llm.py 71.9 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import itertools
5
from collections.abc import Sequence
6
from typing import TYPE_CHECKING, Any, Callable, Optional, Union, 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
13

14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
from vllm.beam_search import (
    BeamSearchInstance,
    BeamSearchOutput,
    BeamSearchSequence,
    create_sort_beams_key_function,
)
from vllm.config import (
    CompilationConfig,
    ModelDType,
    StructuredOutputsConfig,
    TokenizerMode,
    is_init_field,
)
from vllm.engine.arg_utils import (
    ConvertOption,
    EngineArgs,
    HfOverrides,
    PoolerConfig,
    RunnerOption,
)
from vllm.entrypoints.chat_utils import (
    ChatCompletionMessageParam,
    ChatTemplateContentFormatOption,
    apply_hf_chat_template,
    apply_mistral_chat_template,
    parse_chat_messages,
    resolve_chat_template_content_format,
)

43
44
# yapf conflicts with isort for this block
# yapf: disable
45
46
47
48
49
50
51
52
53
from vllm.entrypoints.score_utils import (
    ScoreContentPartParam,
    ScoreMultiModalParam,
    _cosine_similarity,
    _validate_score_input_lens,
    compress_token_type_ids,
    get_score_prompt,
)

54
# yapf: enable
55
56
57
58
59
60
61
62
from vllm.entrypoints.utils import _validate_truncation_size, log_non_default_args
from vllm.inputs import (
    DataPrompt,
    PromptType,
    SingletonPrompt,
    TextPrompt,
    TokensPrompt,
)
63
from vllm.inputs.parse import get_prompt_components
64
from vllm.logger import init_logger
65
from vllm.lora.request import LoRARequest
66
from vllm.model_executor.layers.quantization import QuantizationMethods
67
68
69
70
71
72
73
from vllm.outputs import (
    ClassificationRequestOutput,
    EmbeddingRequestOutput,
    PoolingRequestOutput,
    RequestOutput,
    ScoringRequestOutput,
)
74
from vllm.plugins.io_processors import get_io_processor
75
from vllm.pooling_params import PoolingParams
76
from vllm.sampling_params import BeamSearchParams, RequestOutputKind, SamplingParams
77
from vllm.tasks import PoolingTask
78
79
80
81
82
83
from vllm.transformers_utils.tokenizer import (
    AnyTokenizer,
    MistralTokenizer,
    get_cached_tokenizer,
    init_tokenizer_from_configs,
)
yhu422's avatar
yhu422 committed
84
from vllm.usage.usage_lib import UsageContext
85
from vllm.utils import Counter, Device, as_iter, is_list_of
86
from vllm.v1.engine import EngineCoreRequest
87
from vllm.v1.engine.llm_engine import LLMEngine
88
from vllm.v1.engine.processor import Processor
89
from vllm.v1.sample.logits_processor import LogitsProcessor
90

91
92
93
if TYPE_CHECKING:
    from vllm.v1.metrics.reader import Metric

94
95
logger = init_logger(__name__)

96
97
_R = TypeVar("_R", default=Any)

98
99

class LLM:
Woosuk Kwon's avatar
Woosuk Kwon committed
100
101
102
103
104
105
106
107
108
109
    """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.
110
        tokenizer: The name or path of a HuggingFace Transformers tokenizer.
111
112
        tokenizer_mode: The tokenizer mode. "auto" will use the fast tokenizer
            if available, and "slow" will always use the slow tokenizer.
113
114
115
        skip_tokenizer_init: If true, skip initialization of tokenizer and
            detokenizer. Expect valid prompt_token_ids and None for prompt
            from the input.
116
117
        trust_remote_code: Trust remote code (e.g., from HuggingFace) when
            downloading the model and tokenizer.
118
119
120
121
        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.
122
        allowed_media_domains: If set, only media URLs that belong to this
123
            domain can be used for multi-modal inputs.
Woosuk Kwon's avatar
Woosuk Kwon committed
124
125
126
        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
127
128
129
130
            we support `float32`, `float16`, and `bfloat16`. If `auto`, we use
            the `torch_dtype` attribute specified in the model config file.
            However, if the `torch_dtype` in the config is `float32`, we will
            use `float16` instead.
131
        quantization: The method used to quantize the model weights. Currently,
132
            we support "awq", "gptq", and "fp8" (experimental).
133
134
135
136
            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
137
138
        revision: The specific model version to use. It can be a branch name,
            a tag name, or a commit id.
139
140
        tokenizer_revision: The specific tokenizer version to use. It can be a
            branch name, a tag name, or a commit id.
141
142
143
144
145
146
        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.
147
148
149
150
151
152
153
154
        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
            compared with using gpu_memory_memory_utilization. Note that
            kv_cache_memory_bytes (when not-None) ignores
            gpu_memory_utilization
155
        swap_space: The size (GiB) of CPU memory per GPU to use as swap space.
156
157
158
159
160
            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.
161
162
163
164
        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.
165
166
167
        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.
168
169
        disable_custom_all_reduce: See
            [ParallelConfig][vllm.config.ParallelConfig].
170
        hf_token: The token to use as HTTP bearer authorization for remote files
171
            . If `True`, will use the token generated when running
172
            `huggingface-cli login` (stored in `~/.huggingface`).
173
174
175
        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.
176
177
178
179
180
        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}`.
181
182
183
184
185
        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.
186
187
188
        compilation_config: Either an integer or a dictionary. If it is an
            integer, it is used as the level of compilation optimization. If it
            is a dictionary, it can specify the full compilation configuration.
189
        **kwargs: Arguments for [`EngineArgs`][vllm.EngineArgs].
nunjunj's avatar
nunjunj committed
190

191
192
    Note:
        This class is intended to be used for offline inference. For online
193
        serving, use the [AsyncLLMEngine][vllm.AsyncLLMEngine] class instead.
194
    """
195
196
197
198

    def __init__(
        self,
        model: str,
199
        *,
200
201
        runner: RunnerOption = "auto",
        convert: ConvertOption = "auto",
202
        tokenizer: Optional[str] = None,
203
        tokenizer_mode: TokenizerMode = "auto",
204
        skip_tokenizer_init: bool = False,
205
        trust_remote_code: bool = False,
206
        allowed_local_media_path: str = "",
207
        allowed_media_domains: Optional[list[str]] = None,
208
        tensor_parallel_size: int = 1,
209
210
        dtype: ModelDType = "auto",
        quantization: Optional[QuantizationMethods] = None,
211
        revision: Optional[str] = None,
212
        tokenizer_revision: Optional[str] = None,
213
        seed: Optional[int] = None,
214
        gpu_memory_utilization: float = 0.9,
215
        swap_space: float = 4,
216
        cpu_offload_gb: float = 0,
217
        enforce_eager: bool = False,
218
        disable_custom_all_reduce: bool = False,
219
        hf_token: Optional[Union[bool, str]] = None,
220
        hf_overrides: Optional[HfOverrides] = None,
221
        mm_processor_kwargs: Optional[dict[str, Any]] = None,
222
        pooler_config: Optional[PoolerConfig] = None,
223
        override_pooler_config: Optional[PoolerConfig] = None,
224
225
226
        structured_outputs_config: Optional[
            Union[dict[str, Any], StructuredOutputsConfig]
        ] = None,
227
        kv_cache_memory_bytes: Optional[int] = None,
228
229
230
231
        compilation_config: Optional[
            Union[int, dict[str, Any], CompilationConfig]
        ] = None,
        logits_processors: Optional[list[Union[str, type[LogitsProcessor]]]] = None,
232
        **kwargs: Any,
233
    ) -> None:
234
        """LLM constructor."""
235

236
237
        if "disable_log_stats" not in kwargs:
            kwargs["disable_log_stats"] = True
238

239
240
241
242
243
244
245
        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)

246
        if "kv_transfer_config" in kwargs and isinstance(
247
248
            kwargs["kv_transfer_config"], dict
        ):
249
            from vllm.config.kv_transfer import KVTransferConfig
250

251
252
            raw_config_dict = kwargs["kv_transfer_config"]
            try:
253
                kwargs["kv_transfer_config"] = KVTransferConfig(**raw_config_dict)
254
255
256
257
            except ValidationError as e:
                logger.error(
                    "Failed to convert 'kv_transfer_config' dict to "
                    "KVTransferConfig object. Dict: %s. Error: %s",
258
259
260
                    raw_config_dict,
                    e,
                )
261
262
                # Consider re-raising a more specific vLLM error or ValueError
                # to provide better context to the user.
263
                raise ValueError(f"Invalid 'kv_transfer_config' provided: {e}") from e
264

265
266
267
        if hf_overrides is None:
            hf_overrides = {}

268
        if compilation_config is not None:
269
270
            if isinstance(compilation_config, int):
                compilation_config_instance = CompilationConfig(
271
272
                    level=compilation_config
                )
273
274
            elif isinstance(compilation_config, dict):
                compilation_config_instance = CompilationConfig(
275
276
277
278
                    **{
                        k: v
                        for k, v in compilation_config.items()
                        if is_init_field(CompilationConfig, k)
279
280
                    }
                )
281
282
            else:
                compilation_config_instance = compilation_config
283
        else:
284
            compilation_config_instance = CompilationConfig()
285

286
287
288
289
290
291
292
        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)
293
294
                    }
                )
295
296
297
298
299
            else:
                structured_outputs_instance = structured_outputs_config
        else:
            structured_outputs_instance = StructuredOutputsConfig()

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: Union[dict[str, Any], None] = None
343

344
        supported_tasks = self.llm_engine.get_supported_tasks()  # type: ignore
345
346
347
348
349

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

        self.supported_tasks = supported_tasks

350
351
        # Load the Input/Output processor plugin if any
        io_processor_plugin = self.llm_engine.model_config.io_processor_plugin
352
353
354
        self.io_processor = get_io_processor(
            self.llm_engine.vllm_config, io_processor_plugin
        )
355

356
357
358
359
    @property
    def model_config(self):
        return self.llm_engine.model_config

360
361
    def get_tokenizer(self) -> AnyTokenizer:
        return self.llm_engine.get_tokenizer()
362
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
376
377
378
379
380
381
    def _get_processor(self) -> Processor:
        if not hasattr(self, "_processor"):
            vllm_config = self.llm_engine.vllm_config
            if self.model_config.skip_tokenizer_init:
                tokenizer = None
            else:
                tokenizer = init_tokenizer_from_configs(self.model_config)
            self._processor = Processor(vllm_config, tokenizer)
        return self._processor

382
    def get_default_sampling_params(self) -> SamplingParams:
383
384
        if self.default_sampling_params is None:
            self.default_sampling_params = (
385
386
                self.llm_engine.model_config.get_diff_sampling_param()
            )
387
388
        if self.default_sampling_params:
            return SamplingParams.from_optional(**self.default_sampling_params)
389
390
        return SamplingParams()

391
392
393
    def generate(
        self,
        prompts: Union[PromptType, Sequence[PromptType]],
394
395
396
        sampling_params: Optional[
            Union[SamplingParams, Sequence[SamplingParams]]
        ] = None,
397
        *,
398
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
399
400
401
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
        priority: Optional[list[int]] = None,
    ) -> list[RequestOutput]:
Woosuk Kwon's avatar
Woosuk Kwon committed
402
403
        """Generates the completions for the input prompts.

404
        This class automatically batches the given prompts, considering
Woosuk Kwon's avatar
Woosuk Kwon committed
405
406
407
408
        the memory constraint. For the best performance, put all of your prompts
        into a single list and pass it to this method.

        Args:
409
            prompts: The prompts to the LLM. You may pass a sequence of prompts
410
                for batch inference. See [PromptType][vllm.inputs.PromptType]
411
                for more details about the format of each prompt.
Woosuk Kwon's avatar
Woosuk Kwon committed
412
            sampling_params: The sampling parameters for text generation. If
nunjunj's avatar
nunjunj committed
413
414
415
                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
416
                prompts and it is paired one by one with the prompt.
417
418
419
420
            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.
421
            lora_request: LoRA request to use for generation, if any.
422
423
            priority: The priority of the requests, if any.
                Only applicable when priority scheduling policy is enabled.
Woosuk Kwon's avatar
Woosuk Kwon committed
424
425

        Returns:
426
            A list of `RequestOutput` objects containing the
427
            generated completions in the same order as the input prompts.
428

429
430
431
432
        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.
433
        """
434
435
436
        model_config = self.llm_engine.model_config
        runner_type = model_config.runner_type
        if runner_type != "generate":
437
438
439
            raise ValueError(
                "LLM.generate() is only supported for generative models. "
                "Try passing `--runner generate` to use the model as a "
440
441
                "generative model."
            )
442

443
444
        if sampling_params is None:
            # Use default sampling params.
445
            sampling_params = self.get_default_sampling_params()
446

447
        # Add any modality specific loras to the corresponding prompts
448
        lora_request = self._get_modality_specific_lora_reqs(prompts, lora_request)
449

450
        self._validate_and_add_requests(
451
            prompts=prompts,
452
            params=sampling_params,
453
            use_tqdm=use_tqdm,
454
            lora_request=lora_request,
455
456
            priority=priority,
        )
457

458
        outputs = self._run_engine(use_tqdm=use_tqdm)
Joe Runde's avatar
Joe Runde committed
459
        return self.engine_class.validate_outputs(outputs, RequestOutput)
460

461
    def _get_modality_specific_lora_reqs(
462
463
464
465
        self,
        prompts: Union[PromptType, Sequence[PromptType]],
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]],
    ):
466
467
468
469
470
471
        # 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.
472
473
474
475
476
        if (
            lora_config is None
            or not self.llm_engine.model_config.is_multimodal_model
            or (lora_config and lora_config.default_mm_loras is None)
        ):
477
478
            return lora_request

479
480
        if not isinstance(prompts, Sequence):
            prompts = [prompts]
481

482
483
484
485
486
        optional_loras = (
            [lora_request] * len(prompts)
            if not isinstance(lora_request, Sequence)
            else lora_request
        )
487
488
489

        return [
            self._resolve_single_prompt_mm_lora(
490
                prompt,
491
492
                opt_lora_req,
                lora_config.default_mm_loras,
493
494
            )
            for prompt, opt_lora_req in zip(prompts, optional_loras)
495
496
        ]

497
498
499
500
501
502
503
504
505
506
507
    def _resolve_single_prompt_mm_lora(
        self,
        prompt: PromptType,
        lora_request: Optional[LoRARequest],
        default_mm_loras: Optional[dict[str, str]],
    ):
        if (
            not default_mm_loras
            or not isinstance(prompt, dict)
            or "multi_modal_data" not in prompt
        ):
508
509
            return lora_request

510
        prompt = cast(Union[TextPrompt, TokensPrompt], prompt)
511

512
513
514
        intersection = set(prompt["multi_modal_data"].keys()).intersection(
            default_mm_loras.keys()
        )
515
516
517
518
519
520
521
522
523
        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"
524
525
526
                " will be skipped",
                intersection,
            )
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
            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 "
542
543
                    "lora_request as we only apply one LoRARequest per prompt"
                )
544
545
546
547
548
549
550
551
            return lora_request

        return LoRARequest(
            modality_name,
            modality_lora_id,
            modality_lora_path,
        )

552
553
554
555
556
557
558
    def collective_rpc(
        self,
        method: Union[str, Callable[..., _R]],
        timeout: Optional[float] = None,
        args: tuple = (),
        kwargs: Optional[dict[str, Any]] = None,
    ) -> list[_R]:
559
560
561
562
563
564
565
566
567
568
569
        """
        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
570
                [`TimeoutError`][] on timeout. `None` means wait indefinitely.
571
572
573
574
575
            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.
576

577
578
579
        Note:
            It is recommended to use this API to only pass control messages,
            and set up data-plane communication to pass data.
580
        """
581
582

        return self.llm_engine.collective_rpc(method, timeout, args, kwargs)
583
584

    def apply_model(self, func: Callable[[nn.Module], _R]) -> list[_R]:
585
        """
586
587
        Run a function directly on the model inside each worker,
        returning the result for each of them.
588
589
590
591
592
593

        !!! 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!
594
        """
595
        return self.llm_engine.apply_model(func)
596

597
598
599
600
601
602
    def _get_beam_search_lora_requests(
        self,
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]],
        prompts: list[Union[TokensPrompt, TextPrompt]],
    ) -> list[Optional[LoRARequest]]:
        """Get the optional lora request corresponding to each prompt."""
603
        if isinstance(lora_request, Sequence) and len(lora_request) != len(prompts):
604
            raise ValueError(
605
606
                "Lora request list should be the same length as the prompts"
            )
607
608
609
610
611
612

        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)}")

613
614
    def beam_search(
        self,
615
        prompts: list[Union[TokensPrompt, TextPrompt]],
616
        params: BeamSearchParams,
617
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
618
        use_tqdm: bool = False,
619
        concurrency_limit: Optional[int] = None,
620
    ) -> list[BeamSearchOutput]:
621
622
623
624
625
626
        """
        Generate sequences using beam search.

        Args:
            prompts: A list of prompts. Each prompt can be a string or a list
                of token IDs.
627
            params: The beam search parameters.
628
            lora_request: LoRA request to use for generation, if any.
629
            use_tqdm: Whether to use tqdm to display the progress bar.
630
631
            concurrency_limit: The maximum number of concurrent requests.
                If None, the number of concurrent requests is unlimited.
632
        """
633
634
        # TODO: how does beam search work together with length penalty,
        # frequency, penalty, and stopping criteria, etc.?
635
636
637
638
        beam_width = params.beam_width
        max_tokens = params.max_tokens
        temperature = params.temperature
        ignore_eos = params.ignore_eos
639
640
        length_penalty = params.length_penalty

641
        lora_requests = self._get_beam_search_lora_requests(lora_request, prompts)
642

643
644
645
646
647
        tokenizer = self.get_tokenizer()
        sort_beams_key = create_sort_beams_key_function(
            tokenizer.eos_token_id,
            length_penalty,
        )
648

649
650
651
        if use_tqdm and concurrency_limit is not None:
            logger.warning(
                "Progress bar is not supported when using concurrency_limit. "
652
653
                "Disabling progress bar."
            )
654
655
656
657
658
            use_tqdm = False

        if concurrency_limit is None:
            concurrency_limit = len(prompts)

659
660
        def create_tokens_prompt_from_beam(beam: BeamSearchSequence) -> TokensPrompt:
            token_prompt_kwargs: TokensPrompt = {"prompt_token_ids": beam.tokens}
661
662
663
664
            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:
665
                token_prompt_kwargs["mm_processor_kwargs"] = beam.mm_processor_kwargs
666
            return TokensPrompt(**token_prompt_kwargs)
667

668
669
670
        # 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
671
672
673
        beam_search_params = SamplingParams(
            logprobs=2 * beam_width, max_tokens=1, temperature=temperature
        )
674
        instances: list[BeamSearchInstance] = []
675

676
        for lora_req, prompt in zip(lora_requests, prompts):
677
678
679
680
681
            # 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:
682
                mm_kwargs["mm_processor_kwargs"] = prompt["mm_processor_kwargs"]
683

684
685
            if "prompt_token_ids" in prompt:
                prompt = cast(TokensPrompt, prompt)  # Needed for mypy
686
687
688
                prompt_tokens = prompt["prompt_token_ids"]
            else:
                prompt_tokens = tokenizer.encode(prompt["prompt"])
689

690
            instances.append(
691
692
693
694
695
                BeamSearchInstance(
                    prompt_tokens,
                    lora_request=lora_req,
                    logprobs=None,
                    **mm_kwargs,
696
697
                ),
            )
698

699
        for prompt_start in range(0, len(prompts), concurrency_limit):
700
            instances_batch = instances[prompt_start : prompt_start + concurrency_limit]
701
702
703

            token_iter = range(max_tokens)
            if use_tqdm:
704
705
706
                token_iter = tqdm(
                    token_iter, desc="Beam search", unit="token", unit_scale=False
                )
707
708
709
                logger.warning(
                    "The progress bar shows the upper bound on token steps and "
                    "may finish early due to stopping conditions. It does not "
710
711
                    "reflect instance-level progress."
                )
712
713
            for _ in token_iter:
                all_beams: list[BeamSearchSequence] = list(
714
715
                    sum((instance.beams for instance in instances_batch), [])
                )
716
717
                pos = [0] + list(
                    itertools.accumulate(
718
719
720
                        len(instance.beams) for instance in instances_batch
                    )
                )
721
                instance_start_and_end: list[tuple[int, int]] = list(
722
723
                    zip(pos[:-1], pos[1:])
                )
724
725
726
727
728
729

                if len(all_beams) == 0:
                    break

                # create corresponding batch entries for prompt & optional lora
                prompts_batch, lora_req_batch = zip(
730
731
732
733
734
                    *[
                        (create_tokens_prompt_from_beam(beam), beam.lora_request)
                        for beam in all_beams
                    ]
                )
735
736
737

                # only runs for one step
                # we don't need to use tqdm here
738
739
740
741
742
743
                output = self.generate(
                    prompts_batch,
                    sampling_params=beam_search_params,
                    use_tqdm=False,
                    lora_request=lora_req_batch,
                )
744

745
746
747
                for (start, end), instance in zip(
                    instance_start_and_end, instances_batch
                ):
748
749
750
751
752
753
754
755
756
757
758
759
760
761
                    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],
762
                                    logprobs=current_beam.logprobs + [logprobs],
763
                                    lora_request=current_beam.lora_request,
764
765
766
767
768
769
770
771
772
773
                                    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
                                ):
774
775
776
                                    instance.completed.append(new_beam)
                                else:
                                    instance_new_beams.append(new_beam)
777
778
779
                    sorted_beams = sorted(
                        instance_new_beams, key=sort_beams_key, reverse=True
                    )
780
                    instance.beams = sorted_beams[:beam_width]
781
782
783
784

        outputs = []
        for instance in instances:
            instance.completed.extend(instance.beams)
785
786
787
            sorted_completed = sorted(
                instance.completed, key=sort_beams_key, reverse=True
            )
788
789
790
791
792
793
794
795
            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

796
    def preprocess_chat(
nunjunj's avatar
nunjunj committed
797
        self,
798
799
800
        messages: Union[
            list[ChatCompletionMessageParam], list[list[ChatCompletionMessageParam]]
        ],
nunjunj's avatar
nunjunj committed
801
        chat_template: Optional[str] = None,
802
        chat_template_content_format: ChatTemplateContentFormatOption = "auto",
803
        add_generation_prompt: bool = True,
804
        continue_final_message: bool = False,
805
        tools: Optional[list[dict[str, Any]]] = None,
806
        chat_template_kwargs: Optional[dict[str, Any]] = None,
807
        mm_processor_kwargs: Optional[dict[str, Any]] = None,
808
    ) -> list[TokensPrompt]:
nunjunj's avatar
nunjunj committed
809
        """
810
811
        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
812

813
        Refer to `chat` for a complete description of the arguments.
nunjunj's avatar
nunjunj committed
814
        Returns:
815
816
817
            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
818
        """
819
        list_of_messages: list[list[ChatCompletionMessageParam]]
nunjunj's avatar
nunjunj committed
820

821
822
        # Handle multi and single conversations
        if is_list_of(messages, list):
823
            # messages is list[list[...]]
824
            list_of_messages = cast(list[list[ChatCompletionMessageParam]], messages)
825
        else:
826
            # messages is list[...]
827
            list_of_messages = [cast(list[ChatCompletionMessageParam], messages)]
828

829
        tokenizer = self.get_tokenizer()
830
831
832
        model_config = self.llm_engine.get_model_config()
        resolved_content_format = resolve_chat_template_content_format(
            chat_template,
833
            tools,
834
835
            chat_template_content_format,
            tokenizer,
836
            model_config=model_config,
837
838
        )

839
840
841
842
843
844
845
846
        _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 {})

847
        prompts: list[TokensPrompt] = []
848
849

        for msgs in list_of_messages:
850
851
852
            # 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.
853
            conversation, mm_data, mm_uuids = parse_chat_messages(
854
855
856
857
858
                msgs,
                model_config,
                tokenizer,
                content_format=resolved_content_format,
            )
859
860

            if isinstance(tokenizer, MistralTokenizer):
861
                prompt_token_ids = apply_mistral_chat_template(
862
863
                    tokenizer,
                    messages=msgs,
864
                    **_chat_template_kwargs,
865
866
                )
            else:
867
                prompt_str = apply_hf_chat_template(
868
                    tokenizer=tokenizer,
869
                    conversation=conversation,
870
                    model_config=model_config,
871
                    **_chat_template_kwargs,
872
                )
873
874
                # Special tokens are already included in chat templates so
                # should not be added by the tokenizer in this case.
875
876
877
                prompt_token_ids = tokenizer.encode(
                    prompt_str, add_special_tokens=False
                )
878

879
            prompt = TokensPrompt(prompt_token_ids=prompt_token_ids)
880
881
882
883

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

884
885
886
            if mm_uuids is not None:
                prompt["multi_modal_uuids"] = mm_uuids

887
888
889
            if mm_processor_kwargs is not None:
                prompt["mm_processor_kwargs"] = mm_processor_kwargs

890
            prompts.append(prompt)
891

892
893
894
895
        return prompts

    def chat(
        self,
896
897
898
899
        messages: Union[
            list[ChatCompletionMessageParam], list[list[ChatCompletionMessageParam]]
        ],
        sampling_params: Optional[Union[SamplingParams, list[SamplingParams]]] = None,
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
958
959
960
961
962
963
964
965
966
967
968
969
970
971
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
        lora_request: Optional[LoRARequest] = None,
        chat_template: Optional[str] = None,
        chat_template_content_format: ChatTemplateContentFormatOption = "auto",
        add_generation_prompt: bool = True,
        continue_final_message: bool = False,
        tools: Optional[list[dict[str, Any]]] = None,
        chat_template_kwargs: Optional[dict[str, Any]] = None,
        mm_processor_kwargs: Optional[dict[str, Any]] = None,
    ) -> 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
972
        return self.generate(
973
            prompts,
974
            sampling_params=sampling_params,
nunjunj's avatar
nunjunj committed
975
976
977
978
            use_tqdm=use_tqdm,
            lora_request=lora_request,
        )

979
980
    def encode(
        self,
981
        prompts: Union[PromptType, Sequence[PromptType], DataPrompt],
982
        pooling_params: Optional[Union[PoolingParams, Sequence[PoolingParams]]] = None,
983
        *,
984
        truncate_prompt_tokens: Optional[int] = None,
985
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
986
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
987
        pooling_task: PoolingTask = "encode",
988
        tokenization_kwargs: Optional[dict[str, Any]] = None,
989
    ) -> list[PoolingRequestOutput]:
990
991
        """Apply pooling to the hidden states corresponding to the input
        prompts.
992

993
        This class automatically batches the given prompts, considering
994
995
996
997
        the memory constraint. For the best performance, put all of your prompts
        into a single list and pass it to this method.

        Args:
998
            prompts: The prompts to the LLM. You may pass a sequence of prompts
999
                for batch inference. See [PromptType][vllm.inputs.PromptType]
1000
                for more details about the format of each prompt.
1001
1002
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1003
1004
1005
1006
            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.
1007
            lora_request: LoRA request to use for generation, if any.
1008
            pooling_task: Override the pooling task to use.
1009
1010
            tokenization_kwargs: overrides tokenization_kwargs set in
                pooling_params
1011
1012

        Returns:
1013
            A list of `PoolingRequestOutput` objects containing the
1014
            pooled hidden states in the same order as the input prompts.
1015

1016
1017
1018
1019
        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.
1020
        """
1021
1022
1023
1024

        if self.supported_tasks == ["encode"] and pooling_task is None:
            pooling_task = "encode"

1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
        if pooling_task is None:
            if "embed" in self.supported_tasks:
                pooling_task = "embed"
            else:
                pooling_task = "encode"

            logger.warning_once(
                "`LLM.encode` is currently using `pooling_task = %s`.\n"
                "Please use one of the more specific methods or set the "
                "task directly when using `LLM.encode`:\n"
                "  - For embeddings, use `LLM.embed(...)` "
1036
                'or `pooling_task="embed"`.\n'
1037
                "  - For classification logits, use `LLM.classify(...)` "
1038
                'or `pooling_task="classify"`.\n'
1039
                "  - For rewards, use `LLM.reward(...)` "
1040
                'or `pooling_task="reward"`\n'
1041
                "  - For similarity scores, use `LLM.score(...)`.",
1042
1043
                pooling_task,
            )
1044

1045
1046
        model_config = self.llm_engine.model_config
        runner_type = model_config.runner_type
1047
        if runner_type != "pooling":
1048
1049
1050
            raise ValueError(
                "LLM.encode() is only supported for pooling models. "
                "Try passing `--runner pooling` to use the model as a "
1051
1052
                "pooling model."
            )
1053

1054
        if pooling_task not in self.supported_tasks:
1055
            raise ValueError(f"pooling_task must be one of {self.supported_tasks}.")
1056

1057
1058
1059
        if pooling_params is None:
            # Use default pooling params.
            pooling_params = PoolingParams()
1060

1061
1062
1063
1064
1065
        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
1066

1067
1068
1069
1070
1071
1072
1073
1074
        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' "
1075
1076
                    "offline inference example for more details."
                )
1077
1078
1079
1080
1081
1082
1083

            # 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)

1084
        self._validate_and_add_requests(
1085
            prompts=prompts,
1086
            params=pooling_params,
1087
            use_tqdm=use_tqdm,
1088
            lora_request=lora_request,
1089
1090
        )

1091
        outputs = self._run_engine(use_tqdm=use_tqdm)
1092
1093

        model_outputs = self.engine_class.validate_outputs(
1094
1095
            outputs, PoolingRequestOutput
        )
1096
1097
1098
1099
1100

        if io_processor_prompt:
            # get the post-processed model outputs
            assert self.io_processor is not None
            processed_outputs = self.io_processor.post_process(
1101
1102
                model_output=model_outputs
            )
1103
1104

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

1115
1116
1117
1118
    def embed(
        self,
        prompts: Union[PromptType, Sequence[PromptType]],
        *,
1119
        truncate_prompt_tokens: Optional[int] = None,
1120
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
1121
        pooling_params: Optional[Union[PoolingParams, Sequence[PoolingParams]]] = None,
1122
1123
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
    ) -> 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
1166
1167

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

    def classify(
        self,
        prompts: Union[PromptType, Sequence[PromptType]],
        *,
1168
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
1169
        pooling_params: Optional[Union[PoolingParams, Sequence[PoolingParams]]] = None,
1170
1171
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
    ) -> 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
1212
1213
1214
1215
1216
    def reward(
        self,
        prompts: Union[PromptType, Sequence[PromptType]],
        /,
        *,
        truncate_prompt_tokens: Optional[int] = None,
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
1217
        pooling_params: Optional[Union[PoolingParams, Sequence[PoolingParams]]] = None,
1218
1219
1220
1221
1222
1223
1224
1225
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
    ) -> list[PoolingRequestOutput]:
        """
        Generate rewards for each prompt.

        Args:
            prompts: The prompts to the LLM. You may pass a sequence of prompts
                for batch inference. See [PromptType][vllm.inputs.PromptType]
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
1245
1246
1247
            use_tqdm: If `True`, shows a tqdm progress bar.
                If a callable (e.g., `functools.partial(tqdm, leave=False)`),
                it is used to create the progress bar.
                If `False`, no progress bar is created.
            lora_request: LoRA request to use for generation, if any.
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
        Returns:
            A list of `PoolingRequestOutput` objects containing the
            pooled hidden states in the same order as the input prompts.
        """

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

1248
1249
1250
    def _embedding_score(
        self,
        tokenizer: AnyTokenizer,
1251
1252
        text_1: list[Union[str, TextPrompt, TokensPrompt]],
        text_2: list[Union[str, TextPrompt, TokensPrompt]],
1253
        truncate_prompt_tokens: Optional[int] = None,
1254
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
1255
        pooling_params: Optional[PoolingParams] = None,
1256
1257
1258
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
    ) -> 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
        data_1: Union[list[str], list[ScoreContentPartParam]],
        data_2: Union[list[str], list[ScoreContentPartParam]],
1285
        truncate_prompt_tokens: Optional[int] = None,
1286
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
1287
        pooling_params: Optional[PoolingParams] = None,
1288
1289
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
    ) -> list[ScoringRequestOutput]:
1290
        model_config = self.llm_engine.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
1302
        if pooling_params is None:
            pooling_params = PoolingParams(task="score")

        model_config = self.llm_engine.model_config
        pooling_params.verify("score", model_config)
1303
        pooling_params_list = list[PoolingParams]()
1304

1305
        tokenization_kwargs: dict[str, Any] = {}
1306

1307
1308
1309
        _validate_truncation_size(
            model_config.max_model_len, truncate_prompt_tokens, tokenization_kwargs
        )
1310

1311
        prompts = list[PromptType]()
1312

1313
1314
        input_pairs = [(t1, t2) for t1, t2 in zip(data_1, data_2)]

1315
        model_config = self.llm_engine.model_config
1316

1317
1318
1319
1320
1321
1322
1323
1324
1325
        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,
            )

1326
            if token_type_ids := engine_prompt.pop("token_type_ids", None):
1327
1328
1329
1330
1331
1332
1333
                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)

1334
            prompts.append(engine_prompt)
1335
1336

        self._validate_and_add_requests(
1337
            prompts=prompts,
1338
            params=pooling_params_list,
1339
            use_tqdm=use_tqdm,
1340
1341
1342
1343
            lora_request=lora_request,
        )

        outputs = self._run_engine(use_tqdm=use_tqdm)
1344
        items = self.engine_class.validate_outputs(outputs, PoolingRequestOutput)
1345
1346
1347

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

1348
1349
    def score(
        self,
1350
1351
        data_1: Union[SingletonPrompt, Sequence[SingletonPrompt], ScoreMultiModalParam],
        data_2: Union[SingletonPrompt, Sequence[SingletonPrompt], ScoreMultiModalParam],
1352
        /,
1353
        *,
1354
        truncate_prompt_tokens: Optional[int] = None,
1355
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
1356
        pooling_params: Optional[PoolingParams] = None,
1357
1358
        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
    ) -> list[ScoringRequestOutput]:
1359
1360
        """Generate similarity scores for all pairs `<text,text_pair>` or
          `<multi-modal data, multi-modal data pair>`.
1361

1362
        The inputs can be `1 -> 1`, `1 -> N` or `N -> N`.
1363
1364
        In the `1 - N` case the `data_1` input will be replicated `N`
        times to pair with the `data_2` inputs.
1365
        The input pairs are used to build a list of prompts for the
1366
1367
        cross encoder model. This class automatically batches the prompts,
        considering the memory constraint. For the best performance, put all
1368
1369
1370
        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
1371
        appropriate multi-modal models. For multi-modal inputs, ensure the
1372
        prompt structure matches the model's expected input format.
1373
1374

        Args:
1375
1376
1377
            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
1378
                the `data_2` list.
1379
            data_2: The data to pair with the query to form the input to
1380
                the LLM. Can be text or multi-modal data. See [PromptType]
1381
                [vllm.inputs.PromptType] for more details about the format of
1382
                each prompt.
1383
1384
1385
1386
            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.
1387
            lora_request: LoRA request to use for generation, if any.
1388
1389
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1390
        Returns:
1391
            A list of `ScoringRequestOutput` objects containing the
1392
1393
            generated scores in the same order as the input prompts.
        """
1394
1395
        model_config = self.llm_engine.model_config
        runner_type = model_config.runner_type
1396
        if runner_type != "pooling":
1397
1398
1399
            raise ValueError(
                "LLM.score() is only supported for pooling models. "
                "Try passing `--runner pooling` to use the model as a "
1400
1401
                "pooling model."
            )
1402

1403
1404
        supported_tasks = self.supported_tasks
        if all(t not in supported_tasks for t in ("embed", "classify")):
1405
1406
1407
1408
1409
            raise ValueError(
                "Score API is not supported by this model. "
                "Try converting the model using "
                "`--convert embed` or `--convert classify`."
            )
1410

1411
1412
1413
1414
        if (
            model_config.is_cross_encoder
            and getattr(model_config.hf_config, "num_labels", 0) != 1
        ):
1415
            raise ValueError("Score API is only enabled for num_labels == 1.")
1416
1417
1418
1419

        # 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
1420
        tokenizer = self.get_tokenizer()
1421

1422
        if not model_config.is_multimodal_model:
1423

1424
1425
1426
1427
1428
            def check_data_type(
                data: Union[
                    SingletonPrompt, Sequence[SingletonPrompt], ScoreMultiModalParam
                ],
            ):
1429
                if isinstance(data, dict) and "content" in data:
1430
1431
1432
1433
                    raise ValueError(
                        "ScoreMultiModalParam is not supported "
                        f"for {model_config.architecture}"
                    )
1434
1435
1436
1437
1438
1439
1440

            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:
1441
1442
1443
                        raise ValueError(
                            "Multi-modal prompt is not supported for scoring"
                        )
1444
1445
                    elif "prompt_token_ids" in prompt:
                        prompt = tokenizer.decode(
1446
1447
                            cast(TokensPrompt, prompt)["prompt_token_ids"]
                        )
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
1473
1474
1475
                    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]
1476

1477
        if model_config.is_cross_encoder:
1478
1479
1480
1481
1482
1483
            return self._cross_encoding_score(
                tokenizer,
                data_1,  # type: ignore[arg-type]
                data_2,  # type: ignore[arg-type]
                truncate_prompt_tokens,
                use_tqdm,
1484
                pooling_params,
1485
1486
                lora_request,
            )
1487
        else:
1488
1489
            return self._embedding_score(
                tokenizer,
1490
1491
                data_1,  # type: ignore[arg-type]
                data_2,  # type: ignore[arg-type]
1492
1493
                truncate_prompt_tokens,
                use_tqdm,
1494
                pooling_params,
1495
1496
                lora_request,
            )
1497

1498
1499
1500
1501
1502
1503
    def start_profile(self) -> None:
        self.llm_engine.start_profile()

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

1504
1505
    def reset_prefix_cache(self, device: Optional[Device] = None) -> bool:
        return self.llm_engine.reset_prefix_cache(device)
1506

1507
1508
1509
1510
1511
1512
    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.

1513
        Args:
1514
1515
            level: The sleep level. Level 1 sleep will offload the model
                weights and discard the kv cache. The content of kv cache
1516
                is forgotten. Level 1 sleep is good for sleeping and waking
1517
1518
1519
1520
1521
                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
1522
                sleep is good for sleeping and waking up the engine to run a
1523
                different model or update the model, where previous model
1524
                weights are not needed. It reduces CPU memory pressure.
1525
        """
1526
        self.reset_prefix_cache()
1527
1528
        self.llm_engine.sleep(level=level)

1529
    def wake_up(self, tags: Optional[list[str]] = None):
1530
        """
1531
1532
        Wake up the engine from sleep mode. See the [sleep][vllm.LLM.sleep]
        method for more details.
1533

1534
        Args:
1535
1536
            tags: An optional list of tags to reallocate the engine memory
                for specific memory allocations. Values must be in
1537
                `("weights", "kv_cache")`. If None, all memory is reallocated.
1538
                wake_up should be called with all tags (or None) before the
1539
1540
1541
                engine is used again.
        """
        self.llm_engine.wake_up(tags)
1542

1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
    def get_metrics(self) -> list["Metric"]:
        """Return a snapshot of aggregated metrics from Prometheus.

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

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

1555
1556
    def _validate_and_add_requests(
        self,
1557
        prompts: Union[PromptType, Sequence[PromptType], DataPrompt],
1558
1559
1560
1561
1562
1563
        params: Union[
            SamplingParams,
            Sequence[SamplingParams],
            PoolingParams,
            Sequence[PoolingParams],
        ],
1564
        *,
1565
        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
1566
        lora_request: Optional[Union[Sequence[LoRARequest], LoRARequest]],
1567
        priority: Optional[list[int]] = None,
1568
    ) -> None:
1569
        if isinstance(prompts, (str, dict)):
1570
            # Convert a single prompt to a list.
1571
            prompts = [prompts]  # type: ignore[list-item]
1572

1573
        num_requests = len(prompts)
1574
        if isinstance(params, Sequence) and len(params) != num_requests:
1575
1576
1577
1578
1579
1580
1581
            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,):
1582
1583
1584
            if isinstance(sp, SamplingParams):
                # We only care about the final output
                sp.output_kind = RequestOutputKind.FINAL_ONLY
1585

Zhuohan Li's avatar
Zhuohan Li committed
1586
        # Add requests to the engine.
1587
1588
        it = prompts
        if use_tqdm:
1589
1590
            tqdm_func = use_tqdm if callable(use_tqdm) else tqdm
            it = tqdm_func(it, desc="Adding requests")
1591
1592

        for i, prompt in enumerate(it):
1593
1594
            if isinstance(prompt, dict):
                self._validate_mm_data_and_uuids(
1595
1596
                    prompt.get("multi_modal_data"), prompt.get("multi_modal_uuids")
                )
1597

1598
            self._add_request(
1599
                prompt,
1600
                params[i] if isinstance(params, Sequence) else params,
1601
1602
1603
                lora_request=lora_request[i]
                if isinstance(lora_request, Sequence)
                else lora_request,
1604
                priority=priority[i] if priority else 0,
nunjunj's avatar
nunjunj committed
1605
            )
1606

1607
    def _validate_mm_data_and_uuids(
1608
1609
1610
        self,
        multi_modal_data: Optional[Any],  # MultiModalDataDict
        multi_modal_uuids: Optional[Any],  # MultiModalUUIDDict
1611
1612
1613
    ):
        """
        Validate that if any multi-modal data is skipped (i.e. None),
1614
        then its corresponding UUID must be set.
1615
1616
1617
1618
1619
1620
1621
1622
        """
        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:
1623
1624
1625
1626
1627
1628
1629
1630
                        if (
                            multi_modal_uuids is None
                            or modality not in multi_modal_uuids
                            or multi_modal_uuids[  # noqa: E501
                                modality
                            ]
                            is None
                        ):
1631
1632
                            raise ValueError(
                                f"Multi-modal data for {modality} is None "
1633
1634
                                f"but UUID is not provided"
                            )
1635
                        else:
1636
1637
1638
1639
                            if (
                                len(multi_modal_uuids[modality]) <= i
                                or multi_modal_uuids[modality][i] is None
                            ):
1640
1641
                                raise ValueError(
                                    f"Multi-modal data for {modality} is None "
1642
1643
                                    f"but UUID is not provided"
                                )
1644
            else:
1645
1646
1647
1648
1649
1650
1651
1652
1653
                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"
                    )
1654

1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
    def _process_inputs(
        self,
        request_id: str,
        engine_prompt: PromptType,
        params: Union[SamplingParams, PoolingParams],
        *,
        lora_request: Optional[LoRARequest],
        priority: int,
    ) -> tuple[EngineCoreRequest, dict[str, Any]]:
        """Use the Processor to process inputs for LLMEngine."""
        tokenization_kwargs: dict[str, Any] = {}
1666
1667
1668
1669
1670
        _validate_truncation_size(
            self.model_config.max_model_len,
            params.truncate_prompt_tokens,
            tokenization_kwargs,
        )
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682

        processor = self._get_processor()
        engine_request = processor.process_inputs(
            request_id,
            engine_prompt,
            params,
            lora_request=lora_request,
            tokenization_kwargs=tokenization_kwargs,
            priority=priority,
        )
        return engine_request, tokenization_kwargs

1683
    def _add_request(
nunjunj's avatar
nunjunj committed
1684
        self,
1685
        prompt: PromptType,
nunjunj's avatar
nunjunj committed
1686
        params: Union[SamplingParams, PoolingParams],
1687
        lora_request: Optional[LoRARequest] = None,
1688
        priority: int = 0,
1689
    ) -> None:
1690
        prompt_text, _, _ = get_prompt_components(prompt)
1691
        request_id = str(next(self.request_counter))
1692
1693

        engine_request, tokenization_kwargs = self._process_inputs(
1694
            request_id,
1695
            prompt,
1696
1697
            params,
            lora_request=lora_request,
1698
1699
1700
1701
1702
1703
1704
1705
            priority=priority,
        )

        self.llm_engine.add_request(
            request_id,
            engine_request,
            params,
            lora_request=lora_request,
1706
            tokenization_kwargs=tokenization_kwargs,
1707
            priority=priority,
1708
            prompt_text=prompt_text,
nunjunj's avatar
nunjunj committed
1709
        )
1710

1711
    def _run_engine(
1712
        self, *, use_tqdm: Union[bool, Callable[..., tqdm]] = True
1713
    ) -> list[Union[RequestOutput, PoolingRequestOutput]]:
1714
1715
        # Initialize tqdm.
        if use_tqdm:
Zhuohan Li's avatar
Zhuohan Li committed
1716
            num_requests = self.llm_engine.get_num_unfinished_requests()
1717
1718
            tqdm_func = use_tqdm if callable(use_tqdm) else tqdm
            pbar = tqdm_func(
1719
1720
1721
                total=num_requests,
                desc="Processed prompts",
                dynamic_ncols=True,
1722
                postfix=(f"est. speed input: {0:.2f} toks/s, output: {0:.2f} toks/s"),
1723
            )
1724

Zhuohan Li's avatar
Zhuohan Li committed
1725
        # Run the engine.
1726
        outputs: list[Union[RequestOutput, PoolingRequestOutput]] = []
1727
1728
        total_in_toks = 0
        total_out_toks = 0
Zhuohan Li's avatar
Zhuohan Li committed
1729
1730
        while self.llm_engine.has_unfinished_requests():
            step_outputs = self.llm_engine.step()
1731
            for output in step_outputs:
1732
                if output.finished:
1733
1734
                    outputs.append(output)
                    if use_tqdm:
1735
1736
                        if isinstance(output, RequestOutput):
                            # Calculate tokens only for RequestOutput
1737
                            n = len(output.outputs)
1738
                            assert output.prompt_token_ids is not None
1739
                            total_in_toks += len(output.prompt_token_ids) * n
1740
1741
                            in_spd = total_in_toks / pbar.format_dict["elapsed"]
                            total_out_toks += sum(
1742
1743
1744
                                len(stp.token_ids) for stp in output.outputs
                            )
                            out_spd = total_out_toks / pbar.format_dict["elapsed"]
1745
1746
                            pbar.postfix = (
                                f"est. speed input: {in_spd:.2f} toks/s, "
1747
1748
                                f"output: {out_spd:.2f} toks/s"
                            )
1749
                            pbar.update(n)
1750
1751
                        else:
                            pbar.update(1)
1752
1753
                        if pbar.n == num_requests:
                            pbar.refresh()
1754

1755
1756
        if use_tqdm:
            pbar.close()
1757
1758
1759
        # Sort the outputs by request ID.
        # This is necessary because some requests may be finished earlier than
        # its previous requests.
1760
        return sorted(outputs, key=lambda x: int(x.request_id))