llm.py 71.7 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import itertools
5
6
from collections.abc import Callable, Sequence
from typing import TYPE_CHECKING, Any, cast
7

8
import cloudpickle
9
import torch.nn as nn
10
from pydantic import ValidationError
11
from tqdm.auto import tqdm
12
from typing_extensions import TypeVar, deprecated
13

14
15
16
17
18
19
20
21
from vllm.beam_search import (
    BeamSearchInstance,
    BeamSearchOutput,
    BeamSearchSequence,
    create_sort_beams_key_function,
)
from vllm.config import (
    CompilationConfig,
22
    PoolerConfig,
23
24
25
    StructuredOutputsConfig,
    is_init_field,
)
26
from vllm.config.model import (
27
28
    ConvertOption,
    HfOverrides,
29
    ModelDType,
30
    RunnerOption,
31
    TokenizerMode,
32
)
33
from vllm.engine.arg_utils import EngineArgs
34
from vllm.engine.protocol import Device
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
from vllm.entrypoints.chat_utils import (
    ChatCompletionMessageParam,
    ChatTemplateContentFormatOption,
    apply_hf_chat_template,
    apply_mistral_chat_template,
    parse_chat_messages,
    resolve_chat_template_content_format,
)
from vllm.entrypoints.score_utils import (
    ScoreContentPartParam,
    ScoreMultiModalParam,
    _cosine_similarity,
    _validate_score_input_lens,
    compress_token_type_ids,
    get_score_prompt,
)
from vllm.entrypoints.utils import _validate_truncation_size, log_non_default_args
from vllm.inputs import (
    DataPrompt,
    PromptType,
    SingletonPrompt,
    TextPrompt,
    TokensPrompt,
)
59
from vllm.inputs.parse import get_prompt_components
60
from vllm.logger import init_logger
61
from vllm.lora.request import LoRARequest
62
from vllm.model_executor.layers.quantization import QuantizationMethods
63
64
65
66
67
68
69
from vllm.outputs import (
    ClassificationRequestOutput,
    EmbeddingRequestOutput,
    PoolingRequestOutput,
    RequestOutput,
    ScoringRequestOutput,
)
70
from vllm.pooling_params import PoolingParams
71
from vllm.sampling_params import BeamSearchParams, RequestOutputKind, SamplingParams
72
from vllm.tasks import PoolingTask
73
74
75
76
77
from vllm.transformers_utils.tokenizer import (
    AnyTokenizer,
    MistralTokenizer,
    get_cached_tokenizer,
)
yhu422's avatar
yhu422 committed
78
from vllm.usage.usage_lib import UsageContext
79
from vllm.utils.collection_utils import as_iter, is_list_of
80
from vllm.utils.counter import Counter
81
from vllm.v1.engine import EngineCoreRequest
82
from vllm.v1.engine.llm_engine import LLMEngine
83
from vllm.v1.sample.logits_processor import LogitsProcessor
84

85
86
87
if TYPE_CHECKING:
    from vllm.v1.metrics.reader import Metric

88
89
logger = init_logger(__name__)

90
91
_R = TypeVar("_R", default=Any)

92
93

class LLM:
Woosuk Kwon's avatar
Woosuk Kwon committed
94
95
96
97
98
99
100
101
102
103
    """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.
104
        tokenizer: The name or path of a HuggingFace Transformers tokenizer.
105
106
        tokenizer_mode: The tokenizer mode. "auto" will use the fast tokenizer
            if available, and "slow" will always use the slow tokenizer.
107
108
109
        skip_tokenizer_init: If true, skip initialization of tokenizer and
            detokenizer. Expect valid prompt_token_ids and None for prompt
            from the input.
110
111
        trust_remote_code: Trust remote code (e.g., from HuggingFace) when
            downloading the model and tokenizer.
112
113
114
115
        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.
116
        allowed_media_domains: If set, only media URLs that belong to this
117
            domain can be used for multi-modal inputs.
Woosuk Kwon's avatar
Woosuk Kwon committed
118
119
120
        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
121
            we support `float32`, `float16`, and `bfloat16`. If `auto`, we use
122
123
            the `dtype` attribute of the Transformers model's config. However,
            if the `dtype` in the config is `float32`, we will use `float16` instead.
124
        quantization: The method used to quantize the model weights. Currently,
125
            we support "awq", "gptq", and "fp8" (experimental).
126
127
128
129
            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
130
131
        revision: The specific model version to use. It can be a branch name,
            a tag name, or a commit id.
132
133
        tokenizer_revision: The specific tokenizer version to use. It can be a
            branch name, a tag name, or a commit id.
134
135
136
137
138
139
        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.
140
141
142
143
144
        kv_cache_memory_bytes: Size of KV Cache per GPU in bytes. By default,
            this is set to None and vllm can automatically infer the kv cache
            size based on gpu_memory_utilization. However, users may want to
            manually specify the kv cache memory size. kv_cache_memory_bytes
            allows more fine-grain control of how much memory gets used when
145
            compared with using gpu_memory_utilization. Note that
146
147
            kv_cache_memory_bytes (when not-None) ignores
            gpu_memory_utilization
148
        swap_space: The size (GiB) of CPU memory per GPU to use as swap space.
149
150
151
152
153
            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.
154
155
156
157
        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.
158
159
160
        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.
161
162
        disable_custom_all_reduce: See
            [ParallelConfig][vllm.config.ParallelConfig].
163
        hf_token: The token to use as HTTP bearer authorization for remote files
164
            . If `True`, will use the token generated when running
165
            `huggingface-cli login` (stored in `~/.huggingface`).
166
167
168
        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.
169
170
171
172
173
        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}`.
174
175
176
177
178
        pooler_config: Initialize non-default pooling config for the pooling
            model. e.g. `PoolerConfig(pooling_type="mean", normalize=False)`.
        override_pooler_config: [DEPRECATED] Use `pooler_config` instead. This
            argument is deprecated and will be removed in v0.12.0 or v1.0.0,
            whichever is sooner.
179
        compilation_config: Either an integer or a dictionary. If it is an
180
            integer, it is used as the mode of compilation optimization. If it
181
            is a dictionary, it can specify the full compilation configuration.
182
        **kwargs: Arguments for [`EngineArgs`][vllm.EngineArgs].
nunjunj's avatar
nunjunj committed
183

184
185
    Note:
        This class is intended to be used for offline inference. For online
186
        serving, use the [AsyncLLMEngine][vllm.AsyncLLMEngine] class instead.
187
    """
188
189
190
191

    def __init__(
        self,
        model: str,
192
        *,
193
194
        runner: RunnerOption = "auto",
        convert: ConvertOption = "auto",
195
        tokenizer: str | None = None,
196
        tokenizer_mode: TokenizerMode = "auto",
197
        skip_tokenizer_init: bool = False,
198
        trust_remote_code: bool = False,
199
        allowed_local_media_path: str = "",
200
        allowed_media_domains: list[str] | None = None,
201
        tensor_parallel_size: int = 1,
202
        dtype: ModelDType = "auto",
203
204
205
206
        quantization: QuantizationMethods | None = None,
        revision: str | None = None,
        tokenizer_revision: str | None = None,
        seed: int | None = None,
207
        gpu_memory_utilization: float = 0.9,
208
        swap_space: float = 4,
209
        cpu_offload_gb: float = 0,
210
        enforce_eager: bool = False,
211
        disable_custom_all_reduce: bool = False,
212
213
214
215
216
217
218
219
220
221
222
        hf_token: bool | str | None = None,
        hf_overrides: HfOverrides | None = None,
        mm_processor_kwargs: dict[str, Any] | None = None,
        pooler_config: PoolerConfig | None = None,
        override_pooler_config: PoolerConfig | None = None,
        structured_outputs_config: dict[str, Any]
        | StructuredOutputsConfig
        | None = None,
        kv_cache_memory_bytes: int | None = None,
        compilation_config: int | dict[str, Any] | CompilationConfig | None = None,
        logits_processors: list[str | type[LogitsProcessor]] | None = None,
223
        **kwargs: Any,
224
    ) -> None:
225
        """LLM constructor."""
226

227
228
        if "disable_log_stats" not in kwargs:
            kwargs["disable_log_stats"] = True
229

230
231
232
233
234
235
236
        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)

237
        if "kv_transfer_config" in kwargs and isinstance(
238
239
            kwargs["kv_transfer_config"], dict
        ):
240
            from vllm.config.kv_transfer import KVTransferConfig
241

242
243
            raw_config_dict = kwargs["kv_transfer_config"]
            try:
244
                kwargs["kv_transfer_config"] = KVTransferConfig(**raw_config_dict)
245
246
247
248
            except ValidationError as e:
                logger.error(
                    "Failed to convert 'kv_transfer_config' dict to "
                    "KVTransferConfig object. Dict: %s. Error: %s",
249
250
251
                    raw_config_dict,
                    e,
                )
252
253
                # Consider re-raising a more specific vLLM error or ValueError
                # to provide better context to the user.
254
                raise ValueError(f"Invalid 'kv_transfer_config' provided: {e}") from e
255

256
257
258
        if hf_overrides is None:
            hf_overrides = {}

259
        if compilation_config is not None:
260
            if isinstance(compilation_config, int):
261
                compilation_config_instance = CompilationConfig(mode=compilation_config)
262
263
            elif isinstance(compilation_config, dict):
                compilation_config_instance = CompilationConfig(
264
265
266
267
                    **{
                        k: v
                        for k, v in compilation_config.items()
                        if is_init_field(CompilationConfig, k)
268
269
                    }
                )
270
271
            else:
                compilation_config_instance = compilation_config
272
        else:
273
            compilation_config_instance = CompilationConfig()
274

275
276
277
278
279
280
281
        if structured_outputs_config is not None:
            if isinstance(structured_outputs_config, dict):
                structured_outputs_instance = StructuredOutputsConfig(
                    **{
                        k: v
                        for k, v in structured_outputs_config.items()
                        if is_init_field(StructuredOutputsConfig, k)
282
283
                    }
                )
284
285
286
287
288
            else:
                structured_outputs_instance = structured_outputs_config
        else:
            structured_outputs_instance = StructuredOutputsConfig()

289
        # warn about single-process data parallel usage.
290
291
292
        _dp_size = int(kwargs.get("data_parallel_size", 1))
        _distributed_executor_backend = kwargs.get("distributed_executor_backend")
        if _dp_size > 1 and not _distributed_executor_backend == "external_launcher":
293
            raise ValueError(
294
                f"LLM(data_parallel_size={_dp_size}) is not supported for single-"
295
296
297
298
299
                "process usage and may hang. Please use "
                "the explicit multi-process data-parallel example at "
                "'examples/offline_inference/data_parallel.py'."
            )

Zhuohan Li's avatar
Zhuohan Li committed
300
        engine_args = EngineArgs(
301
            model=model,
302
303
            runner=runner,
            convert=convert,
304
            tokenizer=tokenizer,
305
            tokenizer_mode=tokenizer_mode,
306
            skip_tokenizer_init=skip_tokenizer_init,
307
            trust_remote_code=trust_remote_code,
308
            allowed_local_media_path=allowed_local_media_path,
309
            allowed_media_domains=allowed_media_domains,
310
311
            tensor_parallel_size=tensor_parallel_size,
            dtype=dtype,
312
            quantization=quantization,
313
            revision=revision,
314
            tokenizer_revision=tokenizer_revision,
315
316
            seed=seed,
            gpu_memory_utilization=gpu_memory_utilization,
317
            kv_cache_memory_bytes=kv_cache_memory_bytes,
318
            swap_space=swap_space,
319
            cpu_offload_gb=cpu_offload_gb,
320
            enforce_eager=enforce_eager,
321
            disable_custom_all_reduce=disable_custom_all_reduce,
322
            hf_token=hf_token,
323
            hf_overrides=hf_overrides,
324
            mm_processor_kwargs=mm_processor_kwargs,
325
            pooler_config=pooler_config,
326
            override_pooler_config=override_pooler_config,
327
            structured_outputs_config=structured_outputs_instance,
328
            compilation_config=compilation_config_instance,
329
            logits_processors=logits_processors,
330
331
            **kwargs,
        )
332

333
334
        log_non_default_args(engine_args)

335
336
        # Create the Engine (autoselects V0 vs V1)
        self.llm_engine = LLMEngine.from_engine_args(
337
338
            engine_args=engine_args, usage_context=UsageContext.LLM_CLASS
        )
339
        self.engine_class = type(self.llm_engine)
340

341
        self.request_counter = Counter()
342
        self.default_sampling_params: dict[str, Any] | None = None
343

344
345
        supported_tasks = self.llm_engine.get_supported_tasks()
        logger.info("Supported tasks: %s", supported_tasks)
346
347
        self.supported_tasks = supported_tasks

348
349
350
        self.model_config = self.llm_engine.model_config
        self.processor = self.llm_engine.processor
        self.io_processor = self.llm_engine.io_processor
351

352
353
    def get_tokenizer(self) -> AnyTokenizer:
        return self.llm_engine.get_tokenizer()
354

355
    @deprecated("`set_tokenizer` is deprecated and will be removed in v0.13.")
356
    def set_tokenizer(self, tokenizer: AnyTokenizer) -> None:
357
358
359
360
        # 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"):
361
            self.llm_engine.tokenizer = tokenizer
362
        else:
363
            self.llm_engine.tokenizer = get_cached_tokenizer(tokenizer)
364

365
366
367
368
    def reset_mm_cache(self) -> None:
        self.processor.clear_mm_cache()
        self.llm_engine.reset_mm_cache()

369
    def get_default_sampling_params(self) -> SamplingParams:
370
        if self.default_sampling_params is None:
371
            self.default_sampling_params = self.model_config.get_diff_sampling_param()
372
373
        if self.default_sampling_params:
            return SamplingParams.from_optional(**self.default_sampling_params)
374
375
        return SamplingParams()

376
377
    def generate(
        self,
378
379
        prompts: PromptType | Sequence[PromptType],
        sampling_params: SamplingParams | Sequence[SamplingParams] | None = None,
380
        *,
381
382
383
        use_tqdm: bool | Callable[..., tqdm] = True,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
        priority: list[int] | None = None,
384
    ) -> list[RequestOutput]:
Woosuk Kwon's avatar
Woosuk Kwon committed
385
386
        """Generates the completions for the input prompts.

387
        This class automatically batches the given prompts, considering
Woosuk Kwon's avatar
Woosuk Kwon committed
388
389
390
391
        the memory constraint. For the best performance, put all of your prompts
        into a single list and pass it to this method.

        Args:
392
            prompts: The prompts to the LLM. You may pass a sequence of prompts
393
                for batch inference. See [PromptType][vllm.inputs.PromptType]
394
                for more details about the format of each prompt.
Woosuk Kwon's avatar
Woosuk Kwon committed
395
            sampling_params: The sampling parameters for text generation. If
nunjunj's avatar
nunjunj committed
396
397
398
                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
399
                prompts and it is paired one by one with the prompt.
400
401
402
403
            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.
404
            lora_request: LoRA request to use for generation, if any.
405
406
            priority: The priority of the requests, if any.
                Only applicable when priority scheduling policy is enabled.
Woosuk Kwon's avatar
Woosuk Kwon committed
407
408

        Returns:
409
            A list of `RequestOutput` objects containing the
410
            generated completions in the same order as the input prompts.
411

412
413
414
415
        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.
416
        """
417
        model_config = self.model_config
418
419
        runner_type = model_config.runner_type
        if runner_type != "generate":
420
421
422
            raise ValueError(
                "LLM.generate() is only supported for generative models. "
                "Try passing `--runner generate` to use the model as a "
423
424
                "generative model."
            )
425

426
427
        if sampling_params is None:
            # Use default sampling params.
428
            sampling_params = self.get_default_sampling_params()
429

430
        # Add any modality specific loras to the corresponding prompts
431
        lora_request = self._get_modality_specific_lora_reqs(prompts, lora_request)
432

433
        self._validate_and_add_requests(
434
            prompts=prompts,
435
            params=sampling_params,
436
            use_tqdm=use_tqdm,
437
            lora_request=lora_request,
438
439
            priority=priority,
        )
440

441
        outputs = self._run_engine(use_tqdm=use_tqdm)
Joe Runde's avatar
Joe Runde committed
442
        return self.engine_class.validate_outputs(outputs, RequestOutput)
443

444
    def _get_modality_specific_lora_reqs(
445
        self,
446
447
        prompts: PromptType | Sequence[PromptType],
        lora_request: list[LoRARequest] | LoRARequest | None,
448
    ):
449
450
451
452
453
454
        # 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.
455
456
        if (
            lora_config is None
457
            or not self.model_config.is_multimodal_model
458
459
            or (lora_config and lora_config.default_mm_loras is None)
        ):
460
461
            return lora_request

462
463
        if not isinstance(prompts, Sequence):
            prompts = [prompts]
464

465
466
467
468
469
        optional_loras = (
            [lora_request] * len(prompts)
            if not isinstance(lora_request, Sequence)
            else lora_request
        )
470
471
472

        return [
            self._resolve_single_prompt_mm_lora(
473
                prompt,
474
475
                opt_lora_req,
                lora_config.default_mm_loras,
476
477
            )
            for prompt, opt_lora_req in zip(prompts, optional_loras)
478
479
        ]

480
481
482
    def _resolve_single_prompt_mm_lora(
        self,
        prompt: PromptType,
483
484
        lora_request: LoRARequest | None,
        default_mm_loras: dict[str, str] | None,
485
486
487
488
    ):
        if (
            not default_mm_loras
            or not isinstance(prompt, dict)
489
            or not (mm_data := prompt.get("multi_modal_data") or {})
490
        ):
491
492
            return lora_request

493
494
495
        intersection = set(
            mm_data.keys()  # type: ignore
        ).intersection(default_mm_loras.keys())
496
497
498
499
500
501
502
503
504
        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"
505
506
507
                " will be skipped",
                intersection,
            )
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
            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 "
523
524
                    "lora_request as we only apply one LoRARequest per prompt"
                )
525
526
527
528
529
530
531
532
            return lora_request

        return LoRARequest(
            modality_name,
            modality_lora_id,
            modality_lora_path,
        )

533
534
    def collective_rpc(
        self,
535
536
        method: str | Callable[..., _R],
        timeout: float | None = None,
537
        args: tuple = (),
538
        kwargs: dict[str, Any] | None = None,
539
    ) -> list[_R]:
540
541
542
543
544
545
546
547
548
549
550
        """
        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
551
                [`TimeoutError`][] on timeout. `None` means wait indefinitely.
552
553
554
555
556
            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.
557

558
559
560
        Note:
            It is recommended to use this API to only pass control messages,
            and set up data-plane communication to pass data.
561
        """
562
563

        return self.llm_engine.collective_rpc(method, timeout, args, kwargs)
564
565

    def apply_model(self, func: Callable[[nn.Module], _R]) -> list[_R]:
566
        """
567
568
        Run a function directly on the model inside each worker,
        returning the result for each of them.
569
570
571
572
573
574

        !!! warning
            To reduce the overhead of data transfer, avoid returning large
            arrays or tensors from this method. If you must return them,
            make sure you move them to CPU first to avoid taking up additional
            VRAM!
575
        """
576
        return self.llm_engine.apply_model(func)
577

578
579
    def _get_beam_search_lora_requests(
        self,
580
581
582
        lora_request: list[LoRARequest] | LoRARequest | None,
        prompts: list[TokensPrompt | TextPrompt],
    ) -> list[LoRARequest | None]:
583
        """Get the optional lora request corresponding to each prompt."""
584
        if isinstance(lora_request, Sequence) and len(lora_request) != len(prompts):
585
            raise ValueError(
586
587
                "Lora request list should be the same length as the prompts"
            )
588
589
590
591
592
593

        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)}")

594
595
    def beam_search(
        self,
596
        prompts: list[TokensPrompt | TextPrompt],
597
        params: BeamSearchParams,
598
        lora_request: list[LoRARequest] | LoRARequest | None = None,
599
        use_tqdm: bool = False,
600
        concurrency_limit: int | None = None,
601
    ) -> list[BeamSearchOutput]:
602
603
604
605
606
607
        """
        Generate sequences using beam search.

        Args:
            prompts: A list of prompts. Each prompt can be a string or a list
                of token IDs.
608
            params: The beam search parameters.
609
            lora_request: LoRA request to use for generation, if any.
610
            use_tqdm: Whether to use tqdm to display the progress bar.
611
612
            concurrency_limit: The maximum number of concurrent requests.
                If None, the number of concurrent requests is unlimited.
613
        """
614
615
        # TODO: how does beam search work together with length penalty,
        # frequency, penalty, and stopping criteria, etc.?
616
617
618
619
        beam_width = params.beam_width
        max_tokens = params.max_tokens
        temperature = params.temperature
        ignore_eos = params.ignore_eos
620
621
        length_penalty = params.length_penalty

622
        lora_requests = self._get_beam_search_lora_requests(lora_request, prompts)
623

624
625
626
627
628
        tokenizer = self.get_tokenizer()
        sort_beams_key = create_sort_beams_key_function(
            tokenizer.eos_token_id,
            length_penalty,
        )
629

630
631
632
        if use_tqdm and concurrency_limit is not None:
            logger.warning(
                "Progress bar is not supported when using concurrency_limit. "
633
634
                "Disabling progress bar."
            )
635
636
637
638
639
            use_tqdm = False

        if concurrency_limit is None:
            concurrency_limit = len(prompts)

640
641
        def create_tokens_prompt_from_beam(beam: BeamSearchSequence) -> TokensPrompt:
            token_prompt_kwargs: TokensPrompt = {"prompt_token_ids": beam.tokens}
642
643
644
645
            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:
646
                token_prompt_kwargs["mm_processor_kwargs"] = beam.mm_processor_kwargs
647
            return TokensPrompt(**token_prompt_kwargs)
648

649
650
651
        # 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
652
653
654
        beam_search_params = SamplingParams(
            logprobs=2 * beam_width, max_tokens=1, temperature=temperature
        )
655
        instances: list[BeamSearchInstance] = []
656

657
        for lora_req, prompt in zip(lora_requests, prompts):
658
659
660
661
662
            # 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:
663
                mm_kwargs["mm_processor_kwargs"] = prompt["mm_processor_kwargs"]
664

665
666
            if "prompt_token_ids" in prompt:
                prompt = cast(TokensPrompt, prompt)  # Needed for mypy
667
668
669
                prompt_tokens = prompt["prompt_token_ids"]
            else:
                prompt_tokens = tokenizer.encode(prompt["prompt"])
670

671
            instances.append(
672
673
674
675
676
                BeamSearchInstance(
                    prompt_tokens,
                    lora_request=lora_req,
                    logprobs=None,
                    **mm_kwargs,
677
678
                ),
            )
679

680
        for prompt_start in range(0, len(prompts), concurrency_limit):
681
            instances_batch = instances[prompt_start : prompt_start + concurrency_limit]
682
683
684

            token_iter = range(max_tokens)
            if use_tqdm:
685
686
687
                token_iter = tqdm(
                    token_iter, desc="Beam search", unit="token", unit_scale=False
                )
688
689
690
                logger.warning(
                    "The progress bar shows the upper bound on token steps and "
                    "may finish early due to stopping conditions. It does not "
691
692
                    "reflect instance-level progress."
                )
693
694
            for _ in token_iter:
                all_beams: list[BeamSearchSequence] = list(
695
696
                    sum((instance.beams for instance in instances_batch), [])
                )
697
698
                pos = [0] + list(
                    itertools.accumulate(
699
700
701
                        len(instance.beams) for instance in instances_batch
                    )
                )
702
                instance_start_and_end: list[tuple[int, int]] = list(
703
704
                    zip(pos[:-1], pos[1:])
                )
705
706
707
708
709
710

                if len(all_beams) == 0:
                    break

                # create corresponding batch entries for prompt & optional lora
                prompts_batch, lora_req_batch = zip(
711
712
713
714
715
                    *[
                        (create_tokens_prompt_from_beam(beam), beam.lora_request)
                        for beam in all_beams
                    ]
                )
716
717
718

                # only runs for one step
                # we don't need to use tqdm here
719
720
721
722
723
724
                output = self.generate(
                    prompts_batch,
                    sampling_params=beam_search_params,
                    use_tqdm=False,
                    lora_request=lora_req_batch,
                )
725

726
727
728
                for (start, end), instance in zip(
                    instance_start_and_end, instances_batch
                ):
729
730
731
732
733
734
735
736
737
738
739
740
741
742
                    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],
743
                                    logprobs=current_beam.logprobs + [logprobs],
744
                                    lora_request=current_beam.lora_request,
745
746
747
748
749
750
751
752
753
754
                                    cum_logprob=current_beam.cum_logprob
                                    + logprob_obj.logprob,
                                    multi_modal_data=current_beam.multi_modal_data,
                                    mm_processor_kwargs=current_beam.mm_processor_kwargs,
                                )

                                if (
                                    token_id == tokenizer.eos_token_id
                                    and not ignore_eos
                                ):
755
756
757
                                    instance.completed.append(new_beam)
                                else:
                                    instance_new_beams.append(new_beam)
758
759
760
                    sorted_beams = sorted(
                        instance_new_beams, key=sort_beams_key, reverse=True
                    )
761
                    instance.beams = sorted_beams[:beam_width]
762
763
764
765

        outputs = []
        for instance in instances:
            instance.completed.extend(instance.beams)
766
767
768
            sorted_completed = sorted(
                instance.completed, key=sort_beams_key, reverse=True
            )
769
770
771
772
773
774
775
776
            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

777
    def preprocess_chat(
nunjunj's avatar
nunjunj committed
778
        self,
779
780
781
        messages: list[ChatCompletionMessageParam]
        | list[list[ChatCompletionMessageParam]],
        chat_template: str | None = None,
782
        chat_template_content_format: ChatTemplateContentFormatOption = "auto",
783
        add_generation_prompt: bool = True,
784
        continue_final_message: bool = False,
785
786
787
        tools: list[dict[str, Any]] | None = None,
        chat_template_kwargs: dict[str, Any] | None = None,
        mm_processor_kwargs: dict[str, Any] | None = None,
788
    ) -> list[TokensPrompt]:
nunjunj's avatar
nunjunj committed
789
        """
790
791
        Generate prompt for a chat conversation. The pre-processed
        prompt can then be used as input for the other LLM methods.
nunjunj's avatar
nunjunj committed
792

793
        Refer to `chat` for a complete description of the arguments.
nunjunj's avatar
nunjunj committed
794
        Returns:
795
796
797
            A list of `TokensPrompts` objects containing the tokenized
            prompt after chat template interpolation, and the
            pre-processed multi-modal inputs.
nunjunj's avatar
nunjunj committed
798
        """
799
        list_of_messages: list[list[ChatCompletionMessageParam]]
nunjunj's avatar
nunjunj committed
800

801
802
        # Handle multi and single conversations
        if is_list_of(messages, list):
803
            # messages is list[list[...]]
804
            list_of_messages = cast(list[list[ChatCompletionMessageParam]], messages)
805
        else:
806
            # messages is list[...]
807
            list_of_messages = [cast(list[ChatCompletionMessageParam], messages)]
808

809
        tokenizer = self.get_tokenizer()
810
        model_config = self.model_config
811
812
        resolved_content_format = resolve_chat_template_content_format(
            chat_template,
813
            tools,
814
815
            chat_template_content_format,
            tokenizer,
816
            model_config=model_config,
817
818
        )

819
820
821
822
823
824
825
826
        _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 {})

827
        prompts: list[TokensPrompt] = []
828
829

        for msgs in list_of_messages:
830
831
832
            # 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.
833
            conversation, mm_data, mm_uuids = parse_chat_messages(
834
835
836
837
838
                msgs,
                model_config,
                tokenizer,
                content_format=resolved_content_format,
            )
839
840

            if isinstance(tokenizer, MistralTokenizer):
841
                prompt_token_ids = apply_mistral_chat_template(
842
843
                    tokenizer,
                    messages=msgs,
844
                    **_chat_template_kwargs,
845
846
                )
            else:
847
                prompt_str = apply_hf_chat_template(
848
                    tokenizer=tokenizer,
849
                    conversation=conversation,
850
                    model_config=model_config,
851
                    **_chat_template_kwargs,
852
                )
853
854
                # Special tokens are already included in chat templates so
                # should not be added by the tokenizer in this case.
855
856
857
                prompt_token_ids = tokenizer.encode(
                    prompt_str, add_special_tokens=False
                )
858

859
            prompt = TokensPrompt(prompt_token_ids=prompt_token_ids)
860
861
862
863

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

864
865
866
            if mm_uuids is not None:
                prompt["multi_modal_uuids"] = mm_uuids

867
868
869
            if mm_processor_kwargs is not None:
                prompt["mm_processor_kwargs"] = mm_processor_kwargs

870
            prompts.append(prompt)
871

872
873
874
875
        return prompts

    def chat(
        self,
876
877
878
879
880
881
        messages: list[ChatCompletionMessageParam]
        | list[list[ChatCompletionMessageParam]],
        sampling_params: SamplingParams | list[SamplingParams] | None = None,
        use_tqdm: bool | Callable[..., tqdm] = True,
        lora_request: LoRARequest | None = None,
        chat_template: str | None = None,
882
883
884
        chat_template_content_format: ChatTemplateContentFormatOption = "auto",
        add_generation_prompt: bool = True,
        continue_final_message: bool = False,
885
886
887
        tools: list[dict[str, Any]] | None = None,
        chat_template_kwargs: dict[str, Any] | None = None,
        mm_processor_kwargs: dict[str, Any] | None = None,
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
    ) -> list[RequestOutput]:
        """
        Generate responses for a chat conversation.

        The chat conversation is converted into a text prompt using the
        tokenizer and calls the [generate][vllm.LLM.generate] method to generate
        the responses.

        Multi-modal inputs can be passed in the same way you would pass them
        to the OpenAI API.

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

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

            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.
            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.
            chat_template: The template to use for structuring the chat.
                If not provided, the model's default chat template will be used.
            chat_template_content_format: The format to render message content.

                - "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?"}]`

            add_generation_prompt: If True, adds a generation template
                to each message.
            continue_final_message: If True, continues the final message in
                the conversation instead of starting a new one. Cannot be
                `True` if `add_generation_prompt` is also `True`.
            chat_template_kwargs: Additional kwargs to pass to the chat
                template.
            mm_processor_kwargs: Multimodal processor kwarg overrides for this
                chat request. Only used for offline requests.

        Returns:
            A list of `RequestOutput` objects containing the generated
            responses in the same order as the input messages.
        """

        prompts = self.preprocess_chat(
            messages=messages,
            chat_template=chat_template,
            chat_template_content_format=chat_template_content_format,
            add_generation_prompt=add_generation_prompt,
            continue_final_message=continue_final_message,
            tools=tools,
            chat_template_kwargs=chat_template_kwargs,
            mm_processor_kwargs=mm_processor_kwargs,
        )

nunjunj's avatar
nunjunj committed
951
        return self.generate(
952
            prompts,
953
            sampling_params=sampling_params,
nunjunj's avatar
nunjunj committed
954
955
956
957
            use_tqdm=use_tqdm,
            lora_request=lora_request,
        )

958
959
    def encode(
        self,
960
961
        prompts: PromptType | Sequence[PromptType] | DataPrompt,
        pooling_params: PoolingParams | Sequence[PoolingParams] | None = None,
962
        *,
963
964
965
        truncate_prompt_tokens: int | None = None,
        use_tqdm: bool | Callable[..., tqdm] = True,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
966
        pooling_task: PoolingTask | None = None,
967
        tokenization_kwargs: dict[str, Any] | None = None,
968
    ) -> list[PoolingRequestOutput]:
969
970
        """Apply pooling to the hidden states corresponding to the input
        prompts.
971

972
        This class automatically batches the given prompts, considering
973
974
975
976
        the memory constraint. For the best performance, put all of your prompts
        into a single list and pass it to this method.

        Args:
977
            prompts: The prompts to the LLM. You may pass a sequence of prompts
978
                for batch inference. See [PromptType][vllm.inputs.PromptType]
979
                for more details about the format of each prompt.
980
981
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
982
983
984
985
            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.
986
            lora_request: LoRA request to use for generation, if any.
987
            pooling_task: Override the pooling task to use.
988
989
            tokenization_kwargs: overrides tokenization_kwargs set in
                pooling_params
990
991

        Returns:
992
            A list of `PoolingRequestOutput` objects containing the
993
            pooled hidden states in the same order as the input prompts.
994

995
996
997
998
        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.
999
        """
1000

1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
        error_str = (
            "pooling_task required for `LLM.encode`\n"
            "Please use one of the more specific methods or set the "
            "pooling_task 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 similarity scores, use `LLM.score(...)`.\n"
            "  - For rewards, use `LLM.reward(...)` "
            'or `pooling_task="token_classify"`\n'
            "  - For token classification, "
            'use `pooling_task="token_classify"`\n'
            '  - For multi-vector retrieval, use `pooling_task="token_embed"`'
        )
1016

1017
        if pooling_task is None:
1018
            raise ValueError(error_str)
1019

1020
        model_config = self.model_config
1021
        runner_type = model_config.runner_type
1022
        if runner_type != "pooling":
1023
1024
1025
            raise ValueError(
                "LLM.encode() is only supported for pooling models. "
                "Try passing `--runner pooling` to use the model as a "
1026
1027
                "pooling model."
            )
1028

1029
1030
1031
1032
1033
1034
1035
1036
        io_processor_prompt = False
        if isinstance(prompts, dict) and "data" in prompts:
            io_processor_prompt = True
            if self.io_processor is None:
                raise ValueError(
                    "No IOProcessor plugin installed. Please refer "
                    "to the documentation and to the "
                    "'prithvi_geospatial_mae_io_processor' "
1037
1038
                    "offline inference example for more details."
                )
1039
1040
1041
1042
1043
1044
1045

            # Validate the request data is valid for the loaded plugin
            validated_prompt = self.io_processor.parse_request(prompts)

            # obtain the actual model prompts from the pre-processor
            prompts = self.io_processor.pre_process(prompt=validated_prompt)

1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
        if io_processor_prompt:
            assert self.io_processor is not None
            if is_list_of(pooling_params, PoolingParams):
                validated_pooling_params: list[PoolingParams] = []
                for param in as_iter(pooling_params):
                    validated_pooling_params.append(
                        self.io_processor.validate_or_generate_params(param)
                    )
                pooling_params = validated_pooling_params
            else:
                assert not isinstance(pooling_params, Sequence)
                pooling_params = self.io_processor.validate_or_generate_params(
                    pooling_params
                )
        else:
            if pooling_params is None:
                # Use default pooling params.
                pooling_params = PoolingParams()

        if pooling_task not in self.supported_tasks:
            raise ValueError(f"pooling_task must be one of {self.supported_tasks}.")

        for param in as_iter(pooling_params):
            param.verify(pooling_task, model_config)
            # for backwards compatibility
            if truncate_prompt_tokens is not None:
                param.truncate_prompt_tokens = truncate_prompt_tokens

1074
        self._validate_and_add_requests(
1075
            prompts=prompts,
1076
            params=pooling_params,
1077
            use_tqdm=use_tqdm,
1078
            lora_request=lora_request,
1079
1080
        )

1081
        outputs = self._run_engine(use_tqdm=use_tqdm)
1082
1083

        model_outputs = self.engine_class.validate_outputs(
1084
1085
            outputs, PoolingRequestOutput
        )
1086
1087
1088
1089
1090

        if io_processor_prompt:
            # get the post-processed model outputs
            assert self.io_processor is not None
            processed_outputs = self.io_processor.post_process(
1091
1092
                model_output=model_outputs
            )
1093
1094

            return [
1095
1096
1097
                PoolingRequestOutput[Any](
                    request_id="",
                    outputs=processed_outputs,
1098
1099
1100
                    num_cached_tokens=getattr(
                        processed_outputs, "num_cached_tokens", 0
                    ),
1101
1102
1103
                    prompt_token_ids=[],
                    finished=True,
                )
1104
1105
1106
            ]
        else:
            return model_outputs
1107

1108
1109
    def embed(
        self,
1110
        prompts: PromptType | Sequence[PromptType],
1111
        *,
1112
1113
1114
1115
        truncate_prompt_tokens: int | None = None,
        use_tqdm: bool | Callable[..., tqdm] = True,
        pooling_params: PoolingParams | Sequence[PoolingParams] | None = None,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
1116
    ) -> list[EmbeddingRequestOutput]:
1117
1118
1119
1120
1121
1122
1123
1124
1125
        """
        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
1126
                for batch inference. See [PromptType][vllm.inputs.PromptType]
1127
                for more details about the format of each prompt.
1128
1129
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1130
1131
1132
1133
            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.
1134
1135
1136
            lora_request: LoRA request to use for generation, if any.

        Returns:
1137
            A list of `EmbeddingRequestOutput` objects containing the
1138
1139
            embedding vectors in the same order as the input prompts.
        """
1140
        if "embed" not in self.supported_tasks:
1141
1142
            raise ValueError(
                "Embedding API is not supported by this model. "
1143
1144
                "Try converting the model using `--convert embed`."
            )
1145

1146
1147
1148
1149
1150
1151
1152
1153
        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",
        )
1154
1155
1156
1157
1158

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

    def classify(
        self,
1159
        prompts: PromptType | Sequence[PromptType],
1160
        *,
1161
1162
1163
        use_tqdm: bool | Callable[..., tqdm] = True,
        pooling_params: PoolingParams | Sequence[PoolingParams] | None = None,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
1164
    ) -> list[ClassificationRequestOutput]:
1165
1166
1167
1168
1169
1170
1171
1172
1173
        """
        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
1174
                for batch inference. See [PromptType][vllm.inputs.PromptType]
1175
                for more details about the format of each prompt.
1176
1177
1178
1179
            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.
1180
            lora_request: LoRA request to use for generation, if any.
1181
1182
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1183
        Returns:
1184
            A list of `ClassificationRequestOutput` objects containing the
1185
1186
            embedding vectors in the same order as the input prompts.
        """
1187
        if "classify" not in self.supported_tasks:
1188
            raise ValueError(
1189
                "Classification API is not supported by this model. "
1190
1191
                "Try converting the model using `--convert classify`."
            )
1192

1193
1194
1195
        items = self.encode(
            prompts,
            use_tqdm=use_tqdm,
1196
            pooling_params=pooling_params,
1197
1198
1199
            lora_request=lora_request,
            pooling_task="classify",
        )
1200
1201
1202

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

1203
1204
    def reward(
        self,
1205
        prompts: PromptType | Sequence[PromptType],
1206
1207
        /,
        *,
1208
1209
1210
1211
        truncate_prompt_tokens: int | None = None,
        use_tqdm: bool | Callable[..., tqdm] = True,
        pooling_params: PoolingParams | Sequence[PoolingParams] | None = None,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
1212
1213
1214
1215
1216
1217
1218
    ) -> 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]
1219
                for more details about the format of each prompt.
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
            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,
1238
            pooling_task="token_classify",
1239
1240
        )

1241
1242
1243
    def _embedding_score(
        self,
        tokenizer: AnyTokenizer,
1244
1245
1246
1247
1248
1249
        text_1: list[str | TextPrompt | TokensPrompt],
        text_2: list[str | TextPrompt | TokensPrompt],
        truncate_prompt_tokens: int | None = None,
        use_tqdm: bool | Callable[..., tqdm] = True,
        pooling_params: PoolingParams | None = None,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
1250
1251
    ) -> list[ScoringRequestOutput]:
        encoded_output: list[PoolingRequestOutput] = self.encode(
1252
            text_1 + text_2,
1253
            truncate_prompt_tokens=truncate_prompt_tokens,
1254
1255
            use_tqdm=use_tqdm,
            lora_request=lora_request,
1256
            pooling_params=pooling_params,
1257
1258
            pooling_task="embed",
        )
1259

1260
1261
        encoded_output_1: list[PoolingRequestOutput] = encoded_output[0 : len(text_1)]
        encoded_output_2: list[PoolingRequestOutput] = encoded_output[len(text_1) :]
1262
1263
1264
1265

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

1266
1267
1268
        scores = _cosine_similarity(
            tokenizer=tokenizer, embed_1=encoded_output_1, embed_2=encoded_output_2
        )
1269

1270
        items = self.engine_class.validate_outputs(scores, PoolingRequestOutput)
1271
1272
1273
1274
        return [ScoringRequestOutput.from_base(item) for item in items]

    def _cross_encoding_score(
        self,
1275
        tokenizer: AnyTokenizer,
1276
1277
1278
1279
1280
1281
        data_1: list[str] | list[ScoreContentPartParam],
        data_2: list[str] | list[ScoreContentPartParam],
        truncate_prompt_tokens: int | None = None,
        use_tqdm: bool | Callable[..., tqdm] = True,
        pooling_params: PoolingParams | None = None,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
1282
    ) -> list[ScoringRequestOutput]:
1283
        model_config = self.model_config
1284
1285

        if isinstance(tokenizer, MistralTokenizer):
1286
            raise ValueError("Score API is not supported for Mistral tokenizer")
1287

1288
1289
        if len(data_1) == 1:
            data_1 = data_1 * len(data_2)
1290

1291
1292
1293
1294
        if pooling_params is None:
            pooling_params = PoolingParams(task="score")

        pooling_params.verify("score", model_config)
1295
        pooling_params_list = list[PoolingParams]()
1296

1297
        tokenization_kwargs: dict[str, Any] = {}
1298

1299
1300
1301
        _validate_truncation_size(
            model_config.max_model_len, truncate_prompt_tokens, tokenization_kwargs
        )
1302

1303
        prompts = list[PromptType]()
1304

1305
1306
        input_pairs = [(t1, t2) for t1, t2 in zip(data_1, data_2)]

1307
1308
1309
1310
1311
1312
1313
1314
1315
        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,
            )

1316
            if token_type_ids := engine_prompt.pop("token_type_ids", None):
1317
1318
1319
1320
1321
1322
1323
                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)

1324
            prompts.append(engine_prompt)
1325
1326

        self._validate_and_add_requests(
1327
            prompts=prompts,
1328
            params=pooling_params_list,
1329
            use_tqdm=use_tqdm,
1330
1331
1332
1333
            lora_request=lora_request,
        )

        outputs = self._run_engine(use_tqdm=use_tqdm)
1334
        items = self.engine_class.validate_outputs(outputs, PoolingRequestOutput)
1335
1336
1337

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

1338
1339
    def score(
        self,
1340
1341
        data_1: SingletonPrompt | Sequence[SingletonPrompt] | ScoreMultiModalParam,
        data_2: SingletonPrompt | Sequence[SingletonPrompt] | ScoreMultiModalParam,
1342
        /,
1343
        *,
1344
1345
1346
1347
        truncate_prompt_tokens: int | None = None,
        use_tqdm: bool | Callable[..., tqdm] = True,
        pooling_params: PoolingParams | None = None,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
1348
    ) -> list[ScoringRequestOutput]:
1349
1350
        """Generate similarity scores for all pairs `<text,text_pair>` or
          `<multi-modal data, multi-modal data pair>`.
1351

1352
        The inputs can be `1 -> 1`, `1 -> N` or `N -> N`.
1353
1354
        In the `1 - N` case the `data_1` input will be replicated `N`
        times to pair with the `data_2` inputs.
1355
        The input pairs are used to build a list of prompts for the
1356
1357
        cross encoder model. This class automatically batches the prompts,
        considering the memory constraint. For the best performance, put all
1358
1359
1360
        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
1361
        appropriate multi-modal models. For multi-modal inputs, ensure the
1362
        prompt structure matches the model's expected input format.
1363
1364

        Args:
1365
1366
1367
            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
1368
                the `data_2` list.
1369
            data_2: The data to pair with the query to form the input to
1370
                the LLM. Can be text or multi-modal data. See [PromptType]
1371
                [vllm.inputs.PromptType] for more details about the format of
1372
                each prompt.
1373
1374
1375
1376
            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.
1377
            lora_request: LoRA request to use for generation, if any.
1378
1379
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1380
        Returns:
1381
            A list of `ScoringRequestOutput` objects containing the
1382
1383
            generated scores in the same order as the input prompts.
        """
1384
        model_config = self.model_config
1385
        runner_type = model_config.runner_type
1386
        if runner_type != "pooling":
1387
1388
1389
            raise ValueError(
                "LLM.score() is only supported for pooling models. "
                "Try passing `--runner pooling` to use the model as a "
1390
1391
                "pooling model."
            )
1392

1393
1394
        supported_tasks = self.supported_tasks
        if all(t not in supported_tasks for t in ("embed", "classify")):
1395
1396
1397
1398
1399
            raise ValueError(
                "Score API is not supported by this model. "
                "Try converting the model using "
                "`--convert embed` or `--convert classify`."
            )
1400

1401
1402
1403
1404
        if (
            model_config.is_cross_encoder
            and getattr(model_config.hf_config, "num_labels", 0) != 1
        ):
1405
            raise ValueError("Score API is only enabled for num_labels == 1.")
1406
1407
1408
1409

        # 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
1410
        tokenizer = self.get_tokenizer()
1411

1412
        if not model_config.is_multimodal_model:
1413

1414
            def check_data_type(
1415
1416
1417
                data: SingletonPrompt
                | Sequence[SingletonPrompt]
                | ScoreMultiModalParam,
1418
            ):
1419
                if isinstance(data, dict) and "content" in data:
1420
1421
1422
1423
                    raise ValueError(
                        "ScoreMultiModalParam is not supported "
                        f"for {model_config.architecture}"
                    )
1424
1425
1426
1427
1428
1429
1430

            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:
1431
1432
1433
                        raise ValueError(
                            "Multi-modal prompt is not supported for scoring"
                        )
1434
1435
                    elif "prompt_token_ids" in prompt:
                        prompt = tokenizer.decode(
1436
1437
                            cast(TokensPrompt, prompt)["prompt_token_ids"]
                        )
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
                    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]
1466

1467
        if model_config.is_cross_encoder:
1468
1469
1470
1471
1472
1473
            return self._cross_encoding_score(
                tokenizer,
                data_1,  # type: ignore[arg-type]
                data_2,  # type: ignore[arg-type]
                truncate_prompt_tokens,
                use_tqdm,
1474
                pooling_params,
1475
1476
                lora_request,
            )
1477
        else:
1478
1479
            return self._embedding_score(
                tokenizer,
1480
1481
                data_1,  # type: ignore[arg-type]
                data_2,  # type: ignore[arg-type]
1482
1483
                truncate_prompt_tokens,
                use_tqdm,
1484
                pooling_params,
1485
1486
                lora_request,
            )
1487

1488
1489
1490
1491
1492
1493
    def start_profile(self) -> None:
        self.llm_engine.start_profile()

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

1494
1495
    def reset_prefix_cache(self, device: Device | None = None) -> None:
        self.llm_engine.reset_prefix_cache(device)
1496

1497
1498
1499
1500
1501
1502
    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.

1503
        Args:
1504
1505
            level: The sleep level. Level 1 sleep will offload the model
                weights and discard the kv cache. The content of kv cache
1506
                is forgotten. Level 1 sleep is good for sleeping and waking
1507
1508
1509
1510
1511
                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
1512
                sleep is good for sleeping and waking up the engine to run a
1513
                different model or update the model, where previous model
1514
                weights are not needed. It reduces CPU memory pressure.
1515
        """
1516
        self.reset_prefix_cache()
1517
1518
        self.llm_engine.sleep(level=level)

1519
    def wake_up(self, tags: list[str] | None = None):
1520
        """
1521
1522
        Wake up the engine from sleep mode. See the [sleep][vllm.LLM.sleep]
        method for more details.
1523

1524
        Args:
1525
1526
            tags: An optional list of tags to reallocate the engine memory
                for specific memory allocations. Values must be in
1527
                `("weights", "kv_cache")`. If None, all memory is reallocated.
1528
                wake_up should be called with all tags (or None) before the
1529
1530
1531
                engine is used again.
        """
        self.llm_engine.wake_up(tags)
1532

1533
1534
1535
1536
    def get_metrics(self) -> list["Metric"]:
        """Return a snapshot of aggregated metrics from Prometheus.

        Returns:
1537
            A `MetricSnapshot` instance capturing the current state
1538
1539
1540
1541
1542
1543
1544
            of all aggregated metrics from Prometheus.

        Note:
            This method is only available with the V1 LLM engine.
        """
        return self.llm_engine.get_metrics()

1545
1546
    def _validate_and_add_requests(
        self,
1547
1548
1549
1550
1551
        prompts: PromptType | Sequence[PromptType] | DataPrompt,
        params: SamplingParams
        | Sequence[SamplingParams]
        | PoolingParams
        | Sequence[PoolingParams],
1552
        *,
1553
1554
1555
        use_tqdm: bool | Callable[..., tqdm] = True,
        lora_request: Sequence[LoRARequest] | LoRARequest | None,
        priority: list[int] | None = None,
1556
    ) -> None:
1557
        if isinstance(prompts, (str, dict)):
1558
            # Convert a single prompt to a list.
1559
            prompts = [prompts]  # type: ignore[list-item]
1560

1561
        num_requests = len(prompts)
1562
        if isinstance(params, Sequence) and len(params) != num_requests:
1563
1564
1565
1566
1567
1568
1569
            raise ValueError("The lengths of prompts and params must be the same.")
        if isinstance(lora_request, Sequence) and len(lora_request) != num_requests:
            raise ValueError(
                "The lengths of prompts and lora_request must be the same."
            )

        for sp in params if isinstance(params, Sequence) else (params,):
1570
1571
1572
            if isinstance(sp, SamplingParams):
                # We only care about the final output
                sp.output_kind = RequestOutputKind.FINAL_ONLY
1573

Zhuohan Li's avatar
Zhuohan Li committed
1574
        # Add requests to the engine.
1575
1576
        it = prompts
        if use_tqdm:
1577
1578
            tqdm_func = use_tqdm if callable(use_tqdm) else tqdm
            it = tqdm_func(it, desc="Adding requests")
1579
1580

        for i, prompt in enumerate(it):
1581
1582
            if isinstance(prompt, dict):
                self._validate_mm_data_and_uuids(
1583
1584
                    prompt.get("multi_modal_data"), prompt.get("multi_modal_uuids")
                )
1585

1586
            self._add_request(
1587
                prompt,
1588
                params[i] if isinstance(params, Sequence) else params,
1589
1590
1591
                lora_request=lora_request[i]
                if isinstance(lora_request, Sequence)
                else lora_request,
1592
                priority=priority[i] if priority else 0,
nunjunj's avatar
nunjunj committed
1593
            )
1594

1595
    def _validate_mm_data_and_uuids(
1596
        self,
1597
1598
        multi_modal_data: Any | None,  # MultiModalDataDict
        multi_modal_uuids: Any | None,  # MultiModalUUIDDict
1599
1600
1601
    ):
        """
        Validate that if any multi-modal data is skipped (i.e. None),
1602
        then its corresponding UUID must be set.
1603
1604
1605
1606
1607
1608
1609
1610
        """
        if multi_modal_data is None:
            return

        for modality, data in multi_modal_data.items():
            if isinstance(data, list):
                for i, d in enumerate(data):
                    if d is None:
1611
1612
1613
1614
1615
1616
1617
1618
                        if (
                            multi_modal_uuids is None
                            or modality not in multi_modal_uuids
                            or multi_modal_uuids[  # noqa: E501
                                modality
                            ]
                            is None
                        ):
1619
1620
                            raise ValueError(
                                f"Multi-modal data for {modality} is None "
1621
1622
                                f"but UUID is not provided"
                            )
1623
                        else:
1624
1625
1626
1627
                            if (
                                len(multi_modal_uuids[modality]) <= i
                                or multi_modal_uuids[modality][i] is None
                            ):
1628
1629
                                raise ValueError(
                                    f"Multi-modal data for {modality} is None "
1630
1631
                                    f"but UUID is not provided"
                                )
1632
            else:
1633
1634
1635
1636
1637
1638
1639
1640
1641
                if data is None and (
                    multi_modal_uuids is None
                    or modality not in multi_modal_uuids
                    or multi_modal_uuids[modality] is None
                ):
                    raise ValueError(
                        f"Multi-modal data for {modality} is None"
                        f" but UUID is not provided"
                    )
1642

1643
1644
1645
1646
    def _process_inputs(
        self,
        request_id: str,
        engine_prompt: PromptType,
1647
        params: SamplingParams | PoolingParams,
1648
        *,
1649
        lora_request: LoRARequest | None,
1650
1651
1652
1653
        priority: int,
    ) -> tuple[EngineCoreRequest, dict[str, Any]]:
        """Use the Processor to process inputs for LLMEngine."""
        tokenization_kwargs: dict[str, Any] = {}
1654
1655
1656
1657
1658
        _validate_truncation_size(
            self.model_config.max_model_len,
            params.truncate_prompt_tokens,
            tokenization_kwargs,
        )
1659

1660
        engine_request = self.processor.process_inputs(
1661
1662
1663
1664
1665
1666
1667
1668
1669
            request_id,
            engine_prompt,
            params,
            lora_request=lora_request,
            tokenization_kwargs=tokenization_kwargs,
            priority=priority,
        )
        return engine_request, tokenization_kwargs

1670
    def _add_request(
nunjunj's avatar
nunjunj committed
1671
        self,
1672
        prompt: PromptType,
1673
1674
        params: SamplingParams | PoolingParams,
        lora_request: LoRARequest | None = None,
1675
        priority: int = 0,
1676
    ) -> None:
1677
        prompt_text, _, _ = get_prompt_components(prompt)
1678
        request_id = str(next(self.request_counter))
1679
1680

        engine_request, tokenization_kwargs = self._process_inputs(
1681
            request_id,
1682
            prompt,
1683
1684
            params,
            lora_request=lora_request,
1685
1686
1687
1688
1689
1690
1691
1692
            priority=priority,
        )

        self.llm_engine.add_request(
            request_id,
            engine_request,
            params,
            lora_request=lora_request,
1693
            tokenization_kwargs=tokenization_kwargs,
1694
            priority=priority,
1695
            prompt_text=prompt_text,
nunjunj's avatar
nunjunj committed
1696
        )
1697

1698
    def _run_engine(
1699
1700
        self, *, use_tqdm: bool | Callable[..., tqdm] = True
    ) -> list[RequestOutput | PoolingRequestOutput]:
1701
1702
        # Initialize tqdm.
        if use_tqdm:
Zhuohan Li's avatar
Zhuohan Li committed
1703
            num_requests = self.llm_engine.get_num_unfinished_requests()
1704
1705
            tqdm_func = use_tqdm if callable(use_tqdm) else tqdm
            pbar = tqdm_func(
1706
1707
1708
                total=num_requests,
                desc="Processed prompts",
                dynamic_ncols=True,
1709
                postfix=(f"est. speed input: {0:.2f} toks/s, output: {0:.2f} toks/s"),
1710
            )
1711

Zhuohan Li's avatar
Zhuohan Li committed
1712
        # Run the engine.
1713
        outputs: list[RequestOutput | PoolingRequestOutput] = []
1714
1715
        total_in_toks = 0
        total_out_toks = 0
Zhuohan Li's avatar
Zhuohan Li committed
1716
1717
        while self.llm_engine.has_unfinished_requests():
            step_outputs = self.llm_engine.step()
1718
            for output in step_outputs:
1719
                if output.finished:
1720
1721
                    outputs.append(output)
                    if use_tqdm:
1722
1723
                        if isinstance(output, RequestOutput):
                            # Calculate tokens only for RequestOutput
1724
                            n = len(output.outputs)
1725
                            assert output.prompt_token_ids is not None
1726
                            total_in_toks += len(output.prompt_token_ids) * n
1727
1728
                            in_spd = total_in_toks / pbar.format_dict["elapsed"]
                            total_out_toks += sum(
1729
1730
1731
                                len(stp.token_ids) for stp in output.outputs
                            )
                            out_spd = total_out_toks / pbar.format_dict["elapsed"]
1732
1733
                            pbar.postfix = (
                                f"est. speed input: {in_spd:.2f} toks/s, "
1734
1735
                                f"output: {out_spd:.2f} toks/s"
                            )
1736
                            pbar.update(n)
1737
1738
                        else:
                            pbar.update(1)
1739
1740
                        if pbar.n == num_requests:
                            pbar.refresh()
1741

1742
1743
        if use_tqdm:
            pbar.close()
1744
1745
1746
        # Sort the outputs by request ID.
        # This is necessary because some requests may be finished earlier than
        # its previous requests.
1747
        return sorted(outputs, key=lambda x: int(x.request_id))