def step(self) -> List[RequestOutput]:
Antoni Baum's avatar
Antoni Baum committed
1079
1080
        """Performs one decoding iteration and returns newly generated results.

1081
1082
1083
1084
        <figure markdown="span">
        ![Overview of the step function](https://i.imgur.com/sv2HssD.png)
        <figcaption>Overview of the step function</figcaption>
        </figure>
1085
1086

        Details:
1087
1088
        - Step 1: Schedules the sequences to be executed in the next
            iteration and the token blocks to be swapped in/out/copy.
1089

1090
1091
1092
1093
            - 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.
1094

1095
1096
        - Step 2: Calls the distributed executor to execute the model.
        - Step 3: Processes the model output. This mainly includes:
1097

1098
1099
1100
1101
            - 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.
1102

1103
        - Finally, it creates and returns the newly generated results.
1104
1105

        Example:
1106
1107
1108
1109
1110
1111
1112
        ```
        # 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))]
1113

1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
        # 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
1129
        """
1130
1131
1132
1133
        if self.parallel_config.pipeline_parallel_size > 1:
            raise NotImplementedError(
                "Pipeline parallelism is only supported through AsyncLLMEngine "
                "as performance will be severely degraded otherwise.")
1134

1135
        # For llm_engine, there is no pipeline parallel support, so the engine
1136
        # used is always 0.
1137
1138
        virtual_engine = 0

1139
1140
        # These are cached outputs from previous iterations. None if on first
        # iteration
1141
        cached_outputs = self.cached_scheduler_outputs[virtual_engine]
1142
1143
        seq_group_metadata_list = cached_outputs.seq_group_metadata_list
        scheduler_outputs = cached_outputs.scheduler_outputs
1144
        allow_async_output_proc = cached_outputs.allow_async_output_proc
1145

1146
1147
        ctx = self.scheduler_contexts[virtual_engine]

1148
1149
1150
        # Clear outputs for each new scheduler iteration
        ctx.request_outputs.clear()

1151
1152
1153
        # 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.
1154
1155
1156
1157
1158
        # 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:
1159
            # Schedule iteration
1160
            (seq_group_metadata_list, scheduler_outputs,
1161
1162
             allow_async_output_proc
             ) = self.scheduler[virtual_engine].schedule()
1163

1164
1165
            ctx.seq_group_metadata_list = seq_group_metadata_list
            ctx.scheduler_outputs = scheduler_outputs
1166

1167
1168
            finished_requests_ids = self.scheduler[
                virtual_engine].get_and_reset_finished_requests_ids()
1169
1170
1171
1172
1173
            # 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]
1174

1175
1176
            # Maybe switch from async mode to sync mode
            if not allow_async_output_proc and len(ctx.output_queue) > 0:
1177
                self._process_model_outputs(ctx=ctx)
1178

1179
1180
        else:
            finished_requests_ids = list()
1181
1182
1183

        assert seq_group_metadata_list is not None
        assert scheduler_outputs is not None
Antoni Baum's avatar
Antoni Baum committed
1184

1185
        if not scheduler_outputs.is_empty():
1186
1187
1188
1189
1190
1191

            # 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 = \
1192
                self._get_last_sampled_token_ids(virtual_engine)
1193

1194
            execute_model_req = ExecuteModelRequest(
1195
1196
1197
1198
                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,
1199
1200
                num_lookahead_slots=scheduler_outputs.num_lookahead_slots,
                running_queue_size=scheduler_outputs.running_queue_size,
1201
1202
1203
1204
1205
                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)

1206
            if allow_async_output_proc:
1207
1208
                execute_model_req.async_callback = self.async_callbacks[
                    virtual_engine]
1209

1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
            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
1227

1228
        else:
1229
1230
            # Nothing scheduled => If there is pending async postprocessor,
            # then finish it here.
1231
1232
            if len(ctx.output_queue) > 0:
                self._process_model_outputs(ctx=ctx)
1233
            # No outputs in this case
1234
            outputs = []
Antoni Baum's avatar
Antoni Baum committed
1235

1236
        if not self._has_remaining_steps(seq_group_metadata_list):
1237
            # is_first_step_output is True only when the num_steps of all
1238
            # the sequences are 1.
1239
1240
1241
            is_first_step_output: bool = False if not seq_group_metadata_list \
                else seq_group_metadata_list[0].state.num_steps == 1

1242
            # Add results to the output_queue
1243
1244
1245
1246
            ctx.append_output(outputs=outputs,
                              seq_group_metadata_list=seq_group_metadata_list,
                              scheduler_outputs=scheduler_outputs,
                              is_async=allow_async_output_proc,
1247
1248
                              is_last_step=True,
                              is_first_step_output=is_first_step_output)
1249
1250
1251

            if outputs and allow_async_output_proc:
                assert len(outputs) == 1, (
1252
                    "Async postprocessor expects only a single output set")
1253

1254
                self._advance_to_next_step(
1255
                    outputs[0], seq_group_metadata_list,
1256
                    scheduler_outputs.scheduled_seq_groups)
1257

1258
            # Check if need to run the usual non-async path
1259
            if not allow_async_output_proc:
1260
                self._process_model_outputs(ctx=ctx)
1261

1262
                # Log stats.
1263
                self.do_log_stats(scheduler_outputs, outputs)
1264

1265
1266
1267
                # Tracing
                self.do_tracing(scheduler_outputs)
        else:
1268
            # Multi-step case
1269
            return ctx.request_outputs
1270

1271
        if not self.has_unfinished_requests():
1272
1273
            # Drain async postprocessor (if exists)
            if len(ctx.output_queue) > 0:
1274
                self._process_model_outputs(ctx=ctx)
1275
            assert len(ctx.output_queue) == 0
1276

1277
1278
            # Stop the execute model loop in parallel workers until there are
            # more requests to process. This avoids waiting indefinitely in
1279
            # torch.distributed ops which may otherwise time out, and unblocks
1280
1281
            # the RPC thread in the workers so that they can process any other
            # queued control plane messages, such as add/remove lora adapters.
1282
            logger.debug("Stopping remote worker execution loop.")
1283
1284
            self.model_executor.stop_remote_worker_execution_loop()

1285
        return ctx.request_outputs
Antoni Baum's avatar
Antoni Baum committed
1286

1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
    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)

1319
1320
1321
    def _has_remaining_steps(
        self, seq_group_metadata_list: Optional[List[SequenceGroupMetadata]]
    ) -> bool:
1322
        return False
1323
1324
1325
1326

    def _cache_scheduler_outputs_for_multi_step(
            self, virtual_engine: int,
            seq_group_metadata_list: Optional[List[SequenceGroupMetadata]],
1327
1328
1329
1330
1331
1332
1333
1334
            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
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351

    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]:
        return None
Antoni Baum's avatar
Antoni Baum committed
1352

1353
    def add_logger(self, logger_name: str, logger: StatLoggerBase) -> None:
1354
1355
1356
1357
        if not self.log_stats:
            raise RuntimeError(
                "Stat logging is disabled. Set `disable_log_stats=False` "
                "argument to enable.")
1358
1359
1360
1361
1362
        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:
1363
1364
1365
1366
        if not self.log_stats:
            raise RuntimeError(
                "Stat logging is disabled. Set `disable_log_stats=False` "
                "argument to enable.")
1367
1368
1369
1370
        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]

1371
1372
1373
    def do_log_stats(self,
                     scheduler_outputs: Optional[SchedulerOutputs] = None,
                     model_output: Optional[List[SamplerOutput]] = None,
1374
1375
                     finished_before: Optional[List[int]] = None,
                     skip: Optional[List[int]] = None) -> None:
1376
1377
        """Forced log when no requests active."""
        if self.log_stats:
1378
            stats = self._get_stats(scheduler_outputs, model_output,
1379
                                    finished_before, skip)
1380
            for logger in self.stat_loggers.values():
1381
                logger.log(stats)
1382

1383
1384
1385
    def _get_stats(self,
                   scheduler_outputs: Optional[SchedulerOutputs],
                   model_output: Optional[List[SamplerOutput]] = None,
1386
1387
                   finished_before: Optional[List[int]] = None,
                   skip: Optional[List[int]] = None) -> Stats:
1388
1389
1390
1391
1392
1393
1394
        """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.
1395
1396
1397
1398
            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.
1399
        """
1400
        now = time.time()
Woosuk Kwon's avatar
Woosuk Kwon committed
1401

1402
1403
        # System State
        #   Scheduler State
1404
1405
1406
1407
1408
1409
        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)
1410
1411

        # KV Cache Usage in %
1412
        num_total_gpu = self.cache_config.num_gpu_blocks
1413
        gpu_cache_usage_sys = 0.
1414
        if num_total_gpu:  # Guard against both None and 0
1415
1416
1417
            num_free_gpu = sum(
                scheduler.block_manager.get_num_free_gpu_blocks()
                for scheduler in self.scheduler)
1418
            gpu_cache_usage_sys = 1.0 - (num_free_gpu / num_total_gpu)
Woosuk Kwon's avatar
Woosuk Kwon committed
1419

1420
        num_total_cpu = self.cache_config.num_cpu_blocks
1421
        cpu_cache_usage_sys = 0.
1422
        if num_total_cpu:  # Guard against both None and 0
1423
1424
1425
            num_free_cpu = sum(
                scheduler.block_manager.get_num_free_cpu_blocks()
                for scheduler in self.scheduler)
1426
1427
            cpu_cache_usage_sys = 1.0 - (num_free_cpu / num_total_cpu)

1428
1429
1430
1431
1432
1433
1434
        # 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)

1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
        # 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,
            )

1449
1450
1451
        # Iteration stats
        num_prompt_tokens_iter = 0
        num_generation_tokens_iter = 0
harrywu's avatar
harrywu committed
1452
        num_tokens_iter = 0
1453
1454
        time_to_first_tokens_iter: List[float] = []
        time_per_output_tokens_iter: List[float] = []
1455
1456
        num_preemption_iter = (0 if scheduler_outputs is None else
                               scheduler_outputs.preempted)
1457
1458
1459
1460

        # Request stats
        #   Latency
        time_e2e_requests: List[float] = []
harrywu's avatar
harrywu committed
1461
1462
1463
1464
        time_queue_requests: List[float] = []
        time_inference_requests: List[float] = []
        time_prefill_requests: List[float] = []
        time_decode_requests: List[float] = []
1465
1466
1467
1468
        #   Metadata
        num_prompt_tokens_requests: List[int] = []
        num_generation_tokens_requests: List[int] = []
        n_requests: List[int] = []
harrywu's avatar
harrywu committed
1469
        max_num_generation_tokens_requests: List[int] = []
1470
        max_tokens_requests: List[int] = []
1471
1472
        finished_reason_requests: List[str] = []

1473
        # LoRA requests
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
        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)

1492
1493
        # NOTE: This loop assumes prefill seq_groups are before
        # decode seq_groups in scheduled_seq_groups.
1494
        if scheduler_outputs is not None:
1495
1496
1497
1498
            # 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

1499
            num_generation_tokens_from_prefill_groups = 0
1500
1501
1502
1503
            # 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.
1504
1505
1506

            for idx, scheduled_seq_group in enumerate(
                    scheduler_outputs.scheduled_seq_groups):
1507
1508
1509
1510
1511
                # Skip double logging when using async output proc
                if finished_before and idx in finished_before:
                    actual_num_batched_tokens -= 1
                    continue

1512
1513
1514
1515
                # Currently, skip == preempted sequences, so we need to skip
                # their log stats
                if skip and idx in skip:
                    continue
1516

1517
                group_was_prefill = idx < scheduler_outputs.num_prefill_groups
1518
                seq_group = scheduled_seq_group.seq_group
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530

                # 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():
1531
                        latency = seq_group.get_last_token_latency()
1532
1533
1534
1535
1536
1537
1538
                        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.
1539
                    latency = seq_group.get_last_token_latency()
1540
                    time_per_output_tokens_iter.append(latency)
1541
1542
1543
1544
1545
1546
1547
1548
1549
                    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
1550
1551
1552
1553
1554
1555

                # 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.
1556
                if seq_group.is_finished():
1557
                    # Latency timings
1558
1559
                    time_e2e_requests.append(now -
                                             seq_group.metrics.arrival_time)
harrywu's avatar
harrywu committed
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
                    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)
1572
1573
1574
1575
1576
1577
1578
                    # 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
1579
1580
1581
                    max_num_generation_tokens_requests.append(
                        max(seq.get_output_len()
                            for seq in seq_group.get_seqs()))
1582
1583
                    if seq_group.sampling_params is not None:
                        n_requests.append(seq_group.sampling_params.n)
1584
1585
                        max_tokens_requests.append(
                            seq_group.sampling_params.max_tokens)
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
                    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 = (
1598
                actual_num_batched_tokens - num_prompt_tokens_iter +
1599
                num_generation_tokens_from_prefill_groups)
harrywu's avatar
harrywu committed
1600
1601
            num_tokens_iter = (num_generation_tokens_iter +
                               num_prompt_tokens_iter)
1602

1603
1604
        return Stats(
            now=now,
1605
1606
1607
1608
1609
1610
1611
1612
            # 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,
1613
1614
1615
            #   Prefix Cache Hit Rate
            cpu_prefix_cache_hit_rate=cpu_prefix_cache_hit_rate,
            gpu_prefix_cache_hit_rate=gpu_prefix_cache_hit_rate,
1616
1617
1618
1619

            # Iteration stats
            num_prompt_tokens_iter=num_prompt_tokens_iter,
            num_generation_tokens_iter=num_generation_tokens_iter,
harrywu's avatar
harrywu committed
1620
            num_tokens_iter=num_tokens_iter,
1621
1622
            time_to_first_tokens_iter=time_to_first_tokens_iter,
            time_per_output_tokens_iter=time_per_output_tokens_iter,
1623
            num_preemption_iter=num_preemption_iter,
1624
1625
1626
1627

            # Request stats
            #   Latency
            time_e2e_requests=time_e2e_requests,
harrywu's avatar
harrywu committed
1628
1629
1630
1631
            time_queue_requests=time_queue_requests,
            time_inference_requests=time_inference_requests,
            time_prefill_requests=time_prefill_requests,
            time_decode_requests=time_decode_requests,
1632
1633
1634
            #   Metadata
            num_prompt_tokens_requests=num_prompt_tokens_requests,
            num_generation_tokens_requests=num_generation_tokens_requests,
harrywu's avatar
harrywu committed
1635
1636
            max_num_generation_tokens_requests=
            max_num_generation_tokens_requests,
1637
            n_requests=n_requests,
1638
            max_tokens_requests=max_tokens_requests,
1639
            finished_reason_requests=finished_reason_requests,
1640
1641
1642
            max_lora=str(max_lora_stat),
            waiting_lora_adapters=list(waiting_lora_adapters.keys()),
            running_lora_adapters=list(running_lora_adapters.keys()))
1643

1644
    def add_lora(self, lora_request: LoRARequest) -> bool:
1645
        return self.model_executor.add_lora(lora_request)
1646
1647

    def remove_lora(self, lora_id: int) -> bool:
1648
        return self.model_executor.remove_lora(lora_id)
1649

1650
    def list_loras(self) -> Set[int]:
1651
        return self.model_executor.list_loras()
1652

1653
1654
1655
    def pin_lora(self, lora_id: int) -> bool:
        return self.model_executor.pin_lora(lora_id)

1656
1657
1658
1659
1660
1661
    def start_profile(self) -> None:
        self.model_executor.start_profile()

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

1662
1663
1664
1665
1666
    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)

1667
    def wake_up(self, tags: Optional[list[str]] = None) -> None:
1668
1669
        assert self.vllm_config.model_config.enable_sleep_mode, (
            "Sleep mode is not enabled in the model config")
1670
        self.model_executor.wake_up(tags)
1671

1672
1673
1674
    def is_sleeping(self) -> bool:
        return self.model_executor.is_sleeping

1675
    def check_health(self) -> None:
1676
        self.model_executor.check_health()
1677
1678
1679
1680

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

1681
1682
1683
    def do_tracing(self,
                   scheduler_outputs: SchedulerOutputs,
                   finished_before: Optional[List[int]] = None) -> None:
1684
1685
1686
        if self.tracer is None:
            return

1687
1688
1689
1690
1691
1692
        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

1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
            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
1710
1711
1712
1713
1714
1715
1716
1717

            # Handle potential None values for cancelled/aborted requests
            ttft = (metrics.first_token_time - metrics.arrival_time
                    if metrics.first_token_time is not None else None)

            e2e_time = (metrics.finished_time - metrics.arrival_time
                        if metrics.finished_time is not None else None)

1718
            seq_span.set_attribute(SpanAttributes.GEN_AI_RESPONSE_MODEL,
1719
                                   self.model_config.model)
1720
            seq_span.set_attribute(SpanAttributes.GEN_AI_REQUEST_ID,
1721
                                   seq_group.request_id)
1722
            seq_span.set_attribute(SpanAttributes.GEN_AI_REQUEST_TEMPERATURE,
1723
                                   seq_group.sampling_params.temperature)
1724
            seq_span.set_attribute(SpanAttributes.GEN_AI_REQUEST_TOP_P,
1725
                                   seq_group.sampling_params.top_p)
1726
            seq_span.set_attribute(SpanAttributes.GEN_AI_REQUEST_MAX_TOKENS,
1727
                                   seq_group.sampling_params.max_tokens)
1728
            seq_span.set_attribute(SpanAttributes.GEN_AI_REQUEST_N,
1729
                                   seq_group.sampling_params.n)
1730
            seq_span.set_attribute(SpanAttributes.GEN_AI_USAGE_NUM_SEQUENCES,
1731
                                   seq_group.num_seqs())
1732
            seq_span.set_attribute(SpanAttributes.GEN_AI_USAGE_PROMPT_TOKENS,
1733
1734
                                   len(seq_group.prompt_token_ids))
            seq_span.set_attribute(
1735
                SpanAttributes.GEN_AI_USAGE_COMPLETION_TOKENS,
1736
1737
1738
1739
                sum([
                    seq.get_output_len()
                    for seq in seq_group.get_finished_seqs()
                ]))
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751

            # Only set timing attributes if the values are available
            if metrics.time_in_queue is not None:
                seq_span.set_attribute(
                    SpanAttributes.GEN_AI_LATENCY_TIME_IN_QUEUE,
                    metrics.time_in_queue)
            if ttft is not None:
                seq_span.set_attribute(
                    SpanAttributes.GEN_AI_LATENCY_TIME_TO_FIRST_TOKEN, ttft)
            if e2e_time is not None:
                seq_span.set_attribute(SpanAttributes.GEN_AI_LATENCY_E2E,
                                       e2e_time)
1752
1753
            if metrics.scheduler_time is not None:
                seq_span.set_attribute(
1754
                    SpanAttributes.GEN_AI_LATENCY_TIME_IN_SCHEDULER,
1755
1756
1757
                    metrics.scheduler_time)
            if metrics.model_forward_time is not None:
                seq_span.set_attribute(
1758
                    SpanAttributes.GEN_AI_LATENCY_TIME_IN_MODEL_FORWARD,
1759
1760
1761
                    metrics.model_forward_time / 1000.0)
            if metrics.model_execute_time is not None:
                seq_span.set_attribute(
1762
                    SpanAttributes.GEN_AI_LATENCY_TIME_IN_MODEL_EXECUTE,
1763
                    metrics.model_execute_time)
1764

1765
    def _validate_model_inputs(self, inputs: ProcessorInputs,
1766
                               lora_request: Optional[LoRARequest]):
1767
1768
        encoder_inputs, decoder_inputs = split_enc_dec_inputs(inputs)

1769
1770
1771
1772
        if encoder_inputs is not None:
            self._validate_model_input(encoder_inputs,
                                       lora_request,
                                       prompt_type="encoder")
1773

1774
1775
1776
        self._validate_model_input(decoder_inputs,
                                   lora_request,
                                   prompt_type="decoder")
1777

1778
1779
1780
1781
1782
1783
1784
    def _validate_model_input(
        self,
        prompt_inputs: SingletonInputs,
        lora_request: Optional[LoRARequest],
        *,
        prompt_type: Literal["encoder", "decoder"],
    ):
1785
        model_config = self.model_config
zhuwenwen's avatar
zhuwenwen committed
1786
1787
1788
1789
1790
1791
1792
1793
        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))
1794

1795
        prompt_ids = prompt_inputs.get("prompt_token_ids", [])
1796
1797
1798
        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
1799
            elif prompt_inputs["type"] == "embeds":
1800
                pass
1801
1802
1803
            else:
                raise ValueError(f"The {prompt_type} prompt cannot be empty")

1804
1805
1806
1807
1808
1809
        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")

1810
        max_prompt_len = self.model_config.max_model_len
1811
        if len(prompt_ids) > max_prompt_len:
1812
            if prompt_type == "encoder" and model_config.is_multimodal_model:
1813
1814
                mm_registry = self.input_preprocessor.mm_registry
                mm_processor = mm_registry.create_processor(
1815
1816
1817
                    model_config,
                    tokenizer=tokenizer or object(),  # Dummy if no tokenizer
                )
1818
                assert isinstance(mm_processor, EncDecMultiModalProcessor)
1819

1820
                if mm_processor.pad_dummy_encoder_prompt:
汪志鹏's avatar
汪志鹏 committed
1821
                    return  # Skip encoder length check for Whisper and Donut
1822

1823
            if model_config.is_multimodal_model:
1824
                suggestion = (
1825
1826
1827
1828
                    "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.")
1829
1830
1831
1832
1833
1834
1835
1836
1837
            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}")
1838
1839
1840
1841

            # 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
1842
1843
1844
1845

    def _build_logits_processors(
            self, sampling_params: SamplingParams,
            lora_request: Optional[LoRARequest]) -> SamplingParams:
1846
1847
1848
1849
        """Constructs logits processors based on the 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."""
1850
1851

        logits_processors = []
1852

1853
1854
1855
        if (sampling_params.logit_bias or sampling_params.allowed_token_ids):
            tokenizer = self.get_tokenizer(lora_request=lora_request)

1856
            processors = get_openai_logits_processors(
1857
1858
1859
1860
1861
1862
1863
1864
1865
                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

1866
1867
1868
1869
1870
1871
        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)

1872
1873
1874
1875
1876
1877
1878
        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
1879

1880
1881
1882
1883
1884
1885
1886
1887
    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)

1888

1889
1890
1891
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