llm.py 72.4 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.compilation import CompilationMode
27
from vllm.config.model import (
28
29
    ConvertOption,
    HfOverrides,
30
    ModelDType,
31
    RunnerOption,
32
    TokenizerMode,
33
)
34
from vllm.engine.arg_utils import EngineArgs
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.platforms import current_platform
71
from vllm.pooling_params import PoolingParams
72
from vllm.sampling_params import BeamSearchParams, RequestOutputKind, SamplingParams
73
from vllm.tasks import PoolingTask
74
75
76
77
78
from vllm.transformers_utils.tokenizer import (
    AnyTokenizer,
    MistralTokenizer,
    get_cached_tokenizer,
)
yhu422's avatar
yhu422 committed
79
from vllm.usage.usage_lib import UsageContext
80
from vllm.utils.collection_utils import as_iter, is_list_of
81
from vllm.utils.counter import Counter
82
from vllm.v1.engine import EngineCoreRequest
83
from vllm.v1.engine.llm_engine import LLMEngine
84
from vllm.v1.sample.logits_processor import LogitsProcessor
85

86
87
88
if TYPE_CHECKING:
    from vllm.v1.metrics.reader import Metric

89
90
logger = init_logger(__name__)

91
92
_R = TypeVar("_R", default=Any)

93
94

class LLM:
Woosuk Kwon's avatar
Woosuk Kwon committed
95
96
97
98
99
100
101
102
103
104
    """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.
105
        tokenizer: The name or path of a HuggingFace Transformers tokenizer.
106
107
        tokenizer_mode: The tokenizer mode. "auto" will use the fast tokenizer
            if available, and "slow" will always use the slow tokenizer.
108
109
110
        skip_tokenizer_init: If true, skip initialization of tokenizer and
            detokenizer. Expect valid prompt_token_ids and None for prompt
            from the input.
111
112
        trust_remote_code: Trust remote code (e.g., from HuggingFace) when
            downloading the model and tokenizer.
113
114
115
116
        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.
117
        allowed_media_domains: If set, only media URLs that belong to this
118
            domain can be used for multi-modal inputs.
Woosuk Kwon's avatar
Woosuk Kwon committed
119
120
121
        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
122
            we support `float32`, `float16`, and `bfloat16`. If `auto`, we use
123
124
            the `dtype` attribute of the Transformers model's config. However,
            if the `dtype` in the config is `float32`, we will use `float16` instead.
125
        quantization: The method used to quantize the model weights. Currently,
126
            we support "awq", "gptq", and "fp8" (experimental).
127
128
129
130
            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
131
132
        revision: The specific model version to use. It can be a branch name,
            a tag name, or a commit id.
133
134
        tokenizer_revision: The specific tokenizer version to use. It can be a
            branch name, a tag name, or a commit id.
135
136
137
138
139
140
        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.
141
142
143
144
145
        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
146
            compared with using gpu_memory_utilization. Note that
147
148
            kv_cache_memory_bytes (when not-None) ignores
            gpu_memory_utilization
149
        swap_space: The size (GiB) of CPU memory per GPU to use as swap space.
150
151
152
153
154
            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.
155
156
157
158
        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.
159
160
161
        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.
162
163
        disable_custom_all_reduce: See
            [ParallelConfig][vllm.config.ParallelConfig].
164
        hf_token: The token to use as HTTP bearer authorization for remote files
165
            . If `True`, will use the token generated when running
166
            `huggingface-cli login` (stored in `~/.huggingface`).
167
168
169
        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.
170
171
172
173
174
        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}`.
175
176
177
178
179
        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.
180
        compilation_config: Either an integer or a dictionary. If it is an
181
            integer, it is used as the mode of compilation optimization. If it
182
            is a dictionary, it can specify the full compilation configuration.
183
        **kwargs: Arguments for [`EngineArgs`][vllm.EngineArgs].
nunjunj's avatar
nunjunj committed
184

185
186
    Note:
        This class is intended to be used for offline inference. For online
187
        serving, use the [AsyncLLMEngine][vllm.AsyncLLMEngine] class instead.
188
    """
189
190
191
192

    def __init__(
        self,
        model: str,
193
        *,
194
195
        runner: RunnerOption = "auto",
        convert: ConvertOption = "auto",
196
        tokenizer: str | None = None,
197
        tokenizer_mode: TokenizerMode = "auto",
198
        skip_tokenizer_init: bool = False,
199
        trust_remote_code: bool = False,
200
        allowed_local_media_path: str = "",
201
        allowed_media_domains: list[str] | None = None,
202
        tensor_parallel_size: int = 1,
203
        dtype: ModelDType = "auto",
204
205
206
207
        quantization: QuantizationMethods | None = None,
        revision: str | None = None,
        tokenizer_revision: str | None = None,
        seed: int | None = None,
208
        gpu_memory_utilization: float = 0.9,
209
        swap_space: float = 4,
210
        cpu_offload_gb: float = 0,
211
        enforce_eager: bool = False,
212
        disable_custom_all_reduce: bool = False,
213
214
215
216
217
218
219
220
221
222
223
        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,
224
        **kwargs: Any,
225
    ) -> None:
226
        """LLM constructor."""
227

228
229
        if "disable_log_stats" not in kwargs:
            kwargs["disable_log_stats"] = True
230

231
232
233
234
235
236
237
        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)

238
        if "kv_transfer_config" in kwargs and isinstance(
239
240
            kwargs["kv_transfer_config"], dict
        ):
241
            from vllm.config.kv_transfer import KVTransferConfig
242

243
244
            raw_config_dict = kwargs["kv_transfer_config"]
            try:
245
                kwargs["kv_transfer_config"] = KVTransferConfig(**raw_config_dict)
246
247
248
249
            except ValidationError as e:
                logger.error(
                    "Failed to convert 'kv_transfer_config' dict to "
                    "KVTransferConfig object. Dict: %s. Error: %s",
250
251
252
                    raw_config_dict,
                    e,
                )
253
254
                # Consider re-raising a more specific vLLM error or ValueError
                # to provide better context to the user.
255
                raise ValueError(f"Invalid 'kv_transfer_config' provided: {e}") from e
256

257
258
259
        if hf_overrides is None:
            hf_overrides = {}

260
        if compilation_config is not None:
261
            if isinstance(compilation_config, int):
262
263
264
                compilation_config_instance = CompilationConfig(
                    mode=CompilationMode(compilation_config)
                )
265
266
            elif isinstance(compilation_config, dict):
                compilation_config_instance = CompilationConfig(
267
268
269
270
                    **{
                        k: v
                        for k, v in compilation_config.items()
                        if is_init_field(CompilationConfig, k)
271
272
                    }
                )
273
274
            else:
                compilation_config_instance = compilation_config
275
        else:
276
            compilation_config_instance = CompilationConfig()
277

278
279
280
281
282
283
284
        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)
285
286
                    }
                )
287
288
289
290
291
            else:
                structured_outputs_instance = structured_outputs_config
        else:
            structured_outputs_instance = StructuredOutputsConfig()

292
        # warn about single-process data parallel usage.
293
294
        _dp_size = int(kwargs.get("data_parallel_size", 1))
        _distributed_executor_backend = kwargs.get("distributed_executor_backend")
295
296
297
298
299
        if (
            _dp_size > 1
            and not _distributed_executor_backend == "external_launcher"
            and not current_platform.is_tpu()
        ):
300
            raise ValueError(
301
                f"LLM(data_parallel_size={_dp_size}) is not supported for single-"
302
303
304
305
306
                "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
307
        engine_args = EngineArgs(
308
            model=model,
309
310
            runner=runner,
            convert=convert,
311
            tokenizer=tokenizer,
312
            tokenizer_mode=tokenizer_mode,
313
            skip_tokenizer_init=skip_tokenizer_init,
314
            trust_remote_code=trust_remote_code,
315
            allowed_local_media_path=allowed_local_media_path,
316
            allowed_media_domains=allowed_media_domains,
317
318
            tensor_parallel_size=tensor_parallel_size,
            dtype=dtype,
319
            quantization=quantization,
320
            revision=revision,
321
            tokenizer_revision=tokenizer_revision,
322
323
            seed=seed,
            gpu_memory_utilization=gpu_memory_utilization,
324
            kv_cache_memory_bytes=kv_cache_memory_bytes,
325
            swap_space=swap_space,
326
            cpu_offload_gb=cpu_offload_gb,
327
            enforce_eager=enforce_eager,
328
            disable_custom_all_reduce=disable_custom_all_reduce,
329
            hf_token=hf_token,
330
            hf_overrides=hf_overrides,
331
            mm_processor_kwargs=mm_processor_kwargs,
332
            pooler_config=pooler_config,
333
            override_pooler_config=override_pooler_config,
334
            structured_outputs_config=structured_outputs_instance,
335
            compilation_config=compilation_config_instance,
336
            logits_processors=logits_processors,
337
338
            **kwargs,
        )
339

340
341
        log_non_default_args(engine_args)

342
343
        # Create the Engine (autoselects V0 vs V1)
        self.llm_engine = LLMEngine.from_engine_args(
344
345
            engine_args=engine_args, usage_context=UsageContext.LLM_CLASS
        )
346
        self.engine_class = type(self.llm_engine)
347

348
        self.request_counter = Counter()
349
        self.default_sampling_params: dict[str, Any] | None = None
350

351
352
        supported_tasks = self.llm_engine.get_supported_tasks()
        logger.info("Supported tasks: %s", supported_tasks)
353
354
        self.supported_tasks = supported_tasks

355
356
357
        self.model_config = self.llm_engine.model_config
        self.processor = self.llm_engine.processor
        self.io_processor = self.llm_engine.io_processor
358

359
360
    def get_tokenizer(self) -> AnyTokenizer:
        return self.llm_engine.get_tokenizer()
361

362
    @deprecated("`set_tokenizer` is deprecated and will be removed in v0.13.")
363
    def set_tokenizer(self, tokenizer: AnyTokenizer) -> None:
364
365
366
367
        # 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"):
368
            self.llm_engine.tokenizer = tokenizer
369
        else:
370
            self.llm_engine.tokenizer = get_cached_tokenizer(tokenizer)
371

372
373
374
375
    def reset_mm_cache(self) -> None:
        self.processor.clear_mm_cache()
        self.llm_engine.reset_mm_cache()

376
    def get_default_sampling_params(self) -> SamplingParams:
377
        if self.default_sampling_params is None:
378
            self.default_sampling_params = self.model_config.get_diff_sampling_param()
379
380
        if self.default_sampling_params:
            return SamplingParams.from_optional(**self.default_sampling_params)
381
382
        return SamplingParams()

383
384
    def generate(
        self,
385
386
        prompts: PromptType | Sequence[PromptType],
        sampling_params: SamplingParams | Sequence[SamplingParams] | None = None,
387
        *,
388
389
390
        use_tqdm: bool | Callable[..., tqdm] = True,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
        priority: list[int] | None = None,
391
    ) -> list[RequestOutput]:
Woosuk Kwon's avatar
Woosuk Kwon committed
392
393
        """Generates the completions for the input prompts.

394
        This class automatically batches the given prompts, considering
Woosuk Kwon's avatar
Woosuk Kwon committed
395
396
397
398
        the memory constraint. For the best performance, put all of your prompts
        into a single list and pass it to this method.

        Args:
399
            prompts: The prompts to the LLM. You may pass a sequence of prompts
400
                for batch inference. See [PromptType][vllm.inputs.PromptType]
401
                for more details about the format of each prompt.
Woosuk Kwon's avatar
Woosuk Kwon committed
402
            sampling_params: The sampling parameters for text generation. If
nunjunj's avatar
nunjunj committed
403
404
405
                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
406
                prompts and it is paired one by one with the prompt.
407
408
409
410
            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.
411
            lora_request: LoRA request to use for generation, if any.
412
413
            priority: The priority of the requests, if any.
                Only applicable when priority scheduling policy is enabled.
Woosuk Kwon's avatar
Woosuk Kwon committed
414
415

        Returns:
416
            A list of `RequestOutput` objects containing the
417
            generated completions in the same order as the input prompts.
418

419
420
421
422
        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.
423
        """
424
        model_config = self.model_config
425
426
        runner_type = model_config.runner_type
        if runner_type != "generate":
427
428
429
            raise ValueError(
                "LLM.generate() is only supported for generative models. "
                "Try passing `--runner generate` to use the model as a "
430
431
                "generative model."
            )
432

433
434
        if sampling_params is None:
            # Use default sampling params.
435
            sampling_params = self.get_default_sampling_params()
436

437
        # Add any modality specific loras to the corresponding prompts
438
        lora_request = self._get_modality_specific_lora_reqs(prompts, lora_request)
439

440
        self._validate_and_add_requests(
441
            prompts=prompts,
442
            params=sampling_params,
443
            use_tqdm=use_tqdm,
444
            lora_request=lora_request,
445
446
            priority=priority,
        )
447

448
        outputs = self._run_engine(use_tqdm=use_tqdm)
Joe Runde's avatar
Joe Runde committed
449
        return self.engine_class.validate_outputs(outputs, RequestOutput)
450

451
    def _get_modality_specific_lora_reqs(
452
        self,
453
454
        prompts: PromptType | Sequence[PromptType],
        lora_request: list[LoRARequest] | LoRARequest | None,
455
    ):
456
457
458
459
460
461
        # 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.
462
463
        if (
            lora_config is None
464
            or not self.model_config.is_multimodal_model
465
466
            or (lora_config and lora_config.default_mm_loras is None)
        ):
467
468
            return lora_request

469
470
        if not isinstance(prompts, Sequence):
            prompts = [prompts]
471

472
473
474
475
476
        optional_loras = (
            [lora_request] * len(prompts)
            if not isinstance(lora_request, Sequence)
            else lora_request
        )
477
478
479

        return [
            self._resolve_single_prompt_mm_lora(
480
                prompt,
481
482
                opt_lora_req,
                lora_config.default_mm_loras,
483
484
            )
            for prompt, opt_lora_req in zip(prompts, optional_loras)
485
486
        ]

487
488
489
    def _resolve_single_prompt_mm_lora(
        self,
        prompt: PromptType,
490
491
        lora_request: LoRARequest | None,
        default_mm_loras: dict[str, str] | None,
492
493
494
495
    ):
        if (
            not default_mm_loras
            or not isinstance(prompt, dict)
496
            or not (mm_data := prompt.get("multi_modal_data") or {})
497
        ):
498
499
            return lora_request

500
501
502
        intersection = set(
            mm_data.keys()  # type: ignore
        ).intersection(default_mm_loras.keys())
503
504
505
506
507
508
509
510
511
        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"
512
513
514
                " will be skipped",
                intersection,
            )
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
            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 "
530
531
                    "lora_request as we only apply one LoRARequest per prompt"
                )
532
533
534
535
536
537
538
539
            return lora_request

        return LoRARequest(
            modality_name,
            modality_lora_id,
            modality_lora_path,
        )

540
541
    def collective_rpc(
        self,
542
543
        method: str | Callable[..., _R],
        timeout: float | None = None,
544
        args: tuple = (),
545
        kwargs: dict[str, Any] | None = None,
546
    ) -> list[_R]:
547
548
549
550
551
552
553
554
555
556
557
        """
        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
558
                [`TimeoutError`][] on timeout. `None` means wait indefinitely.
559
560
561
562
563
            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.
564

565
566
567
        Note:
            It is recommended to use this API to only pass control messages,
            and set up data-plane communication to pass data.
568
        """
569
570

        return self.llm_engine.collective_rpc(method, timeout, args, kwargs)
571
572

    def apply_model(self, func: Callable[[nn.Module], _R]) -> list[_R]:
573
        """
574
575
        Run a function directly on the model inside each worker,
        returning the result for each of them.
576
577
578
579
580
581

        !!! 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!
582
        """
583
        return self.llm_engine.apply_model(func)
584

585
586
    def _get_beam_search_lora_requests(
        self,
587
588
589
        lora_request: list[LoRARequest] | LoRARequest | None,
        prompts: list[TokensPrompt | TextPrompt],
    ) -> list[LoRARequest | None]:
590
        """Get the optional lora request corresponding to each prompt."""
591
        if isinstance(lora_request, Sequence) and len(lora_request) != len(prompts):
592
            raise ValueError(
593
594
                "Lora request list should be the same length as the prompts"
            )
595
596
597
598
599
600

        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)}")

601
602
    def beam_search(
        self,
603
        prompts: list[TokensPrompt | TextPrompt],
604
        params: BeamSearchParams,
605
        lora_request: list[LoRARequest] | LoRARequest | None = None,
606
        use_tqdm: bool = False,
607
        concurrency_limit: int | None = None,
608
    ) -> list[BeamSearchOutput]:
609
610
611
612
613
614
        """
        Generate sequences using beam search.

        Args:
            prompts: A list of prompts. Each prompt can be a string or a list
                of token IDs.
615
            params: The beam search parameters.
616
            lora_request: LoRA request to use for generation, if any.
617
            use_tqdm: Whether to use tqdm to display the progress bar.
618
619
            concurrency_limit: The maximum number of concurrent requests.
                If None, the number of concurrent requests is unlimited.
620
        """
621
622
        # TODO: how does beam search work together with length penalty,
        # frequency, penalty, and stopping criteria, etc.?
623
624
625
626
        beam_width = params.beam_width
        max_tokens = params.max_tokens
        temperature = params.temperature
        ignore_eos = params.ignore_eos
627
628
        length_penalty = params.length_penalty

629
        lora_requests = self._get_beam_search_lora_requests(lora_request, prompts)
630

631
632
633
634
635
        tokenizer = self.get_tokenizer()
        sort_beams_key = create_sort_beams_key_function(
            tokenizer.eos_token_id,
            length_penalty,
        )
636

637
638
639
        if use_tqdm and concurrency_limit is not None:
            logger.warning(
                "Progress bar is not supported when using concurrency_limit. "
640
641
                "Disabling progress bar."
            )
642
643
644
645
646
            use_tqdm = False

        if concurrency_limit is None:
            concurrency_limit = len(prompts)

647
648
        def create_tokens_prompt_from_beam(beam: BeamSearchSequence) -> TokensPrompt:
            token_prompt_kwargs: TokensPrompt = {"prompt_token_ids": beam.tokens}
649
650
651
652
            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:
653
                token_prompt_kwargs["mm_processor_kwargs"] = beam.mm_processor_kwargs
654
            return TokensPrompt(**token_prompt_kwargs)
655

656
657
658
        # 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
659
660
661
        beam_search_params = SamplingParams(
            logprobs=2 * beam_width, max_tokens=1, temperature=temperature
        )
662
        instances: list[BeamSearchInstance] = []
663

664
        for lora_req, prompt in zip(lora_requests, prompts):
665
666
667
668
669
            # 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:
670
                mm_kwargs["mm_processor_kwargs"] = prompt["mm_processor_kwargs"]
671

672
673
            if "prompt_token_ids" in prompt:
                prompt = cast(TokensPrompt, prompt)  # Needed for mypy
674
675
676
                prompt_tokens = prompt["prompt_token_ids"]
            else:
                prompt_tokens = tokenizer.encode(prompt["prompt"])
677

678
            instances.append(
679
680
681
682
683
                BeamSearchInstance(
                    prompt_tokens,
                    lora_request=lora_req,
                    logprobs=None,
                    **mm_kwargs,
684
685
                ),
            )
686

687
        for prompt_start in range(0, len(prompts), concurrency_limit):
688
            instances_batch = instances[prompt_start : prompt_start + concurrency_limit]
689
690
691

            token_iter = range(max_tokens)
            if use_tqdm:
692
693
694
                token_iter = tqdm(
                    token_iter, desc="Beam search", unit="token", unit_scale=False
                )
695
696
697
                logger.warning(
                    "The progress bar shows the upper bound on token steps and "
                    "may finish early due to stopping conditions. It does not "
698
699
                    "reflect instance-level progress."
                )
700
701
            for _ in token_iter:
                all_beams: list[BeamSearchSequence] = list(
702
703
                    sum((instance.beams for instance in instances_batch), [])
                )
704
705
                pos = [0] + list(
                    itertools.accumulate(
706
707
708
                        len(instance.beams) for instance in instances_batch
                    )
                )
709
                instance_start_and_end: list[tuple[int, int]] = list(
710
711
                    zip(pos[:-1], pos[1:])
                )
712
713
714
715
716
717

                if len(all_beams) == 0:
                    break

                # create corresponding batch entries for prompt & optional lora
                prompts_batch, lora_req_batch = zip(
718
719
720
721
722
                    *[
                        (create_tokens_prompt_from_beam(beam), beam.lora_request)
                        for beam in all_beams
                    ]
                )
723
724
725

                # only runs for one step
                # we don't need to use tqdm here
726
727
728
729
730
731
                output = self.generate(
                    prompts_batch,
                    sampling_params=beam_search_params,
                    use_tqdm=False,
                    lora_request=lora_req_batch,
                )
732

733
734
735
                for (start, end), instance in zip(
                    instance_start_and_end, instances_batch
                ):
736
737
738
739
740
741
742
743
744
745
746
747
748
749
                    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],
750
                                    logprobs=current_beam.logprobs + [logprobs],
751
                                    lora_request=current_beam.lora_request,
752
753
754
755
756
757
758
759
760
761
                                    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
                                ):
762
763
764
                                    instance.completed.append(new_beam)
                                else:
                                    instance_new_beams.append(new_beam)
765
766
767
                    sorted_beams = sorted(
                        instance_new_beams, key=sort_beams_key, reverse=True
                    )
768
                    instance.beams = sorted_beams[:beam_width]
769
770
771
772

        outputs = []
        for instance in instances:
            instance.completed.extend(instance.beams)
773
774
775
            sorted_completed = sorted(
                instance.completed, key=sort_beams_key, reverse=True
            )
776
777
778
779
780
781
782
783
            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

784
    def preprocess_chat(
nunjunj's avatar
nunjunj committed
785
        self,
786
787
788
        messages: list[ChatCompletionMessageParam]
        | list[list[ChatCompletionMessageParam]],
        chat_template: str | None = None,
789
        chat_template_content_format: ChatTemplateContentFormatOption = "auto",
790
        add_generation_prompt: bool = True,
791
        continue_final_message: bool = False,
792
793
794
        tools: list[dict[str, Any]] | None = None,
        chat_template_kwargs: dict[str, Any] | None = None,
        mm_processor_kwargs: dict[str, Any] | None = None,
795
    ) -> list[TokensPrompt]:
nunjunj's avatar
nunjunj committed
796
        """
797
798
        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
799

800
        Refer to `chat` for a complete description of the arguments.
nunjunj's avatar
nunjunj committed
801
        Returns:
802
803
804
            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
805
        """
806
        list_of_messages: list[list[ChatCompletionMessageParam]]
nunjunj's avatar
nunjunj committed
807

808
809
        # Handle multi and single conversations
        if is_list_of(messages, list):
810
            # messages is list[list[...]]
811
            list_of_messages = cast(list[list[ChatCompletionMessageParam]], messages)
812
        else:
813
            # messages is list[...]
814
            list_of_messages = [cast(list[ChatCompletionMessageParam], messages)]
815

816
        tokenizer = self.get_tokenizer()
817
        model_config = self.model_config
818
819
        resolved_content_format = resolve_chat_template_content_format(
            chat_template,
820
            tools,
821
822
            chat_template_content_format,
            tokenizer,
823
            model_config=model_config,
824
825
        )

826
827
828
829
830
831
832
833
        _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 {})

834
        prompts: list[TokensPrompt] = []
835
836

        for msgs in list_of_messages:
837
838
839
            # 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.
840
            conversation, mm_data, mm_uuids = parse_chat_messages(
841
842
843
844
845
                msgs,
                model_config,
                tokenizer,
                content_format=resolved_content_format,
            )
846
847

            if isinstance(tokenizer, MistralTokenizer):
848
                prompt_token_ids = apply_mistral_chat_template(
849
850
                    tokenizer,
                    messages=msgs,
851
                    **_chat_template_kwargs,
852
853
                )
            else:
854
                prompt_str = apply_hf_chat_template(
855
                    tokenizer=tokenizer,
856
                    conversation=conversation,
857
                    model_config=model_config,
858
                    **_chat_template_kwargs,
859
                )
860
861
                # Special tokens are already included in chat templates so
                # should not be added by the tokenizer in this case.
862
863
864
                prompt_token_ids = tokenizer.encode(
                    prompt_str, add_special_tokens=False
                )
865

866
            prompt = TokensPrompt(prompt_token_ids=prompt_token_ids)
867
868
869
870

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

871
872
873
            if mm_uuids is not None:
                prompt["multi_modal_uuids"] = mm_uuids

874
875
876
            if mm_processor_kwargs is not None:
                prompt["mm_processor_kwargs"] = mm_processor_kwargs

877
            prompts.append(prompt)
878

879
880
881
882
        return prompts

    def chat(
        self,
883
884
885
886
887
888
        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,
889
890
891
        chat_template_content_format: ChatTemplateContentFormatOption = "auto",
        add_generation_prompt: bool = True,
        continue_final_message: bool = False,
892
893
894
        tools: list[dict[str, Any]] | None = None,
        chat_template_kwargs: dict[str, Any] | None = None,
        mm_processor_kwargs: dict[str, Any] | None = None,
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
951
952
953
954
955
956
957
    ) -> 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
958
        return self.generate(
959
            prompts,
960
            sampling_params=sampling_params,
nunjunj's avatar
nunjunj committed
961
962
963
964
            use_tqdm=use_tqdm,
            lora_request=lora_request,
        )

965
966
    def encode(
        self,
967
968
        prompts: PromptType | Sequence[PromptType] | DataPrompt,
        pooling_params: PoolingParams | Sequence[PoolingParams] | None = None,
969
        *,
970
971
972
        truncate_prompt_tokens: int | None = None,
        use_tqdm: bool | Callable[..., tqdm] = True,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
973
        pooling_task: PoolingTask | None = None,
974
        tokenization_kwargs: dict[str, Any] | None = None,
975
    ) -> list[PoolingRequestOutput]:
976
977
        """Apply pooling to the hidden states corresponding to the input
        prompts.
978

979
        This class automatically batches the given prompts, considering
980
981
982
983
        the memory constraint. For the best performance, put all of your prompts
        into a single list and pass it to this method.

        Args:
984
            prompts: The prompts to the LLM. You may pass a sequence of prompts
985
                for batch inference. See [PromptType][vllm.inputs.PromptType]
986
                for more details about the format of each prompt.
987
988
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
989
990
991
992
            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.
993
            lora_request: LoRA request to use for generation, if any.
994
            pooling_task: Override the pooling task to use.
995
996
            tokenization_kwargs: overrides tokenization_kwargs set in
                pooling_params
997
998

        Returns:
999
            A list of `PoolingRequestOutput` objects containing the
1000
            pooled hidden states in the same order as the input prompts.
1001

1002
1003
1004
1005
        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.
1006
        """
1007

1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
        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"`'
        )
1023

1024
        if pooling_task is None:
1025
            raise ValueError(error_str)
1026

1027
        model_config = self.model_config
1028
        runner_type = model_config.runner_type
1029
        if runner_type != "pooling":
1030
1031
1032
            raise ValueError(
                "LLM.encode() is only supported for pooling models. "
                "Try passing `--runner pooling` to use the model as a "
1033
1034
                "pooling model."
            )
1035

1036
1037
1038
1039
1040
1041
1042
1043
        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' "
1044
1045
                    "offline inference example for more details."
                )
1046
1047
1048
1049
1050
1051
1052

            # 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)

1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
        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

1081
        self._validate_and_add_requests(
1082
            prompts=prompts,
1083
            params=pooling_params,
1084
            use_tqdm=use_tqdm,
1085
            lora_request=lora_request,
1086
1087
        )

1088
        outputs = self._run_engine(use_tqdm=use_tqdm)
1089
1090

        model_outputs = self.engine_class.validate_outputs(
1091
1092
            outputs, PoolingRequestOutput
        )
1093
1094
1095
1096
1097

        if io_processor_prompt:
            # get the post-processed model outputs
            assert self.io_processor is not None
            processed_outputs = self.io_processor.post_process(
1098
1099
                model_output=model_outputs
            )
1100
1101

            return [
1102
1103
1104
                PoolingRequestOutput[Any](
                    request_id="",
                    outputs=processed_outputs,
1105
1106
1107
                    num_cached_tokens=getattr(
                        processed_outputs, "num_cached_tokens", 0
                    ),
1108
1109
1110
                    prompt_token_ids=[],
                    finished=True,
                )
1111
1112
1113
            ]
        else:
            return model_outputs
1114

1115
1116
    def embed(
        self,
1117
        prompts: PromptType | Sequence[PromptType],
1118
        *,
1119
1120
1121
1122
        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,
1123
    ) -> list[EmbeddingRequestOutput]:
1124
1125
1126
1127
1128
1129
1130
1131
1132
        """
        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
1133
                for batch inference. See [PromptType][vllm.inputs.PromptType]
1134
                for more details about the format of each prompt.
1135
1136
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1137
1138
1139
1140
            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.
1141
1142
1143
            lora_request: LoRA request to use for generation, if any.

        Returns:
1144
            A list of `EmbeddingRequestOutput` objects containing the
1145
1146
            embedding vectors in the same order as the input prompts.
        """
1147
        if "embed" not in self.supported_tasks:
1148
1149
            raise ValueError(
                "Embedding API is not supported by this model. "
1150
1151
                "Try converting the model using `--convert embed`."
            )
1152

1153
1154
1155
1156
1157
1158
1159
1160
        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",
        )
1161
1162
1163
1164
1165

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

    def classify(
        self,
1166
        prompts: PromptType | Sequence[PromptType],
1167
        *,
1168
1169
1170
        use_tqdm: bool | Callable[..., tqdm] = True,
        pooling_params: PoolingParams | Sequence[PoolingParams] | None = None,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
1171
    ) -> list[ClassificationRequestOutput]:
1172
1173
1174
1175
1176
1177
1178
1179
1180
        """
        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
1181
                for batch inference. See [PromptType][vllm.inputs.PromptType]
1182
                for more details about the format of each prompt.
1183
1184
1185
1186
            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.
1187
            lora_request: LoRA request to use for generation, if any.
1188
1189
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1190
        Returns:
1191
            A list of `ClassificationRequestOutput` objects containing the
1192
1193
            embedding vectors in the same order as the input prompts.
        """
1194
        if "classify" not in self.supported_tasks:
1195
            raise ValueError(
1196
                "Classification API is not supported by this model. "
1197
1198
                "Try converting the model using `--convert classify`."
            )
1199

1200
1201
1202
        items = self.encode(
            prompts,
            use_tqdm=use_tqdm,
1203
            pooling_params=pooling_params,
1204
1205
1206
            lora_request=lora_request,
            pooling_task="classify",
        )
1207
1208
1209

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

1210
1211
    def reward(
        self,
1212
        prompts: PromptType | Sequence[PromptType],
1213
1214
        /,
        *,
1215
1216
1217
1218
        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,
1219
1220
1221
1222
1223
1224
1225
    ) -> 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]
1226
                for more details about the format of each prompt.
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
            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,
1245
            pooling_task="token_classify",
1246
1247
        )

1248
1249
1250
    def _embedding_score(
        self,
        tokenizer: AnyTokenizer,
1251
1252
1253
1254
1255
1256
        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,
1257
1258
    ) -> list[ScoringRequestOutput]:
        encoded_output: list[PoolingRequestOutput] = self.encode(
1259
            text_1 + text_2,
1260
            truncate_prompt_tokens=truncate_prompt_tokens,
1261
1262
            use_tqdm=use_tqdm,
            lora_request=lora_request,
1263
            pooling_params=pooling_params,
1264
1265
            pooling_task="embed",
        )
1266

1267
1268
        encoded_output_1: list[PoolingRequestOutput] = encoded_output[0 : len(text_1)]
        encoded_output_2: list[PoolingRequestOutput] = encoded_output[len(text_1) :]
1269
1270
1271
1272

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

1273
1274
1275
        scores = _cosine_similarity(
            tokenizer=tokenizer, embed_1=encoded_output_1, embed_2=encoded_output_2
        )
1276

1277
        items = self.engine_class.validate_outputs(scores, PoolingRequestOutput)
1278
1279
1280
1281
        return [ScoringRequestOutput.from_base(item) for item in items]

    def _cross_encoding_score(
        self,
1282
        tokenizer: AnyTokenizer,
1283
1284
1285
1286
1287
1288
        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,
1289
    ) -> list[ScoringRequestOutput]:
1290
        model_config = self.model_config
1291
1292

        if isinstance(tokenizer, MistralTokenizer):
1293
            raise ValueError("Score API is not supported for Mistral tokenizer")
1294

1295
1296
        if len(data_1) == 1:
            data_1 = data_1 * len(data_2)
1297

1298
1299
1300
1301
        if pooling_params is None:
            pooling_params = PoolingParams(task="score")

        pooling_params.verify("score", model_config)
1302
        pooling_params_list = list[PoolingParams]()
1303

1304
        tokenization_kwargs: dict[str, Any] = {}
1305

1306
1307
1308
        _validate_truncation_size(
            model_config.max_model_len, truncate_prompt_tokens, tokenization_kwargs
        )
1309

1310
        prompts = list[PromptType]()
1311

1312
1313
        input_pairs = [(t1, t2) for t1, t2 in zip(data_1, data_2)]

1314
1315
1316
1317
1318
1319
1320
1321
1322
        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,
            )

1323
            if token_type_ids := engine_prompt.pop("token_type_ids", None):
1324
1325
1326
1327
1328
1329
1330
                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)

1331
            prompts.append(engine_prompt)
1332
1333

        self._validate_and_add_requests(
1334
            prompts=prompts,
1335
            params=pooling_params_list,
1336
            use_tqdm=use_tqdm,
1337
1338
1339
1340
            lora_request=lora_request,
        )

        outputs = self._run_engine(use_tqdm=use_tqdm)
1341
        items = self.engine_class.validate_outputs(outputs, PoolingRequestOutput)
1342
1343
1344

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

1345
1346
    def score(
        self,
1347
1348
        data_1: SingletonPrompt | Sequence[SingletonPrompt] | ScoreMultiModalParam,
        data_2: SingletonPrompt | Sequence[SingletonPrompt] | ScoreMultiModalParam,
1349
        /,
1350
        *,
1351
1352
1353
1354
        truncate_prompt_tokens: int | None = None,
        use_tqdm: bool | Callable[..., tqdm] = True,
        pooling_params: PoolingParams | None = None,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
1355
    ) -> list[ScoringRequestOutput]:
1356
1357
        """Generate similarity scores for all pairs `<text,text_pair>` or
          `<multi-modal data, multi-modal data pair>`.
1358

1359
        The inputs can be `1 -> 1`, `1 -> N` or `N -> N`.
1360
1361
        In the `1 - N` case the `data_1` input will be replicated `N`
        times to pair with the `data_2` inputs.
1362
        The input pairs are used to build a list of prompts for the
1363
1364
        cross encoder model. This class automatically batches the prompts,
        considering the memory constraint. For the best performance, put all
1365
1366
1367
        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
1368
        appropriate multi-modal models. For multi-modal inputs, ensure the
1369
        prompt structure matches the model's expected input format.
1370
1371

        Args:
1372
1373
1374
            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
1375
                the `data_2` list.
1376
            data_2: The data to pair with the query to form the input to
1377
                the LLM. Can be text or multi-modal data. See [PromptType]
1378
                [vllm.inputs.PromptType] for more details about the format of
1379
                each prompt.
1380
1381
1382
1383
            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.
1384
            lora_request: LoRA request to use for generation, if any.
1385
1386
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1387
        Returns:
1388
            A list of `ScoringRequestOutput` objects containing the
1389
1390
            generated scores in the same order as the input prompts.
        """
1391
        model_config = self.model_config
1392
        runner_type = model_config.runner_type
1393
        if runner_type != "pooling":
1394
1395
1396
            raise ValueError(
                "LLM.score() is only supported for pooling models. "
                "Try passing `--runner pooling` to use the model as a "
1397
1398
                "pooling model."
            )
1399

1400
1401
        supported_tasks = self.supported_tasks
        if all(t not in supported_tasks for t in ("embed", "classify")):
1402
1403
1404
1405
1406
            raise ValueError(
                "Score API is not supported by this model. "
                "Try converting the model using "
                "`--convert embed` or `--convert classify`."
            )
1407

1408
1409
1410
1411
        if (
            model_config.is_cross_encoder
            and getattr(model_config.hf_config, "num_labels", 0) != 1
        ):
1412
            raise ValueError("Score API is only enabled for num_labels == 1.")
1413
1414
1415
1416

        # 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
1417
        tokenizer = self.get_tokenizer()
1418

1419
        if not model_config.is_multimodal_model:
1420

1421
            def check_data_type(
1422
1423
1424
                data: SingletonPrompt
                | Sequence[SingletonPrompt]
                | ScoreMultiModalParam,
1425
            ):
1426
                if isinstance(data, dict) and "content" in data:
1427
1428
1429
1430
                    raise ValueError(
                        "ScoreMultiModalParam is not supported "
                        f"for {model_config.architecture}"
                    )
1431
1432
1433
1434
1435
1436
1437

            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:
1438
1439
1440
                        raise ValueError(
                            "Multi-modal prompt is not supported for scoring"
                        )
1441
1442
                    elif "prompt_token_ids" in prompt:
                        prompt = tokenizer.decode(
1443
1444
                            cast(TokensPrompt, prompt)["prompt_token_ids"]
                        )
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
                    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]
1473

1474
        if model_config.is_cross_encoder:
1475
1476
1477
1478
1479
1480
            return self._cross_encoding_score(
                tokenizer,
                data_1,  # type: ignore[arg-type]
                data_2,  # type: ignore[arg-type]
                truncate_prompt_tokens,
                use_tqdm,
1481
                pooling_params,
1482
1483
                lora_request,
            )
1484
        else:
1485
1486
            return self._embedding_score(
                tokenizer,
1487
1488
                data_1,  # type: ignore[arg-type]
                data_2,  # type: ignore[arg-type]
1489
1490
                truncate_prompt_tokens,
                use_tqdm,
1491
                pooling_params,
1492
1493
                lora_request,
            )
1494

1495
1496
1497
1498
1499
1500
    def start_profile(self) -> None:
        self.llm_engine.start_profile()

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

1501
1502
    def reset_prefix_cache(self) -> None:
        self.llm_engine.reset_prefix_cache()
1503

1504
1505
1506
1507
1508
1509
    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.

1510
        Args:
1511
1512
            level: The sleep level. Level 1 sleep will offload the model
                weights and discard the kv cache. The content of kv cache
1513
                is forgotten. Level 1 sleep is good for sleeping and waking
1514
1515
1516
1517
1518
                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
1519
                sleep is good for sleeping and waking up the engine to run a
1520
                different model or update the model, where previous model
1521
                weights are not needed. It reduces CPU memory pressure.
1522
        """
1523
        self.reset_prefix_cache()
1524
1525
        self.llm_engine.sleep(level=level)

1526
    def wake_up(self, tags: list[str] | None = None):
1527
        """
1528
1529
        Wake up the engine from sleep mode. See the [sleep][vllm.LLM.sleep]
        method for more details.
1530

1531
        Args:
1532
1533
            tags: An optional list of tags to reallocate the engine memory
                for specific memory allocations. Values must be in
1534
                `("weights", "kv_cache")`. If None, all memory is reallocated.
1535
                wake_up should be called with all tags (or None) before the
1536
1537
1538
                engine is used again.
        """
        self.llm_engine.wake_up(tags)
1539

1540
1541
1542
1543
    def get_metrics(self) -> list["Metric"]:
        """Return a snapshot of aggregated metrics from Prometheus.

        Returns:
1544
            A `MetricSnapshot` instance capturing the current state
1545
1546
1547
1548
1549
1550
1551
            of all aggregated metrics from Prometheus.

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

1552
1553
    def _validate_and_add_requests(
        self,
1554
1555
1556
1557
1558
        prompts: PromptType | Sequence[PromptType] | DataPrompt,
        params: SamplingParams
        | Sequence[SamplingParams]
        | PoolingParams
        | Sequence[PoolingParams],
1559
        *,
1560
1561
1562
        use_tqdm: bool | Callable[..., tqdm] = True,
        lora_request: Sequence[LoRARequest] | LoRARequest | None,
        priority: list[int] | None = None,
1563
    ) -> None:
1564
        if isinstance(prompts, (str, dict)):
1565
            # Convert a single prompt to a list.
1566
            prompts = [prompts]  # type: ignore[list-item]
1567

1568
        num_requests = len(prompts)
1569
        if isinstance(params, Sequence) and len(params) != num_requests:
1570
1571
1572
1573
1574
            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."
            )
1575
1576
1577
1578
1579
1580
        if priority is not None and len(priority) != num_requests:
            raise ValueError(
                "The lengths of prompts "
                f"({num_requests}) and priority ({len(priority)}) "
                "must be the same."
            )
1581
1582

        for sp in params if isinstance(params, Sequence) else (params,):
1583
1584
1585
            if isinstance(sp, SamplingParams):
                # We only care about the final output
                sp.output_kind = RequestOutputKind.FINAL_ONLY
1586

Zhuohan Li's avatar
Zhuohan Li committed
1587
        # Add requests to the engine.
1588
1589
        it = prompts
        if use_tqdm:
1590
1591
            tqdm_func = use_tqdm if callable(use_tqdm) else tqdm
            it = tqdm_func(it, desc="Adding requests")
1592

1593
        added_request_ids: list[str] = []
1594

1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
        try:
            for i, prompt in enumerate(it):
                if isinstance(prompt, dict):
                    self._validate_mm_data_and_uuids(
                        prompt.get("multi_modal_data"), prompt.get("multi_modal_uuids")
                    )
                request_id = self._add_request(
                    prompt,
                    params[i] if isinstance(params, Sequence) else params,
                    lora_request=lora_request[i]
                    if isinstance(lora_request, Sequence)
                    else lora_request,
                    priority=priority[i] if priority else 0,
                )
                added_request_ids.append(request_id)
        except Exception as e:
            if added_request_ids:
                self.llm_engine.abort_request(added_request_ids)
            raise e
1614

1615
    def _validate_mm_data_and_uuids(
1616
        self,
1617
1618
        multi_modal_data: Any | None,  # MultiModalDataDict
        multi_modal_uuids: Any | None,  # MultiModalUUIDDict
1619
1620
1621
    ):
        """
        Validate that if any multi-modal data is skipped (i.e. None),
1622
        then its corresponding UUID must be set.
1623
1624
1625
1626
1627
1628
1629
1630
        """
        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:
1631
1632
1633
1634
1635
1636
1637
1638
                        if (
                            multi_modal_uuids is None
                            or modality not in multi_modal_uuids
                            or multi_modal_uuids[  # noqa: E501
                                modality
                            ]
                            is None
                        ):
1639
1640
                            raise ValueError(
                                f"Multi-modal data for {modality} is None "
1641
1642
                                f"but UUID is not provided"
                            )
1643
                        else:
1644
1645
1646
1647
                            if (
                                len(multi_modal_uuids[modality]) <= i
                                or multi_modal_uuids[modality][i] is None
                            ):
1648
1649
                                raise ValueError(
                                    f"Multi-modal data for {modality} is None "
1650
1651
                                    f"but UUID is not provided"
                                )
1652
            else:
1653
1654
1655
1656
1657
1658
1659
1660
1661
                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"
                    )
1662

1663
1664
1665
1666
    def _process_inputs(
        self,
        request_id: str,
        engine_prompt: PromptType,
1667
        params: SamplingParams | PoolingParams,
1668
        *,
1669
        lora_request: LoRARequest | None,
1670
1671
1672
1673
        priority: int,
    ) -> tuple[EngineCoreRequest, dict[str, Any]]:
        """Use the Processor to process inputs for LLMEngine."""
        tokenization_kwargs: dict[str, Any] = {}
1674
1675
1676
1677
1678
        _validate_truncation_size(
            self.model_config.max_model_len,
            params.truncate_prompt_tokens,
            tokenization_kwargs,
        )
1679

1680
        engine_request = self.processor.process_inputs(
1681
1682
1683
1684
1685
1686
1687
1688
1689
            request_id,
            engine_prompt,
            params,
            lora_request=lora_request,
            tokenization_kwargs=tokenization_kwargs,
            priority=priority,
        )
        return engine_request, tokenization_kwargs

1690
    def _add_request(
nunjunj's avatar
nunjunj committed
1691
        self,
1692
        prompt: PromptType,
1693
1694
        params: SamplingParams | PoolingParams,
        lora_request: LoRARequest | None = None,
1695
        priority: int = 0,
1696
    ) -> str:
1697
        prompt_text, _, _ = get_prompt_components(prompt)
1698
        request_id = str(next(self.request_counter))
1699
1700

        engine_request, tokenization_kwargs = self._process_inputs(
1701
            request_id,
1702
            prompt,
1703
1704
            params,
            lora_request=lora_request,
1705
1706
1707
1708
1709
1710
1711
1712
            priority=priority,
        )

        self.llm_engine.add_request(
            request_id,
            engine_request,
            params,
            lora_request=lora_request,
1713
            tokenization_kwargs=tokenization_kwargs,
1714
            priority=priority,
1715
            prompt_text=prompt_text,
nunjunj's avatar
nunjunj committed
1716
        )
1717
        return request_id
1718

1719
    def _run_engine(
1720
1721
        self, *, use_tqdm: bool | Callable[..., tqdm] = True
    ) -> list[RequestOutput | PoolingRequestOutput]:
1722
1723
        # Initialize tqdm.
        if use_tqdm:
Zhuohan Li's avatar
Zhuohan Li committed
1724
            num_requests = self.llm_engine.get_num_unfinished_requests()
1725
1726
            tqdm_func = use_tqdm if callable(use_tqdm) else tqdm
            pbar = tqdm_func(
1727
1728
1729
                total=num_requests,
                desc="Processed prompts",
                dynamic_ncols=True,
1730
                postfix=(f"est. speed input: {0:.2f} toks/s, output: {0:.2f} toks/s"),
1731
            )
1732

Zhuohan Li's avatar
Zhuohan Li committed
1733
        # Run the engine.
1734
        outputs: list[RequestOutput | PoolingRequestOutput] = []
1735
1736
        total_in_toks = 0
        total_out_toks = 0
Zhuohan Li's avatar
Zhuohan Li committed
1737
1738
        while self.llm_engine.has_unfinished_requests():
            step_outputs = self.llm_engine.step()
1739
            for output in step_outputs:
1740
                if output.finished:
1741
1742
                    outputs.append(output)
                    if use_tqdm:
1743
1744
                        if isinstance(output, RequestOutput):
                            # Calculate tokens only for RequestOutput
1745
                            n = len(output.outputs)
1746
                            assert output.prompt_token_ids is not None
1747
                            total_in_toks += len(output.prompt_token_ids) * n
1748
1749
                            in_spd = total_in_toks / pbar.format_dict["elapsed"]
                            total_out_toks += sum(
1750
1751
1752
                                len(stp.token_ids) for stp in output.outputs
                            )
                            out_spd = total_out_toks / pbar.format_dict["elapsed"]
1753
1754
                            pbar.postfix = (
                                f"est. speed input: {in_spd:.2f} toks/s, "
1755
1756
                                f"output: {out_spd:.2f} toks/s"
                            )
1757
                            pbar.update(n)
1758
1759
                        else:
                            pbar.update(1)
1760
1761
                        if pbar.n == num_requests:
                            pbar.refresh()
1762

1763
1764
        if use_tqdm:
            pbar.close()
1765
1766
1767
        # Sort the outputs by request ID.
        # This is necessary because some requests may be finished earlier than
        # its previous requests.
1768
        return sorted(outputs, key=lambda x: int(x.request_id))