llm.py 74.7 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import itertools
5
from collections.abc import Sequence
6
from contextlib import contextmanager
7
8
from typing import (TYPE_CHECKING, Any, Callable, ClassVar, Optional, Union,
                    cast, overload)
9

10
import cloudpickle
11
import torch.nn as nn
12
from pydantic import ValidationError
13
from tqdm.auto import tqdm
14
from typing_extensions import TypeVar, deprecated
15

16
import vllm.envs as envs
17
from vllm.beam_search import (BeamSearchInstance, BeamSearchOutput,
18
19
                              BeamSearchSequence,
                              create_sort_beams_key_function)
20
21
from vllm.config import (CompilationConfig, ModelDType, TokenizerMode,
                         is_init_field)
22
23
from vllm.engine.arg_utils import (ConvertOption, EngineArgs, HfOverrides,
                                   PoolerConfig, RunnerOption)
Joe Runde's avatar
Joe Runde committed
24
from vllm.engine.llm_engine import LLMEngine
nunjunj's avatar
nunjunj committed
25
from vllm.entrypoints.chat_utils import (ChatCompletionMessageParam,
26
                                         ChatTemplateContentFormatOption,
27
28
                                         apply_hf_chat_template,
                                         apply_mistral_chat_template,
29
30
                                         parse_chat_messages,
                                         resolve_chat_template_content_format)
31
32
# yapf conflicts with isort for this block
# yapf: disable
33
34
35
36
from vllm.entrypoints.score_utils import (ScoreContentPartParam,
                                          ScoreMultiModalParam,
                                          _cosine_similarity,
                                          _validate_score_input_lens,
37
                                          compress_token_type_ids,
38
                                          get_score_prompt)
39
# yapf: enable
40
41
from vllm.entrypoints.utils import (_validate_truncation_size,
                                    log_non_default_args)
42
from vllm.inputs import PromptType, SingletonPrompt, TextPrompt, TokensPrompt
43
from vllm.inputs.parse import parse_and_batch_prompt
44
from vllm.logger import init_logger
45
from vllm.lora.request import LoRARequest
46
from vllm.model_executor.layers.quantization import QuantizationMethods
47
48
49
from vllm.outputs import (ClassificationRequestOutput, EmbeddingRequestOutput,
                          PoolingRequestOutput, RequestOutput,
                          ScoringRequestOutput)
50
from vllm.pooling_params import PoolingParams
51
52
from vllm.sampling_params import (BeamSearchParams, RequestOutputKind,
                                  SamplingParams)
53
from vllm.tasks import PoolingTask
54
from vllm.transformers_utils.tokenizer import (AnyTokenizer, MistralTokenizer,
55
                                               get_cached_tokenizer)
yhu422's avatar
yhu422 committed
56
from vllm.usage.usage_lib import UsageContext
57
from vllm.utils import Counter, Device, deprecate_kwargs, is_list_of
58

59
60
61
if TYPE_CHECKING:
    from vllm.v1.metrics.reader import Metric

62
63
logger = init_logger(__name__)

64
65
_R = TypeVar("_R", default=Any)

66
67

class LLM:
Woosuk Kwon's avatar
Woosuk Kwon committed
68
69
70
71
72
73
74
75
76
77
    """An LLM for generating texts from given prompts and sampling parameters.

    This class includes a tokenizer, a language model (possibly distributed
    across multiple GPUs), and GPU memory space allocated for intermediate
    states (aka KV cache). Given a batch of prompts and sampling parameters,
    this class generates texts from the model, using an intelligent batching
    mechanism and efficient memory management.

    Args:
        model: The name or path of a HuggingFace Transformers model.
78
        tokenizer: The name or path of a HuggingFace Transformers tokenizer.
79
80
        tokenizer_mode: The tokenizer mode. "auto" will use the fast tokenizer
            if available, and "slow" will always use the slow tokenizer.
81
82
83
        skip_tokenizer_init: If true, skip initialization of tokenizer and
            detokenizer. Expect valid prompt_token_ids and None for prompt
            from the input.
84
85
        trust_remote_code: Trust remote code (e.g., from HuggingFace) when
            downloading the model and tokenizer.
86
87
88
89
        allowed_local_media_path: Allowing API requests to read local images
            or videos from directories specified by the server file system.
            This is a security risk. Should only be enabled in trusted
            environments.
Woosuk Kwon's avatar
Woosuk Kwon committed
90
91
92
        tensor_parallel_size: The number of GPUs to use for distributed
            execution with tensor parallelism.
        dtype: The data type for the model weights and activations. Currently,
Woosuk Kwon's avatar
Woosuk Kwon committed
93
94
95
96
            we support `float32`, `float16`, and `bfloat16`. If `auto`, we use
            the `torch_dtype` attribute specified in the model config file.
            However, if the `torch_dtype` in the config is `float32`, we will
            use `float16` instead.
97
        quantization: The method used to quantize the model weights. Currently,
98
            we support "awq", "gptq", and "fp8" (experimental).
99
100
101
102
            If None, we first check the `quantization_config` attribute in the
            model config file. If that is None, we assume the model weights are
            not quantized and use `dtype` to determine the data type of
            the weights.
Jasmond L's avatar
Jasmond L committed
103
104
        revision: The specific model version to use. It can be a branch name,
            a tag name, or a commit id.
105
106
        tokenizer_revision: The specific tokenizer version to use. It can be a
            branch name, a tag name, or a commit id.
107
108
109
110
111
112
113
        seed: The seed to initialize the random number generator for sampling.
        gpu_memory_utilization: The ratio (between 0 and 1) of GPU memory to
            reserve for the model weights, activations, and KV cache. Higher
            values will increase the KV cache size and thus improve the model's
            throughput. However, if the value is too high, it may cause out-of-
            memory (OOM) errors.
        swap_space: The size (GiB) of CPU memory per GPU to use as swap space.
114
115
116
117
118
            This can be used for temporarily storing the states of the requests
            when their `best_of` sampling parameters are larger than 1. If all
            requests will have `best_of=1`, you can safely set this to 0.
            Noting that `best_of` is only supported in V0. Otherwise, too small
            values may cause out-of-memory (OOM) errors.
119
120
121
122
        cpu_offload_gb: The size (GiB) of CPU memory to use for offloading
            the model weights. This virtually increases the GPU memory space
            you can use to hold the model weights, at the cost of CPU-GPU data
            transfer for every forward pass.
123
124
125
        enforce_eager: Whether to enforce eager execution. If True, we will
            disable CUDA graph and always execute the model in eager mode.
            If False, we will use CUDA graph and eager execution in hybrid.
126
        max_seq_len_to_capture: Maximum sequence len covered by CUDA graphs.
127
            When a sequence has context length larger than this, we fall back
128
129
130
            to eager mode. Additionally for encoder-decoder models, if the
            sequence length of the encoder input is larger than this, we fall
            back to the eager mode.
131
132
        disable_custom_all_reduce: See
            [ParallelConfig][vllm.config.ParallelConfig].
133
134
        disable_async_output_proc: Disable async output processing.
            This may result in lower performance.
135
        hf_token: The token to use as HTTP bearer authorization for remote files
136
            . If `True`, will use the token generated when running
137
            `huggingface-cli login` (stored in `~/.huggingface`).
138
139
140
        hf_overrides: If a dictionary, contains arguments to be forwarded to the
            HuggingFace config. If a callable, it is called to update the
            HuggingFace config.
141
142
143
144
145
146
147
148
        mm_processor_kwargs: Arguments to be forwarded to the model's processor
            for multi-modal data, e.g., image processor. Overrides for the
            multi-modal processor obtained from `AutoProcessor.from_pretrained`.
            The available overrides depend on the model that is being run.
            For example, for Phi-3-Vision: `{"num_crops": 4}`.
        override_pooler_config: Initialize non-default pooling config or
            override default pooling config for the pooling model.
            e.g. `PoolerConfig(pooling_type="mean", normalize=False)`.
149
150
151
        compilation_config: Either an integer or a dictionary. If it is an
            integer, it is used as the level of compilation optimization. If it
            is a dictionary, it can specify the full compilation configuration.
152
        **kwargs: Arguments for [`EngineArgs`][vllm.EngineArgs].
nunjunj's avatar
nunjunj committed
153

154
155
    Note:
        This class is intended to be used for offline inference. For online
156
        serving, use the [AsyncLLMEngine][vllm.AsyncLLMEngine] class instead.
157
    """
158

159
    DEPRECATE_LEGACY: ClassVar[bool] = True
160
161
162
163
164
165
166
167
168
169
170
    """A flag to toggle whether to deprecate the legacy generate/encode API."""

    @classmethod
    @contextmanager
    def deprecate_legacy_api(cls):
        cls.DEPRECATE_LEGACY = True

        yield

        cls.DEPRECATE_LEGACY = False

171
172
173
    def __init__(
        self,
        model: str,
174
        *,
175
176
        runner: RunnerOption = "auto",
        convert: ConvertOption = "auto",
177
        tokenizer: Optional[str] = None,
178
        tokenizer_mode: TokenizerMode = "auto",
179
        skip_tokenizer_init: bool = False,
180
        trust_remote_code: bool = False,
181
        allowed_local_media_path: str = "",
182
        tensor_parallel_size: int = 1,
183
184
        dtype: ModelDType = "auto",
        quantization: Optional[QuantizationMethods] = None,
185
        revision: Optional[str] = None,
186
        tokenizer_revision: Optional[str] = None,
187
        seed: Optional[int] = None,
188
        gpu_memory_utilization: float = 0.9,
189
        swap_space: float = 4,
190
        cpu_offload_gb: float = 0,
191
        enforce_eager: bool = False,
192
        max_seq_len_to_capture: int = 8192,
193
        disable_custom_all_reduce: bool = False,
194
        disable_async_output_proc: bool = False,
195
        hf_token: Optional[Union[bool, str]] = None,
196
        hf_overrides: Optional[HfOverrides] = None,
197
        mm_processor_kwargs: Optional[dict[str, Any]] = None,
198
        override_pooler_config: Optional[PoolerConfig] = None,
199
200
        compilation_config: Optional[Union[int, dict[str, Any],
                                           CompilationConfig]] = None,
201
202
        **kwargs,
    ) -> None:
203
        """LLM constructor."""
204

205
206
        if "disable_log_stats" not in kwargs:
            kwargs["disable_log_stats"] = True
207

208
209
210
211
212
213
214
        if "worker_cls" in kwargs:
            worker_cls = kwargs["worker_cls"]
            # if the worker_cls is not qualified string name,
            # we serialize it using cloudpickle to avoid pickling issues
            if isinstance(worker_cls, type):
                kwargs["worker_cls"] = cloudpickle.dumps(worker_cls)

215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
        if "kv_transfer_config" in kwargs and isinstance(
                kwargs["kv_transfer_config"], dict):
            from vllm.config import KVTransferConfig
            raw_config_dict = kwargs["kv_transfer_config"]
            try:
                kwargs["kv_transfer_config"] = KVTransferConfig(
                    **raw_config_dict)
            except ValidationError as e:
                logger.error(
                    "Failed to convert 'kv_transfer_config' dict to "
                    "KVTransferConfig object. Dict: %s. Error: %s",
                    raw_config_dict, e)
                # Consider re-raising a more specific vLLM error or ValueError
                # to provide better context to the user.
                raise ValueError(
                    f"Invalid 'kv_transfer_config' provided: {e}") from e

232
233
234
        if hf_overrides is None:
            hf_overrides = {}

235
        if compilation_config is not None:
236
237
238
239
240
241
242
            if isinstance(compilation_config, int):
                compilation_config_instance = CompilationConfig(
                    level=compilation_config)
            elif isinstance(compilation_config, dict):
                predicate = lambda x: is_init_field(CompilationConfig, x[0])
                compilation_config_instance = CompilationConfig(
                    **dict(filter(predicate, compilation_config.items())))
243
244
            else:
                compilation_config_instance = compilation_config
245
        else:
246
            compilation_config_instance = CompilationConfig()
247

Zhuohan Li's avatar
Zhuohan Li committed
248
        engine_args = EngineArgs(
249
            model=model,
250
251
            runner=runner,
            convert=convert,
252
            tokenizer=tokenizer,
253
            tokenizer_mode=tokenizer_mode,
254
            skip_tokenizer_init=skip_tokenizer_init,
255
            trust_remote_code=trust_remote_code,
256
            allowed_local_media_path=allowed_local_media_path,
257
258
            tensor_parallel_size=tensor_parallel_size,
            dtype=dtype,
259
            quantization=quantization,
260
            revision=revision,
261
            tokenizer_revision=tokenizer_revision,
262
263
264
            seed=seed,
            gpu_memory_utilization=gpu_memory_utilization,
            swap_space=swap_space,
265
            cpu_offload_gb=cpu_offload_gb,
266
            enforce_eager=enforce_eager,
267
            max_seq_len_to_capture=max_seq_len_to_capture,
268
            disable_custom_all_reduce=disable_custom_all_reduce,
269
            disable_async_output_proc=disable_async_output_proc,
270
            hf_token=hf_token,
271
            hf_overrides=hf_overrides,
272
            mm_processor_kwargs=mm_processor_kwargs,
273
            override_pooler_config=override_pooler_config,
274
            compilation_config=compilation_config_instance,
275
276
            **kwargs,
        )
277

278
279
        log_non_default_args(engine_args)

280
281
282
283
        # Create the Engine (autoselects V0 vs V1)
        self.llm_engine = LLMEngine.from_engine_args(
            engine_args=engine_args, usage_context=UsageContext.LLM_CLASS)
        self.engine_class = type(self.llm_engine)
284

285
        self.request_counter = Counter()
286
        self.default_sampling_params: Union[dict[str, Any], None] = None
287

288
289
290
291
292
293
294
295
296
297
        if envs.VLLM_USE_V1:
            supported_tasks = self.llm_engine \
                .get_supported_tasks()  # type: ignore
        else:
            supported_tasks = self.llm_engine.model_config.supported_tasks

        logger.info("Supported_tasks: %s", supported_tasks)

        self.supported_tasks = supported_tasks

298
299
300
301
302
303
    def get_tokenizer(
        self,
        lora_request: Optional[LoRARequest] = None,
    ) -> AnyTokenizer:
        return self.llm_engine.get_tokenizer_group().get_lora_tokenizer(
            lora_request)
304
305

    def set_tokenizer(self, tokenizer: AnyTokenizer) -> None:
306
        tokenizer_group = self.llm_engine.get_tokenizer_group()
307

308
309
310
311
        # While CachedTokenizer is dynamic, have no choice but
        # compare class name. Misjudgment will arise from
        # user-defined tokenizer started with 'Cached'
        if tokenizer.__class__.__name__.startswith("Cached"):
312
            tokenizer_group.tokenizer = tokenizer
313
        else:
314
            tokenizer_group.tokenizer = get_cached_tokenizer(tokenizer)
315

316
    def get_default_sampling_params(self) -> SamplingParams:
317
318
319
320
321
        if self.default_sampling_params is None:
            self.default_sampling_params = (
                self.llm_engine.model_config.get_diff_sampling_param())
        if self.default_sampling_params:
            return SamplingParams.from_optional(**self.default_sampling_params)
322
323
        return SamplingParams()

324
325
326
327
328
329
330
    @overload
    def generate(
        self,
        prompts: Union[PromptType, Sequence[PromptType]],
        /,
        sampling_params: Optional[Union[SamplingParams,
                                        Sequence[SamplingParams]]] = None,
331
        *,
332
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
333
334
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
    ) -> list[RequestOutput]:
335
336
        ...

337
    @overload  # LEGACY: single (prompt + optional token ids)
338
    @deprecated("'prompt_token_ids' will become part of 'prompts'")
339
340
341
342
    def generate(
        self,
        prompts: str,
        sampling_params: Optional[Union[SamplingParams,
343
344
                                        list[SamplingParams]]] = None,
        prompt_token_ids: Optional[list[int]] = None,
345
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
346
347
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
    ) -> list[RequestOutput]:
348
349
350
        ...

    @overload  # LEGACY: multi (prompt + optional token ids)
351
    @deprecated("'prompt_token_ids' will become part of 'prompts'")
352
353
    def generate(
        self,
354
        prompts: list[str],
355
        sampling_params: Optional[Union[SamplingParams,
356
357
                                        list[SamplingParams]]] = None,
        prompt_token_ids: Optional[list[list[int]]] = None,
358
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
359
360
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
    ) -> list[RequestOutput]:
361
362
363
        ...

    @overload  # LEGACY: single (token ids + optional prompt)
364
    @deprecated("'prompt_token_ids' will become part of 'prompts'")
365
366
367
368
    def generate(
        self,
        prompts: Optional[str] = None,
        sampling_params: Optional[Union[SamplingParams,
369
                                        list[SamplingParams]]] = None,
370
        *,
371
        prompt_token_ids: list[int],
372
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
373
374
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
    ) -> list[RequestOutput]:
375
376
377
        ...

    @overload  # LEGACY: multi (token ids + optional prompt)
378
    @deprecated("'prompt_token_ids' will become part of 'prompts'")
379
380
    def generate(
        self,
381
        prompts: Optional[list[str]] = None,
382
        sampling_params: Optional[Union[SamplingParams,
383
                                        list[SamplingParams]]] = None,
384
        *,
385
        prompt_token_ids: list[list[int]],
386
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
387
388
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
    ) -> list[RequestOutput]:
389
390
391
        ...

    @overload  # LEGACY: single or multi token ids [pos-only]
392
    @deprecated("'prompt_token_ids' will become part of 'prompts'")
393
394
395
396
    def generate(
        self,
        prompts: None,
        sampling_params: None,
397
        prompt_token_ids: Union[list[int], list[list[int]]],
398
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
399
400
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
    ) -> list[RequestOutput]:
401
402
        ...

nunjunj's avatar
nunjunj committed
403
404
405
    @deprecate_kwargs(
        "prompt_token_ids",
        is_deprecated=lambda: LLM.DEPRECATE_LEGACY,
406
        additional_message="Please use the 'prompts' parameter instead.",
nunjunj's avatar
nunjunj committed
407
    )
408
409
    def generate(
        self,
410
        prompts: Union[Union[PromptType, Sequence[PromptType]],
411
                       Optional[Union[str, list[str]]]] = None,
412
413
        sampling_params: Optional[Union[SamplingParams,
                                        Sequence[SamplingParams]]] = None,
414
        prompt_token_ids: Optional[Union[list[int], list[list[int]]]] = None,
415
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
416
417
418
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
        priority: Optional[list[int]] = None,
    ) -> list[RequestOutput]:
Woosuk Kwon's avatar
Woosuk Kwon committed
419
420
        """Generates the completions for the input prompts.

421
        This class automatically batches the given prompts, considering
Woosuk Kwon's avatar
Woosuk Kwon committed
422
423
424
425
        the memory constraint. For the best performance, put all of your prompts
        into a single list and pass it to this method.

        Args:
426
            prompts: The prompts to the LLM. You may pass a sequence of prompts
427
                for batch inference. See [PromptType][vllm.inputs.PromptType]
428
                for more details about the format of each prompts.
Woosuk Kwon's avatar
Woosuk Kwon committed
429
            sampling_params: The sampling parameters for text generation. If
nunjunj's avatar
nunjunj committed
430
431
432
                None, we use the default sampling parameters.
                When it is a single value, it is applied to every prompt.
                When it is a list, the list must have the same length as the
433
                prompts and it is paired one by one with the prompt.
434
435
436
437
            use_tqdm: If `True`, shows a tqdm progress bar.
                If a callable (e.g., `functools.partial(tqdm, leave=False)`),
                it is used to create the progress bar.
                If `False`, no progress bar is created.
438
            lora_request: LoRA request to use for generation, if any.
439
440
            priority: The priority of the requests, if any.
                Only applicable when priority scheduling policy is enabled.
Woosuk Kwon's avatar
Woosuk Kwon committed
441
442

        Returns:
443
            A list of `RequestOutput` objects containing the
444
            generated completions in the same order as the input prompts.
445

446
447
448
449
        Note:
            Using `prompts` and `prompt_token_ids` as keyword parameters is
            considered legacy and may be deprecated in the future. You should
            instead pass them via the `inputs` parameter.
450
        """
451
452
453
        model_config = self.llm_engine.model_config
        runner_type = model_config.runner_type
        if runner_type != "generate":
454
455
456
457
            raise ValueError(
                "LLM.generate() is only supported for generative models. "
                "Try passing `--runner generate` to use the model as a "
                "generative model.")
458

459
        if prompt_token_ids is not None:
460
            parsed_prompts = self._convert_v1_inputs(
461
                prompts=cast(Optional[Union[str, list[str]]], prompts),
462
463
464
                prompt_token_ids=prompt_token_ids,
            )
        else:
465
466
            parsed_prompts = cast(Union[PromptType, Sequence[PromptType]],
                                  prompts)
467

468
469
        if sampling_params is None:
            # Use default sampling params.
470
            sampling_params = self.get_default_sampling_params()
471

472
473
474
475
        tokenization_kwargs: dict[str, Any] = {}
        truncate_prompt_tokens = None
        if isinstance(sampling_params, SamplingParams):
            truncate_prompt_tokens = sampling_params.truncate_prompt_tokens
476
477

        _validate_truncation_size(model_config.max_model_len,
478
479
                                  truncate_prompt_tokens, tokenization_kwargs)

480
481
482
483
        # Add any modality specific loras to the corresponding prompts
        lora_request = self._get_modality_specific_lora_reqs(
            parsed_prompts, lora_request)

484
        self._validate_and_add_requests(
485
            prompts=parsed_prompts,
486
            params=sampling_params,
487
            use_tqdm=use_tqdm,
488
            lora_request=lora_request,
489
            tokenization_kwargs=tokenization_kwargs,
490
491
            priority=priority,
        )
492

493
        outputs = self._run_engine(use_tqdm=use_tqdm)
Joe Runde's avatar
Joe Runde committed
494
        return self.engine_class.validate_outputs(outputs, RequestOutput)
495

496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
    def _get_modality_specific_lora_reqs(
            self, parsed_prompts: Union[PromptType, Sequence[PromptType]],
            lora_request: Optional[Union[list[LoRARequest], LoRARequest]]):
        # Grab the lora config off the vllm config on the engine,
        # since this is the same for both v0 & v1.
        lora_config = self.llm_engine.vllm_config.lora_config

        # If there's no lora config / default_mm_loras, or the model
        # isn't multimodal, leave the lora as is.
        if (lora_config is None
                or not self.llm_engine.model_config.is_multimodal_model
                or (lora_config and lora_config.default_mm_loras is None)):
            return lora_request

        if not isinstance(parsed_prompts, Sequence):
            parsed_prompts = [parsed_prompts]

        optional_loras = ([lora_request] * len(parsed_prompts)
                          if not isinstance(lora_request, Sequence) else
                          lora_request)

        return [
            self._resolve_single_prompt_mm_lora(
                parsed_prompt,
                opt_lora_req,
                lora_config.default_mm_loras,
            ) for parsed_prompt, opt_lora_req in zip(parsed_prompts,
                                                     optional_loras)
        ]

    def _resolve_single_prompt_mm_lora(self, parsed_prompt: PromptType,
                                       lora_request: Optional[LoRARequest],
                                       default_mm_loras: Optional[dict[str,
                                                                       str]]):
        if (not default_mm_loras or not isinstance(parsed_prompt, dict)
                or "multi_modal_data" not in parsed_prompt):
            return lora_request

        parsed_prompt = cast(Union[TextPrompt, TokensPrompt], parsed_prompt)

        intersection = set(
            parsed_prompt["multi_modal_data"].keys()).intersection(
                default_mm_loras.keys())
        if not intersection:
            return lora_request
        if len(intersection) > 1:
            # TODO: Would be nice to be able to have multiple loras per prompt
            logger.warning(
                "Multiple modality specific loras were registered and would be"
                " used by a single prompt consuming several modalities; "
                " currently we only support one lora per request; as such,"
                " lora(s) registered with modalities: %s"
                " will be skipped", intersection)
            return lora_request

        # Build the LoRA request; the ID of the default mm lora is the
        # index of the modality name sorted alphabetically + 1.
        modality_name = intersection.pop()
        modality_lora_path = default_mm_loras[modality_name]
        modality_lora_id = sorted(default_mm_loras).index(modality_name) + 1

        # If we have a collision, warn if there is a collision,
        # but always send the explicitly provided request.
        if lora_request:
            if lora_request.lora_int_id != modality_lora_id:
                logger.warning(
                    "A modality with a registered lora and a lora_request "
                    "with a different ID were provided; falling back to the "
                    "lora_request as we only apply one LoRARequest per prompt")
            return lora_request

        return LoRARequest(
            modality_name,
            modality_lora_id,
            modality_lora_path,
        )

573
    def collective_rpc(self,
574
                       method: Union[str, Callable[..., _R]],
575
                       timeout: Optional[float] = None,
576
577
                       args: tuple = (),
                       kwargs: Optional[dict[str, Any]] = None) -> list[_R]:
578
579
580
581
582
583
584
585
586
587
588
        """
        Execute an RPC call on all workers.

        Args:
            method: Name of the worker method to execute, or a callable that
                is serialized and sent to all workers to execute.

                If the method is a callable, it should accept an additional
                `self` argument, in addition to the arguments passed in `args`
                and `kwargs`. The `self` argument will be the worker object.
            timeout: Maximum time in seconds to wait for execution. Raises a
589
                [`TimeoutError`][] on timeout. `None` means wait indefinitely.
590
591
592
593
594
            args: Positional arguments to pass to the worker method.
            kwargs: Keyword arguments to pass to the worker method.

        Returns:
            A list containing the results from each worker.
595

596
597
598
        Note:
            It is recommended to use this API to only pass control messages,
            and set up data-plane communication to pass data.
599
        """
600
601

        return self.llm_engine.collective_rpc(method, timeout, args, kwargs)
602
603

    def apply_model(self, func: Callable[[nn.Module], _R]) -> list[_R]:
604
        """
605
606
        Run a function directly on the model inside each worker,
        returning the result for each of them.
607
        """
608
609
        executor = self.llm_engine.model_executor
        return executor.apply_model(func)
610

611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
    def _get_beam_search_lora_requests(
        self,
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]],
        prompts: list[Union[TokensPrompt, TextPrompt]],
    ) -> list[Optional[LoRARequest]]:
        """Get the optional lora request corresponding to each prompt."""
        if isinstance(lora_request,
                      Sequence) and len(lora_request) != len(prompts):
            raise ValueError(
                "Lora request list should be the same length as the prompts")

        if lora_request is None or isinstance(lora_request, LoRARequest):
            return [lora_request] * len(prompts)

        raise TypeError(f"Invalid lora_request type {type(lora_request)}")

627
628
    def beam_search(
        self,
629
        prompts: list[Union[TokensPrompt, TextPrompt]],
630
        params: BeamSearchParams,
631
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
632
        use_tqdm: bool = False,
633
    ) -> list[BeamSearchOutput]:
634
635
636
637
638
639
        """
        Generate sequences using beam search.

        Args:
            prompts: A list of prompts. Each prompt can be a string or a list
                of token IDs.
640
            params: The beam search parameters.
641
            lora_request: LoRA request to use for generation, if any.
642
            use_tqdm: Whether to use tqdm to display the progress bar.
643
        """
644
645
        # TODO: how does beam search work together with length penalty,
        # frequency, penalty, and stopping criteria, etc.?
646
647
648
649
        beam_width = params.beam_width
        max_tokens = params.max_tokens
        temperature = params.temperature
        ignore_eos = params.ignore_eos
650
651
        length_penalty = params.length_penalty

652
653
654
        lora_requests = self._get_beam_search_lora_requests(
            lora_request, prompts)

655
656
657
658
659
        tokenizer = self.get_tokenizer()
        sort_beams_key = create_sort_beams_key_function(
            tokenizer.eos_token_id,
            length_penalty,
        )
660

661
662
663
664
665
666
667
668
669
670
671
672
        def create_tokens_prompt_from_beam(
                beam: BeamSearchSequence) -> TokensPrompt:
            token_prompt_kwargs: TokensPrompt = {
                "prompt_token_ids": beam.tokens
            }
            if beam.multi_modal_data is not None:
                token_prompt_kwargs["multi_modal_data"] = beam.multi_modal_data

            if beam.mm_processor_kwargs is not None:
                token_prompt_kwargs[
                    "mm_processor_kwargs"] = beam.mm_processor_kwargs
            return TokensPrompt(**token_prompt_kwargs)
673

674
675
676
677
678
        # generate 2 * beam_width candidates at each step
        # following the huggingface transformers implementation
        # at https://github.com/huggingface/transformers/blob/e15687fffe5c9d20598a19aeab721ae0a7580f8a/src/transformers/generation/beam_search.py#L534 # noqa
        beam_search_params = SamplingParams(logprobs=2 * beam_width,
                                            max_tokens=1,
679
                                            temperature=temperature)
680
        instances: list[BeamSearchInstance] = []
681

682
        for lora_req, prompt in zip(lora_requests, prompts):
683
684
685
686
687
688
689
690
            # Add multimodal processor kwargs & data
            mm_kwargs = {}
            if "multi_modal_data" in prompt:
                mm_kwargs["multi_modal_data"] = prompt["multi_modal_data"]
            if "mm_processor_kwargs" in prompt:
                mm_kwargs["mm_processor_kwargs"] = prompt[
                    "mm_processor_kwargs"]

691
692
            if "prompt_token_ids" in prompt:
                prompt = cast(TokensPrompt, prompt)  # Needed for mypy
693
694
695
                prompt_tokens = prompt["prompt_token_ids"]
            else:
                prompt_tokens = tokenizer.encode(prompt["prompt"])
696

697
            instances.append(
698
699
700
701
702
703
                BeamSearchInstance(
                    prompt_tokens,
                    lora_request=lora_req,
                    logprobs=None,
                    **mm_kwargs,
                ), )
704

705
706
707
708
709
710
711
712
713
714
715
716
        token_iter = range(max_tokens)
        if use_tqdm:
            token_iter = tqdm(token_iter,
                              desc="Beam search",
                              unit="token",
                              unit_scale=False)
            logger.warning(
                "The progress bar shows the upper bound on token steps and "
                "may finish early due to stopping conditions. It does not "
                "reflect instance-level progress.")

        for _ in token_iter:
717
            all_beams: list[BeamSearchSequence] = list(
718
719
720
721
                sum((instance.beams for instance in instances), []))
            pos = [0] + list(
                itertools.accumulate(
                    len(instance.beams) for instance in instances))
722
            instance_start_and_end: list[tuple[int, int]] = list(
723
724
725
726
727
                zip(pos[:-1], pos[1:]))

            if len(all_beams) == 0:
                break

728
729
730
731
            # create the corresponding batch entries for prompt & optional lora
            prompts_batch, lora_req_batch = zip(
                *[(create_tokens_prompt_from_beam(beam), beam.lora_request)
                  for beam in all_beams])
732
733
734
735
736

            # only runs for one step
            # we don't need to use tqdm here
            output = self.generate(prompts_batch,
                                   sampling_params=beam_search_params,
737
738
                                   use_tqdm=False,
                                   lora_request=lora_req_batch)
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754

            for (start, end), instance in zip(instance_start_and_end,
                                              instances):
                instance_new_beams = []
                for i in range(start, end):
                    current_beam = all_beams[i]
                    result = output[i]

                    if result.outputs[0].logprobs is not None:
                        # if `result.outputs[0].logprobs` is None, it means
                        # the sequence is completed because of the max-model-len
                        # or abortion. we don't need to add it to the new beams.
                        logprobs = result.outputs[0].logprobs[0]
                        for token_id, logprob_obj in logprobs.items():
                            new_beam = BeamSearchSequence(
                                tokens=current_beam.tokens + [token_id],
755
                                logprobs=current_beam.logprobs + [logprobs],
756
                                lora_request=current_beam.lora_request,
757
                                cum_logprob=current_beam.cum_logprob +
758
759
760
761
                                logprob_obj.logprob,
                                multi_modal_data=current_beam.multi_modal_data,
                                mm_processor_kwargs=current_beam.
                                mm_processor_kwargs)
762
763
764
765
766
767
768

                            if token_id == tokenizer.eos_token_id and \
                                not ignore_eos:
                                instance.completed.append(new_beam)
                            else:
                                instance_new_beams.append(new_beam)
                sorted_beams = sorted(instance_new_beams,
769
                                      key=sort_beams_key,
770
771
772
773
774
775
776
                                      reverse=True)
                instance.beams = sorted_beams[:beam_width]

        outputs = []
        for instance in instances:
            instance.completed.extend(instance.beams)
            sorted_completed = sorted(instance.completed,
777
                                      key=sort_beams_key,
778
779
780
781
782
783
784
785
786
                                      reverse=True)
            best_beams = sorted_completed[:beam_width]

            for beam in best_beams:
                beam.text = tokenizer.decode(beam.tokens)
            outputs.append(BeamSearchOutput(sequences=best_beams))

        return outputs

nunjunj's avatar
nunjunj committed
787
788
    def chat(
        self,
789
790
        messages: Union[list[ChatCompletionMessageParam],
                        list[list[ChatCompletionMessageParam]]],
nunjunj's avatar
nunjunj committed
791
        sampling_params: Optional[Union[SamplingParams,
792
                                        list[SamplingParams]]] = None,
793
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
nunjunj's avatar
nunjunj committed
794
795
        lora_request: Optional[LoRARequest] = None,
        chat_template: Optional[str] = None,
796
        chat_template_content_format: ChatTemplateContentFormatOption = "auto",
797
        add_generation_prompt: bool = True,
798
        continue_final_message: bool = False,
799
        tools: Optional[list[dict[str, Any]]] = None,
800
        chat_template_kwargs: Optional[dict[str, Any]] = None,
801
802
        mm_processor_kwargs: Optional[dict[str, Any]] = None,
    ) -> list[RequestOutput]:
nunjunj's avatar
nunjunj committed
803
        """
804
        Generate responses for a chat conversation.
nunjunj's avatar
nunjunj committed
805

806
        The chat conversation is converted into a text prompt using the
807
        tokenizer and calls the [generate][] method to generate the
808
809
810
811
        responses.

        Multi-modal inputs can be passed in the same way you would pass them
        to the OpenAI API.
nunjunj's avatar
nunjunj committed
812
813

        Args:
814
815
            messages: A list of conversations or a single conversation.

816
817
                - Each conversation is represented as a list of messages.
                - Each message is a dictionary with 'role' and 'content' keys.
818

nunjunj's avatar
nunjunj committed
819
820
821
822
823
            sampling_params: The sampling parameters for text generation.
                If None, we use the default sampling parameters. When it
                is a single value, it is applied to every prompt. When it
                is a list, the list must have the same length as the
                prompts and it is paired one by one with the prompt.
824
825
826
827
            use_tqdm: If `True`, shows a tqdm progress bar.
                If a callable (e.g., `functools.partial(tqdm, leave=False)`),
                it is used to create the progress bar.
                If `False`, no progress bar is created.
nunjunj's avatar
nunjunj committed
828
829
            lora_request: LoRA request to use for generation, if any.
            chat_template: The template to use for structuring the chat.
830
                If not provided, the model's default chat template will be used.
831
832
            chat_template_content_format: The format to render message content.

833
834
835
836
837
                - "string" will render the content as a string.
                  Example: `"Who are you?"`
                - "openai" will render the content as a list of dictionaries,
                  similar to OpenAI schema.
                  Example: `[{"type": "text", "text": "Who are you?"}]`
838

839
            add_generation_prompt: If True, adds a generation template
nunjunj's avatar
nunjunj committed
840
                to each message.
841
            continue_final_message: If True, continues the final message in
842
                the conversation instead of starting a new one. Cannot be
843
                `True` if `add_generation_prompt` is also `True`.
844
845
            chat_template_kwargs: Additional kwargs to pass to the chat
                template.
846
847
            mm_processor_kwargs: Multimodal processor kwarg overrides for this
                chat request. Only used for offline requests.
nunjunj's avatar
nunjunj committed
848
849

        Returns:
850
            A list of `RequestOutput` objects containing the generated
nunjunj's avatar
nunjunj committed
851
852
            responses in the same order as the input messages.
        """
853
        list_of_messages: list[list[ChatCompletionMessageParam]]
nunjunj's avatar
nunjunj committed
854

855
856
        # Handle multi and single conversations
        if is_list_of(messages, list):
857
858
            # messages is list[list[...]]
            list_of_messages = cast(list[list[ChatCompletionMessageParam]],
859
                                    messages)
860
        else:
861
            # messages is list[...]
862
            list_of_messages = [
863
                cast(list[ChatCompletionMessageParam], messages)
864
            ]
865

866
        tokenizer = self.get_tokenizer(lora_request)
867
868
869
        model_config = self.llm_engine.get_model_config()
        resolved_content_format = resolve_chat_template_content_format(
            chat_template,
870
            tools,
871
872
            chat_template_content_format,
            tokenizer,
873
            model_config=model_config,
874
875
        )

876
877
878
879
880
881
882
883
        _chat_template_kwargs: dict[str, Any] = dict(
            chat_template=chat_template,
            add_generation_prompt=add_generation_prompt,
            continue_final_message=continue_final_message,
            tools=tools,
        )
        _chat_template_kwargs.update(chat_template_kwargs or {})

884
        prompts: list[Union[TokensPrompt, TextPrompt]] = []
885
886

        for msgs in list_of_messages:
887
888
889
            # NOTE: _parse_chat_message_content_parts() currently doesn't
            # handle mm_processor_kwargs, since there is no implementation in
            # the chat message parsing for it.
890
            conversation, mm_data = parse_chat_messages(
891
892
893
894
895
                msgs,
                model_config,
                tokenizer,
                content_format=resolved_content_format,
            )
896
897

            if isinstance(tokenizer, MistralTokenizer):
898
                prompt_token_ids = apply_mistral_chat_template(
899
900
                    tokenizer,
                    messages=msgs,
901
                    **_chat_template_kwargs,
902
903
                )
            else:
904
                prompt_str = apply_hf_chat_template(
905
                    tokenizer=tokenizer,
906
                    conversation=conversation,
907
                    model_config=model_config,
908
                    **_chat_template_kwargs,
909
                )
910
911
912
913
                # Special tokens are already included in chat templates so
                # should not be added by the tokenizer in this case.
                prompt_token_ids = tokenizer.encode(prompt_str,
                                                    add_special_tokens=False)
914

915
            prompt = TokensPrompt(prompt_token_ids=prompt_token_ids)
916
917
918
919

            if mm_data is not None:
                prompt["multi_modal_data"] = mm_data

920
921
922
            if mm_processor_kwargs is not None:
                prompt["mm_processor_kwargs"] = mm_processor_kwargs

923
            prompts.append(prompt)
924

nunjunj's avatar
nunjunj committed
925
        return self.generate(
926
            prompts,
927
            sampling_params=sampling_params,
nunjunj's avatar
nunjunj committed
928
929
930
931
            use_tqdm=use_tqdm,
            lora_request=lora_request,
        )

932
933
934
935
936
937
938
    @overload
    def encode(
        self,
        prompts: Union[PromptType, Sequence[PromptType]],
        /,
        pooling_params: Optional[Union[PoolingParams,
                                       Sequence[PoolingParams]]] = None,
939
        *,
940
        truncate_prompt_tokens: Optional[int] = None,
941
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
942
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
943
        pooling_task: PoolingTask = "encode",
944
        tokenization_kwargs: Optional[dict[str, Any]] = None,
945
    ) -> list[PoolingRequestOutput]:
946
947
        ...

948
    @overload  # LEGACY: single (prompt + optional token ids)
949
    @deprecated("'prompt_token_ids' will become part of 'prompts'")
950
951
952
953
954
    def encode(
        self,
        prompts: str,
        pooling_params: Optional[Union[PoolingParams,
                                       Sequence[PoolingParams]]] = None,
955
        prompt_token_ids: Optional[list[int]] = None,
956
        truncate_prompt_tokens: Optional[int] = None,
957
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
958
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
959
        pooling_task: PoolingTask = "encode",
960
        tokenization_kwargs: Optional[dict[str, Any]] = None,
961
    ) -> list[PoolingRequestOutput]:
962
        ...
963

964
    @overload  # LEGACY: multi (prompt + optional token ids)
965
    @deprecated("'prompt_token_ids' will become part of 'prompts'")
966
967
    def encode(
        self,
968
        prompts: list[str],
969
        pooling_params: Optional[Union[PoolingParams,
970
                                       Sequence[PoolingParams]]] = None,
971
        prompt_token_ids: Optional[list[list[int]]] = None,
972
        truncate_prompt_tokens: Optional[int] = None,
973
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
974
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
975
        pooling_task: PoolingTask = "encode",
976
        tokenization_kwargs: Optional[dict[str, Any]] = None,
977
    ) -> list[PoolingRequestOutput]:
978
979
980
        ...

    @overload  # LEGACY: single (token ids + optional prompt)
981
    @deprecated("'prompt_token_ids' will become part of 'prompts'")
982
983
984
985
986
987
    def encode(
        self,
        prompts: Optional[str] = None,
        pooling_params: Optional[Union[PoolingParams,
                                       Sequence[PoolingParams]]] = None,
        *,
988
        prompt_token_ids: list[int],
989
        truncate_prompt_tokens: Optional[int] = None,
990
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
991
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
992
        pooling_task: PoolingTask = "encode",
993
        tokenization_kwargs: Optional[dict[str, Any]] = None,
994
    ) -> list[PoolingRequestOutput]:
995
996
997
        ...

    @overload  # LEGACY: multi (token ids + optional prompt)
998
    @deprecated("'prompt_token_ids' will become part of 'prompts'")
999
1000
    def encode(
        self,
1001
        prompts: Optional[list[str]] = None,
1002
1003
1004
        pooling_params: Optional[Union[PoolingParams,
                                       Sequence[PoolingParams]]] = None,
        *,
1005
        prompt_token_ids: list[list[int]],
1006
        truncate_prompt_tokens: Optional[int] = None,
1007
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
1008
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
1009
        pooling_task: PoolingTask = "encode",
1010
        tokenization_kwargs: Optional[dict[str, Any]] = None,
1011
    ) -> list[PoolingRequestOutput]:
1012
1013
1014
        ...

    @overload  # LEGACY: single or multi token ids [pos-only]
1015
    @deprecated("'prompt_token_ids' will become part of 'prompts'")
1016
1017
1018
1019
    def encode(
        self,
        prompts: None,
        pooling_params: None,
1020
        prompt_token_ids: Union[list[int], list[list[int]]],
1021
        truncate_prompt_tokens: Optional[int] = None,
1022
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
1023
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
1024
        pooling_task: PoolingTask = "encode",
1025
        tokenization_kwargs: Optional[dict[str, Any]] = None,
1026
    ) -> list[PoolingRequestOutput]:
1027
1028
        ...

nunjunj's avatar
nunjunj committed
1029
1030
1031
    @deprecate_kwargs(
        "prompt_token_ids",
        is_deprecated=lambda: LLM.DEPRECATE_LEGACY,
1032
        additional_message="Please use the 'prompts' parameter instead.",
nunjunj's avatar
nunjunj committed
1033
    )
1034
1035
    def encode(
        self,
1036
        prompts: Union[Union[PromptType, Sequence[PromptType]],
1037
                       Optional[Union[str, list[str]]]] = None,
1038
1039
        pooling_params: Optional[Union[PoolingParams,
                                       Sequence[PoolingParams]]] = None,
1040
        prompt_token_ids: Optional[Union[list[int], list[list[int]]]] = None,
1041
        truncate_prompt_tokens: Optional[int] = None,
1042
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
1043
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
1044
        pooling_task: Optional[PoolingTask] = None,
1045
        tokenization_kwargs: Optional[dict[str, Any]] = None,
1046
    ) -> list[PoolingRequestOutput]:
1047
1048
        """Apply pooling to the hidden states corresponding to the input
        prompts.
1049

1050
        This class automatically batches the given prompts, considering
1051
1052
1053
1054
        the memory constraint. For the best performance, put all of your prompts
        into a single list and pass it to this method.

        Args:
1055
            prompts: The prompts to the LLM. You may pass a sequence of prompts
1056
                for batch inference. See [PromptType][vllm.inputs.PromptType]
1057
                for more details about the format of each prompts.
1058
1059
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1060
1061
1062
1063
            use_tqdm: If `True`, shows a tqdm progress bar.
                If a callable (e.g., `functools.partial(tqdm, leave=False)`),
                it is used to create the progress bar.
                If `False`, no progress bar is created.
1064
            lora_request: LoRA request to use for generation, if any.
1065
            pooling_task: Override the pooling task to use.
1066
1067

        Returns:
1068
            A list of `PoolingRequestOutput` objects containing the
1069
            pooled hidden states in the same order as the input prompts.
1070

1071
1072
1073
1074
        Note:
            Using `prompts` and `prompt_token_ids` as keyword parameters is
            considered legacy and may be deprecated in the future. You should
            instead pass them via the `inputs` parameter.
1075
        """
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
        if pooling_task is None:
            if "embed" in self.supported_tasks:
                pooling_task = "embed"
            else:
                pooling_task = "encode"

            logger.warning_once(
                "`LLM.encode` is currently using `pooling_task = %s`.\n"
                "Please use one of the more specific methods or set the "
                "task directly when using `LLM.encode`:\n"
                "  - For embeddings, use `LLM.embed(...)` "
                "or `pooling_task=\"embed\"`.\n"
                "  - For classification logits, use `LLM.classify(...)` "
                "or `pooling_task=\"classify\"`.\n"
                "  - For rewards, use `LLM.reward(...)` "
                "or `pooling_task=\"reward\"`\n"
                "  - For similarity scores, use `LLM.score(...)`.",
                pooling_task)

1095
1096
        model_config = self.llm_engine.model_config
        runner_type = model_config.runner_type
1097
        if runner_type != "pooling":
1098
1099
1100
1101
            raise ValueError(
                "LLM.encode() is only supported for pooling models. "
                "Try passing `--runner pooling` to use the model as a "
                "pooling model.")
1102

1103
        if prompt_token_ids is not None:
1104
            parsed_prompts = self._convert_v1_inputs(
1105
                prompts=cast(Optional[Union[str, list[str]]], prompts),
1106
1107
1108
                prompt_token_ids=prompt_token_ids,
            )
        else:
1109
1110
            parsed_prompts = cast(Union[PromptType, Sequence[PromptType]],
                                  prompts)
1111

1112
1113
1114
        if pooling_params is None:
            # Use default pooling params.
            pooling_params = PoolingParams()
1115
1116
1117

        if isinstance(pooling_params, PoolingParams):
            pooling_params.verify(pooling_task, model_config)
1118
1119
        else:
            for pooling_param in pooling_params:
1120
                pooling_param.verify(pooling_task, model_config)
1121

1122
1123
1124
1125
1126
        if tokenization_kwargs is None:
            tokenization_kwargs = dict[str, Any]()
            _validate_truncation_size(model_config.max_model_len,
                                      truncate_prompt_tokens,
                                      tokenization_kwargs)
1127

1128
        self._validate_and_add_requests(
1129
            prompts=parsed_prompts,
1130
            params=pooling_params,
1131
            use_tqdm=use_tqdm,
1132
            lora_request=lora_request,
1133
            tokenization_kwargs=tokenization_kwargs,
1134
1135
        )

1136
        outputs = self._run_engine(use_tqdm=use_tqdm)
Joe Runde's avatar
Joe Runde committed
1137
        return self.engine_class.validate_outputs(outputs,
1138
                                                  PoolingRequestOutput)
1139

1140
1141
1142
1143
1144
    def embed(
        self,
        prompts: Union[PromptType, Sequence[PromptType]],
        /,
        *,
1145
        truncate_prompt_tokens: Optional[int] = None,
1146
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
1147
1148
        pooling_params: Optional[Union[PoolingParams,
                                       Sequence[PoolingParams]]] = None,
1149
1150
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
    ) -> list[EmbeddingRequestOutput]:
1151
1152
1153
1154
1155
1156
1157
1158
1159
        """
        Generate an embedding vector for each prompt.

        This class automatically batches the given prompts, considering
        the memory constraint. For the best performance, put all of your prompts
        into a single list and pass it to this method.

        Args:
            prompts: The prompts to the LLM. You may pass a sequence of prompts
1160
                for batch inference. See [PromptType][vllm.inputs.PromptType]
1161
                for more details about the format of each prompts.
1162
1163
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1164
1165
1166
1167
            use_tqdm: If `True`, shows a tqdm progress bar.
                If a callable (e.g., `functools.partial(tqdm, leave=False)`),
                it is used to create the progress bar.
                If `False`, no progress bar is created.
1168
1169
1170
            lora_request: LoRA request to use for generation, if any.

        Returns:
1171
            A list of `EmbeddingRequestOutput` objects containing the
1172
1173
            embedding vectors in the same order as the input prompts.
        """
1174
        if "embed" not in self.supported_tasks:
1175
1176
1177
            raise ValueError(
                "Embedding API is not supported by this model. "
                "Try converting the model using `--convert embed`.")
1178

1179
1180
1181
1182
1183
1184
1185
1186
        items = self.encode(
            prompts,
            truncate_prompt_tokens=truncate_prompt_tokens,
            use_tqdm=use_tqdm,
            pooling_params=pooling_params,
            lora_request=lora_request,
            pooling_task="embed",
        )
1187
1188
1189
1190
1191
1192
1193
1194

        return [EmbeddingRequestOutput.from_base(item) for item in items]

    def classify(
        self,
        prompts: Union[PromptType, Sequence[PromptType]],
        /,
        *,
1195
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
1196
1197
        pooling_params: Optional[Union[PoolingParams,
                                       Sequence[PoolingParams]]] = None,
1198
1199
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
    ) -> list[ClassificationRequestOutput]:
1200
1201
1202
1203
1204
1205
1206
1207
1208
        """
        Generate class logits for each prompt.

        This class automatically batches the given prompts, considering
        the memory constraint. For the best performance, put all of your prompts
        into a single list and pass it to this method.

        Args:
            prompts: The prompts to the LLM. You may pass a sequence of prompts
1209
                for batch inference. See [PromptType][vllm.inputs.PromptType]
1210
                for more details about the format of each prompts.
1211
1212
1213
1214
            use_tqdm: If `True`, shows a tqdm progress bar.
                If a callable (e.g., `functools.partial(tqdm, leave=False)`),
                it is used to create the progress bar.
                If `False`, no progress bar is created.
1215
            lora_request: LoRA request to use for generation, if any.
1216
1217
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1218
        Returns:
1219
            A list of `ClassificationRequestOutput` objects containing the
1220
1221
            embedding vectors in the same order as the input prompts.
        """
1222
        if "classify" not in self.supported_tasks:
1223
            raise ValueError(
1224
                "Classification API is not supported by this model. "
1225
                "Try converting the model using `--convert classify`.")
1226

1227
1228
1229
        items = self.encode(
            prompts,
            use_tqdm=use_tqdm,
1230
            pooling_params=pooling_params,
1231
1232
1233
            lora_request=lora_request,
            pooling_task="classify",
        )
1234
1235
1236

        return [ClassificationRequestOutput.from_base(item) for item in items]

1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
    def reward(
        self,
        prompts: Union[PromptType, Sequence[PromptType]],
        /,
        *,
        truncate_prompt_tokens: Optional[int] = None,
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
        pooling_params: Optional[Union[PoolingParams,
                                       Sequence[PoolingParams]]] = None,
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
    ) -> list[PoolingRequestOutput]:
        """
        Generate rewards for each prompt.

        Args:
            prompts: The prompts to the LLM. You may pass a sequence of prompts
                for batch inference. See [PromptType][vllm.inputs.PromptType]
                for more details about the format of each prompts.
            use_tqdm: If `True`, shows a tqdm progress bar.
                If a callable (e.g., `functools.partial(tqdm, leave=False)`),
                it is used to create the progress bar.
                If `False`, no progress bar is created.
            lora_request: LoRA request to use for generation, if any.
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
        Returns:
            A list of `PoolingRequestOutput` objects containing the
            pooled hidden states in the same order as the input prompts.
        """

        return self.encode(
            prompts,
            use_tqdm=use_tqdm,
            lora_request=lora_request,
            pooling_params=pooling_params,
            truncate_prompt_tokens=truncate_prompt_tokens,
            pooling_task="encode",
        )

1276
1277
1278
    def _embedding_score(
        self,
        tokenizer: AnyTokenizer,
1279
1280
        text_1: list[Union[str, TextPrompt, TokensPrompt]],
        text_2: list[Union[str, TextPrompt, TokensPrompt]],
1281
        truncate_prompt_tokens: Optional[int] = None,
1282
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
1283
        pooling_params: Optional[PoolingParams] = None,
1284
1285
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
    ) -> list[ScoringRequestOutput]:
1286

1287
        encoded_output: list[PoolingRequestOutput] = self.encode(
1288
            text_1 + text_2,
1289
            truncate_prompt_tokens=truncate_prompt_tokens,
1290
1291
            use_tqdm=use_tqdm,
            lora_request=lora_request,
1292
            pooling_params=pooling_params,
1293
1294
            pooling_task="embed",
        )
1295

1296
        encoded_output_1: list[PoolingRequestOutput] = encoded_output[
1297
            0:len(text_1)]
1298
        encoded_output_2: list[PoolingRequestOutput] = encoded_output[
1299
            len(text_1):]
1300
1301
1302
1303

        if len(encoded_output_1) == 1:
            encoded_output_1 = encoded_output_1 * len(encoded_output_2)

1304
1305
1306
        scores = _cosine_similarity(tokenizer=tokenizer,
                                    embed_1=encoded_output_1,
                                    embed_2=encoded_output_2)
1307
1308
1309
1310
1311
1312
1313

        items = self.engine_class.validate_outputs(scores,
                                                   PoolingRequestOutput)
        return [ScoringRequestOutput.from_base(item) for item in items]

    def _cross_encoding_score(
        self,
1314
        tokenizer: AnyTokenizer,
1315
1316
        data_1: Union[list[str], list[ScoreContentPartParam]],
        data_2: Union[list[str], list[ScoreContentPartParam]],
1317
        truncate_prompt_tokens: Optional[int] = None,
1318
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
1319
        pooling_params: Optional[PoolingParams] = None,
1320
1321
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
    ) -> list[ScoringRequestOutput]:
1322
        model_config = self.llm_engine.model_config
1323
1324
1325

        if isinstance(tokenizer, MistralTokenizer):
            raise ValueError(
1326
                "Score API is not supported for Mistral tokenizer")
1327

1328
1329
        if len(data_1) == 1:
            data_1 = data_1 * len(data_2)
1330

1331
1332
1333
1334
1335
        if pooling_params is None:
            pooling_params = PoolingParams(task="score")

        model_config = self.llm_engine.model_config
        pooling_params.verify("score", model_config)
1336
        pooling_params_list = list[PoolingParams]()
1337

1338
        tokenization_kwargs: dict[str, Any] = {}
1339
1340

        _validate_truncation_size(model_config.max_model_len,
1341
                                  truncate_prompt_tokens, tokenization_kwargs)
1342
1343
1344

        parsed_prompts = []

1345
1346
        input_pairs = [(t1, t2) for t1, t2 in zip(data_1, data_2)]

1347
        model_config = self.llm_engine.model_config
1348

1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
        for q, d in input_pairs:
            _, engine_prompt = get_score_prompt(
                model_config=model_config,
                data_1=q,
                data_2=d,
                tokenizer=tokenizer,
                tokenization_kwargs=tokenization_kwargs,
            )

            if envs.VLLM_USE_V1 and (token_type_ids := engine_prompt.pop(
                    "token_type_ids", None)):
                params = pooling_params.clone()
                compressed = compress_token_type_ids(token_type_ids)
                params.extra_kwargs = {"compressed_token_type_ids": compressed}
                pooling_params_list.append(params)
            else:
                pooling_params_list.append(pooling_params)

            parsed_prompts.append(engine_prompt)
1368
1369
1370

        self._validate_and_add_requests(
            prompts=parsed_prompts,
1371
            params=pooling_params_list,
1372
            use_tqdm=use_tqdm,
1373
1374
1375
1376
1377
1378
1379
1380
1381
            lora_request=lora_request,
        )

        outputs = self._run_engine(use_tqdm=use_tqdm)
        items = self.engine_class.validate_outputs(outputs,
                                                   PoolingRequestOutput)

        return [ScoringRequestOutput.from_base(item) for item in items]

1382
1383
    def score(
        self,
1384
1385
1386
1387
        data_1: Union[SingletonPrompt, Sequence[SingletonPrompt],
                      ScoreMultiModalParam],
        data_2: Union[SingletonPrompt, Sequence[SingletonPrompt],
                      ScoreMultiModalParam],
1388
        /,
1389
        *,
1390
        truncate_prompt_tokens: Optional[int] = None,
1391
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
1392
        pooling_params: Optional[PoolingParams] = None,
1393
1394
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
    ) -> list[ScoringRequestOutput]:
1395
1396
        """Generate similarity scores for all pairs `<text,text_pair>` or
          `<multi-modal data, multi-modal data pair>`.
1397

1398
        The inputs can be `1 -> 1`, `1 -> N` or `N -> N`.
1399
1400
        In the `1 - N` case the `data_1` input will be replicated `N`
        times to pair with the `data_2` inputs.
1401
        The input pairs are used to build a list of prompts for the
1402
1403
        cross encoder model. This class automatically batches the prompts,
        considering the memory constraint. For the best performance, put all
1404
1405
1406
        of your inputs into a single list and pass it to this method.

        Supports both text and multi-modal data (images, etc.) when used with
1407
        appropriate multi-modal models. For multi-modal inputs, ensure the
1408
        prompt structure matches the model's expected input format.
1409
1410

        Args:
1411
1412
1413
            data_1: Can be a single prompt, a list of prompts or
                `ScoreMultiModalParam`, which can contain either text or
                multi-modal data. When a list, it must have the same length as
1414
                the `data_2` list.
1415
            data_2: The data to pair with the query to form the input to
1416
                the LLM. Can be text or multi-modal data. See [PromptType]
1417
                [vllm.inputs.PromptType] for more details about the format of
1418
                each prompt.
1419
1420
1421
1422
            use_tqdm: If `True`, shows a tqdm progress bar.
                If a callable (e.g., `functools.partial(tqdm, leave=False)`),
                it is used to create the progress bar.
                If `False`, no progress bar is created.
1423
            lora_request: LoRA request to use for generation, if any.
1424
1425
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1426
        Returns:
1427
            A list of `ScoringRequestOutput` objects containing the
1428
1429
            generated scores in the same order as the input prompts.
        """
1430
1431
        model_config = self.llm_engine.model_config
        runner_type = model_config.runner_type
1432
        if runner_type != "pooling":
1433
1434
1435
1436
            raise ValueError(
                "LLM.score() is only supported for pooling models. "
                "Try passing `--runner pooling` to use the model as a "
                "pooling model.")
1437

1438
1439
        supported_tasks = self.supported_tasks
        if all(t not in supported_tasks for t in ("embed", "classify")):
1440
            raise ValueError("Score API is not supported by this model. "
1441
1442
                             "Try converting the model using "
                             "`--convert embed` or `--convert classify`.")
1443

1444
        if (model_config.is_cross_encoder
1445
                and getattr(model_config.hf_config, "num_labels", 0) != 1):
1446
            raise ValueError("Score API is only enabled for num_labels == 1.")
1447
1448
1449
1450

        # the tokenizer for models such as
        # "cross-encoder/ms-marco-MiniLM-L-6-v2" doesn't support passing
        # lists of tokens to the `text` and `text_pair` kwargs
1451
        tokenizer = self.get_tokenizer()
1452

1453
        if not model_config.is_multimodal_model:
1454
1455
1456
1457
1458

            def check_data_type(data: Union[SingletonPrompt,
                                            Sequence[SingletonPrompt],
                                            ScoreMultiModalParam]):
                if isinstance(data, dict) and "content" in data:
1459
1460
                    raise ValueError("ScoreMultiModalParam is not supported "
                                     f"for {model_config.architecture}")
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
1492
1493
1494
1495
1496
1497
1498
1499
1500

            check_data_type(data_1)
            check_data_type(data_2)

            def ensure_str(prompt: SingletonPrompt):
                if isinstance(prompt, dict):
                    if "multi_modal_data" in prompt:
                        raise ValueError("Multi-modal prompt is not "
                                         "supported for scoring")
                    elif "prompt_token_ids" in prompt:
                        prompt = tokenizer.decode(
                            cast(TokensPrompt, prompt)["prompt_token_ids"])
                    elif "prompt" in prompt:
                        prompt = cast(TextPrompt, prompt)["prompt"]
                assert type(prompt) is str
                return prompt

            if isinstance(data_1, (str, dict)):
                # Convert a single prompt to a list.
                data_1 = [data_1]  # type: ignore[list-item]

            data_1 = [ensure_str(t) for t in data_1]

            if isinstance(data_2, (str, dict)):
                # Convert a single prompt to a list.
                data_2 = [data_2]  # type: ignore[list-item]

            data_2 = [ensure_str(t) for t in data_2]

        if isinstance(data_1, dict) and "content" in data_1:
            data_1 = data_1.get("content")  # type: ignore[assignment]
        elif isinstance(data_1, str):
            data_1 = [data_1]

        if isinstance(data_2, dict) and "content" in data_2:
            data_2 = data_2.get("content")  # type: ignore[assignment]
        elif isinstance(data_2, str):
            data_2 = [data_2]

        _validate_score_input_lens(data_1, data_2)  # type: ignore[arg-type]
1501

1502
        if model_config.is_cross_encoder:
1503
1504
1505
1506
1507
1508
            return self._cross_encoding_score(
                tokenizer,
                data_1,  # type: ignore[arg-type]
                data_2,  # type: ignore[arg-type]
                truncate_prompt_tokens,
                use_tqdm,
1509
                pooling_params,
1510
                lora_request)
1511
        else:
1512
1513
            return self._embedding_score(
                tokenizer,
1514
1515
                data_1,  # type: ignore[arg-type]
                data_2,  # type: ignore[arg-type]
1516
1517
                truncate_prompt_tokens,
                use_tqdm,
1518
                pooling_params,
1519
                lora_request)
1520

1521
1522
1523
1524
1525
1526
    def start_profile(self) -> None:
        self.llm_engine.start_profile()

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

1527
1528
    def reset_prefix_cache(self, device: Optional[Device] = None) -> bool:
        return self.llm_engine.reset_prefix_cache(device)
1529

1530
1531
1532
1533
1534
1535
    def sleep(self, level: int = 1):
        """
        Put the engine to sleep. The engine should not process any requests.
        The caller should guarantee that no requests are being processed
        during the sleep period, before `wake_up` is called.

1536
        Args:
1537
1538
            level: The sleep level. Level 1 sleep will offload the model
                weights and discard the kv cache. The content of kv cache
1539
                is forgotten. Level 1 sleep is good for sleeping and waking
1540
1541
1542
1543
1544
                up the engine to run the same model again. The model weights
                are backed up in CPU memory. Please make sure there's enough
                CPU memory to store the model weights. Level 2 sleep will
                discard both the model weights and the kv cache. The content
                of both the model weights and kv cache is forgotten. Level 2
1545
                sleep is good for sleeping and waking up the engine to run a
1546
                different model or update the model, where previous model
1547
                weights are not needed. It reduces CPU memory pressure.
1548
        """
1549
        self.reset_prefix_cache()
1550
1551
        self.llm_engine.sleep(level=level)

1552
    def wake_up(self, tags: Optional[list[str]] = None):
1553
        """
1554
        Wake up the engine from sleep mode. See the [sleep][] method
1555
        for more details.
1556

1557
        Args:
1558
1559
            tags: An optional list of tags to reallocate the engine memory
                for specific memory allocations. Values must be in
1560
                `("weights", "kv_cache")`. If None, all memory is reallocated.
1561
                wake_up should be called with all tags (or None) before the
1562
1563
1564
                engine is used again.
        """
        self.llm_engine.wake_up(tags)
1565

1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
    def get_metrics(self) -> list["Metric"]:
        """Return a snapshot of aggregated metrics from Prometheus.

        Returns:
            A ``MetricSnapshot`` instance capturing the current state
            of all aggregated metrics from Prometheus.

        Note:
            This method is only available with the V1 LLM engine.
        """
        from vllm.v1.engine.llm_engine import LLMEngine as V1LLMEngine
        assert isinstance(self.llm_engine, V1LLMEngine)
        return self.llm_engine.get_metrics()

1580
1581
    # LEGACY
    def _convert_v1_inputs(
1582
        self,
1583
1584
        prompts: Optional[Union[str, list[str]]],
        prompt_token_ids: Optional[Union[list[int], list[list[int]]]],
1585
1586
    ):
        # skip_tokenizer_init is now checked in engine
1587

1588
1589
1590
1591
1592
1593
1594
1595
1596
        if prompts is None and prompt_token_ids is None:
            raise ValueError(
                "Either prompts or prompt_token_ids must be provided.")
        if prompts is not None and prompt_token_ids is not None \
                and len(prompts) != len(prompt_token_ids):
            raise ValueError(
                "The lengths of prompts and prompt_token_ids must be the same."
            )

1597
1598
1599
1600
1601
1602
        if prompts is not None:
            prompts = [p["content"] for p in parse_and_batch_prompt(prompts)]
        if prompt_token_ids is not None:
            prompt_token_ids = [
                p["content"] for p in parse_and_batch_prompt(prompt_token_ids)
            ]
1603
1604
        if prompts is not None:
            num_requests = len(prompts)
1605
        elif prompt_token_ids is not None:
1606
            num_requests = len(prompt_token_ids)
1607
        parsed_prompts: list[PromptType] = []
1608
        for i in range(num_requests):
1609
            item: PromptType
1610

1611
            if prompts is not None:
1612
1613
1614
                item = TextPrompt(prompt=prompts[i])
            elif prompt_token_ids is not None:
                item = TokensPrompt(prompt_token_ids=prompt_token_ids[i])
1615
            else:
1616
                raise AssertionError
1617

1618
            parsed_prompts.append(item)
1619

1620
        return parsed_prompts
1621
1622
1623

    def _validate_and_add_requests(
        self,
1624
        prompts: Union[PromptType, Sequence[PromptType]],
1625
1626
        params: Union[SamplingParams, Sequence[SamplingParams], PoolingParams,
                      Sequence[PoolingParams]],
1627
        *,
1628
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
1629
        lora_request: Optional[Union[Sequence[LoRARequest], LoRARequest]],
1630
        tokenization_kwargs: Optional[dict[str, Any]] = None,
1631
        priority: Optional[list[int]] = None,
1632
    ) -> None:
1633
        if isinstance(prompts, (str, dict)):
1634
            # Convert a single prompt to a list.
1635
            prompts = [prompts]
1636

1637
        num_requests = len(prompts)
1638
        if isinstance(params, Sequence) and len(params) != num_requests:
1639
            raise ValueError("The lengths of prompts and params "
1640
                             "must be the same.")
1641
        if isinstance(lora_request,
1642
                      Sequence) and len(lora_request) != num_requests:
1643
1644
            raise ValueError("The lengths of prompts and lora_request "
                             "must be the same.")
1645

1646
        for sp in params if isinstance(params, Sequence) else (params, ):
1647
1648
1649
            if isinstance(sp, SamplingParams):
                # We only care about the final output
                sp.output_kind = RequestOutputKind.FINAL_ONLY
1650

Zhuohan Li's avatar
Zhuohan Li committed
1651
        # Add requests to the engine.
1652
1653
        it = prompts
        if use_tqdm:
1654
1655
            tqdm_func = use_tqdm if callable(use_tqdm) else tqdm
            it = tqdm_func(it, desc="Adding requests")
1656
1657

        for i, prompt in enumerate(it):
1658
            self._add_request(
1659
                prompt,
1660
                params[i] if isinstance(params, Sequence) else params,
1661
                tokenization_kwargs=tokenization_kwargs,
1662
1663
                lora_request=lora_request[i] if isinstance(
                    lora_request, Sequence) else lora_request,
1664
                priority=priority[i] if priority else 0,
nunjunj's avatar
nunjunj committed
1665
            )
1666

1667
    def _add_request(
nunjunj's avatar
nunjunj committed
1668
        self,
1669
        prompt: PromptType,
nunjunj's avatar
nunjunj committed
1670
        params: Union[SamplingParams, PoolingParams],
1671
        tokenization_kwargs: Optional[dict[str, Any]] = None,
1672
        lora_request: Optional[LoRARequest] = None,
1673
        priority: int = 0,
1674
1675
    ) -> None:
        request_id = str(next(self.request_counter))
1676
1677
        self.llm_engine.add_request(
            request_id,
1678
            prompt,
1679
1680
            params,
            lora_request=lora_request,
1681
            tokenization_kwargs=tokenization_kwargs,
1682
            priority=priority,
nunjunj's avatar
nunjunj committed
1683
        )
1684

1685
    def _run_engine(
1686
1687
1688
        self,
        *,
        use_tqdm: Union[bool, Callable[..., tqdm]] = True
1689
    ) -> list[Union[RequestOutput, PoolingRequestOutput]]:
1690
1691
        # Initialize tqdm.
        if use_tqdm:
Zhuohan Li's avatar
Zhuohan Li committed
1692
            num_requests = self.llm_engine.get_num_unfinished_requests()
1693
1694
            tqdm_func = use_tqdm if callable(use_tqdm) else tqdm
            pbar = tqdm_func(
1695
1696
1697
                total=num_requests,
                desc="Processed prompts",
                dynamic_ncols=True,
1698
1699
                postfix=(f"est. speed input: {0:.2f} toks/s, "
                         f"output: {0:.2f} toks/s"),
1700
            )
1701

Zhuohan Li's avatar
Zhuohan Li committed
1702
        # Run the engine.
1703
        outputs: list[Union[RequestOutput, PoolingRequestOutput]] = []
1704
1705
        total_in_toks = 0
        total_out_toks = 0
Zhuohan Li's avatar
Zhuohan Li committed
1706
1707
        while self.llm_engine.has_unfinished_requests():
            step_outputs = self.llm_engine.step()
1708
            for output in step_outputs:
1709
                if output.finished:
1710
1711
                    outputs.append(output)
                    if use_tqdm:
1712
1713
                        if isinstance(output, RequestOutput):
                            # Calculate tokens only for RequestOutput
1714
                            n = len(output.outputs)
1715
                            assert output.prompt_token_ids is not None
1716
                            total_in_toks += len(output.prompt_token_ids) * n
1717
1718
                            in_spd = total_in_toks / pbar.format_dict["elapsed"]
                            total_out_toks += sum(
1719
                                len(stp.token_ids) for stp in output.outputs)
nunjunj's avatar
nunjunj committed
1720
1721
                            out_spd = (total_out_toks /
                                       pbar.format_dict["elapsed"])
1722
1723
1724
                            pbar.postfix = (
                                f"est. speed input: {in_spd:.2f} toks/s, "
                                f"output: {out_spd:.2f} toks/s")
1725
                            pbar.update(n)
1726
1727
                        else:
                            pbar.update(1)
1728
1729
                        if pbar.n == num_requests:
                            pbar.refresh()
1730

1731
1732
        if use_tqdm:
            pbar.close()
1733
1734
1735
        # Sort the outputs by request ID.
        # This is necessary because some requests may be finished earlier than
        # its previous requests.
1736
        return sorted(outputs, key=lambda x: int(x.request_id))