"vllm/vscode:/vscode.git/clone" did not exist on "0283f303d8c6f19f670306c50dba00e897ae94f9"
llm.py 78.1 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 Callable, Iterable, Sequence
6
from pathlib import Path
7
from typing import TYPE_CHECKING, Any
8

9
import cloudpickle
10
import torch.nn as nn
11
from pydantic import ValidationError
12
from tqdm.auto import tqdm
13
from typing_extensions import TypeVar, overload
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 (
22
    AttentionConfig,
23
    CompilationConfig,
24
    PoolerConfig,
25
    ProfilerConfig,
26
27
28
    StructuredOutputsConfig,
    is_init_field,
)
29
from vllm.config.compilation import CompilationMode
30
from vllm.config.model import (
31
32
    ConvertOption,
    HfOverrides,
33
    ModelDType,
34
    RunnerOption,
35
    TokenizerMode,
36
)
37
38
39
from vllm.config.quantization import (
    OnlineQuantizationConfigArgs,
)
40
41
42
43
from vllm.distributed.weight_transfer.base import (
    WeightTransferInitRequest,
    WeightTransferUpdateRequest,
)
44
from vllm.engine.arg_utils import EngineArgs
45
46
from vllm.entrypoints.chat_utils import (
    ChatCompletionMessageParam,
47
    ChatTemplateConfig,
48
    ChatTemplateContentFormatOption,
49
    load_chat_template,
50
)
51
from vllm.entrypoints.pooling.io_processor_factories import init_pooling_io_processors
52
53
from vllm.entrypoints.pooling.scoring.io_processor import (
    ScoringIOProcessor,
54
)
55
56
from vllm.entrypoints.pooling.scoring.typing import ScoreInput
from vllm.entrypoints.pooling.typing import OfflineInputsContext, OfflineOutputsContext
57
from vllm.entrypoints.utils import log_non_default_args
58
from vllm.inputs import (
59
    DataPrompt,
60
    EngineInput,
61
62
63
64
    PromptType,
    TextPrompt,
    TokensPrompt,
)
65
from vllm.logger import init_logger
66
from vllm.lora.request import LoRARequest
67
from vllm.model_executor.layers.quantization import QuantizationMethods
68
69
70
71
72
73
74
from vllm.outputs import (
    ClassificationRequestOutput,
    EmbeddingRequestOutput,
    PoolingRequestOutput,
    RequestOutput,
    ScoringRequestOutput,
)
75
from vllm.platforms import current_platform
76
from vllm.pooling_params import PoolingParams
77
from vllm.renderers import ChatParams, merge_kwargs
78
79
80
81
82
from vllm.renderers.inputs.preprocess import (
    conversation_to_seq,
    parse_model_prompt,
    prompt_to_seq,
)
83
from vllm.sampling_params import BeamSearchParams, RequestOutputKind, SamplingParams
84
from vllm.tasks import PoolingTask
85
from vllm.tokenizers import TokenizerLike
yhu422's avatar
yhu422 committed
86
from vllm.usage.usage_lib import UsageContext
87
from vllm.utils.counter import Counter
88
from vllm.utils.mistral import is_mistral_tokenizer
89
from vllm.utils.tqdm_utils import maybe_tqdm
90
from vllm.v1.engine import PauseMode
91
from vllm.v1.engine.llm_engine import LLMEngine
92
from vllm.v1.sample.logits_processor import LogitsProcessor
93

94
95
96
if TYPE_CHECKING:
    from vllm.v1.metrics.reader import Metric

97
98
logger = init_logger(__name__)

99
100
101
102
103
_O = TypeVar(
    "_O",
    bound=RequestOutput | PoolingRequestOutput,
    default=RequestOutput | PoolingRequestOutput,
)
104
_P = TypeVar("_P", bound=SamplingParams | PoolingParams | None)
105
106
_R = TypeVar("_R", default=Any)

107
108

class LLM:
Woosuk Kwon's avatar
Woosuk Kwon committed
109
110
111
112
113
114
115
116
117
118
    """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.
119
        tokenizer: The name or path of a HuggingFace Transformers tokenizer.
120
121
        tokenizer_mode: The tokenizer mode. "auto" will use the fast tokenizer
            if available, and "slow" will always use the slow tokenizer.
122
123
124
        skip_tokenizer_init: If true, skip initialization of tokenizer and
            detokenizer. Expect valid prompt_token_ids and None for prompt
            from the input.
125
126
        trust_remote_code: Trust remote code (e.g., from HuggingFace) when
            downloading the model and tokenizer.
127
128
129
130
        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.
131
        allowed_media_domains: If set, only media URLs that belong to this
132
            domain can be used for multi-modal inputs.
Woosuk Kwon's avatar
Woosuk Kwon committed
133
134
135
        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
136
            we support `float32`, `float16`, and `bfloat16`. If `auto`, we use
137
138
            the `dtype` attribute of the Transformers model's config. However,
            if the `dtype` in the config is `float32`, we will use `float16` instead.
139
        quantization: The method used to quantize the model weights. Currently,
140
            we support "awq", "gptq", and "fp8" (experimental).
141
142
143
144
            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
145
146
        revision: The specific model version to use. It can be a branch name,
            a tag name, or a commit id.
147
148
        tokenizer_revision: The specific tokenizer version to use. It can be a
            branch name, a tag name, or a commit id.
149
        chat_template: The chat template to apply.
150
151
152
153
154
155
        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.
156
157
158
159
160
        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
161
            compared with using gpu_memory_utilization. Note that
162
163
            kv_cache_memory_bytes (when not-None) ignores
            gpu_memory_utilization
164
165
166
167
        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.
168
169
170
171
172
173
174
175
176
177
178
179
180
        offload_group_size: Prefetch offloading: Group every N layers
            together. Offload last `offload_num_in_group` layers of each group.
            Default is 0 (disabled).
        offload_num_in_group: Prefetch offloading: Number of layers to
            offload per group. Default is 1.
        offload_prefetch_step: Prefetch offloading: Number of layers to
            prefetch ahead. Higher values hide more latency but use more GPU
            memory. Default is 1.
        offload_params: Prefetch offloading: Set of parameter name segments
            to selectively offload. Only parameters whose names contain one of
            these segments will be offloaded (e.g., {"gate_up_proj", "down_proj"}
            for MLP weights, or {"w13_weight", "w2_weight"} for MoE expert
            weights). If None or empty, all parameters are offloaded.
181
182
183
        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.
184
        enable_return_routed_experts: Whether to return routed experts.
185
186
        disable_custom_all_reduce: See
            [ParallelConfig][vllm.config.ParallelConfig].
187
        hf_token: The token to use as HTTP bearer authorization for remote files
188
            . If `True`, will use the token generated when running
189
            `hf auth login` (stored in `~/.cache/huggingface/token`).
190
191
192
        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.
193
194
195
196
197
        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}`.
198
199
        pooler_config: Initialize non-default pooling config for the pooling model,
            e.g., `PoolerConfig(seq_pooling_type="MEAN", use_activation=False)`.
200
        compilation_config: Either an integer or a dictionary. If it is an
201
            integer, it is used as the mode of compilation optimization. If it
202
            is a dictionary, it can specify the full compilation configuration.
203
204
205
206
        attention_config: Configuration for attention mechanisms. Can be a
            dictionary or an AttentionConfig instance. If a dictionary, it will
            be converted to an AttentionConfig. Allows specifying the attention
            backend and other attention-related settings.
207
        **kwargs: Arguments for [`EngineArgs`][vllm.EngineArgs].
nunjunj's avatar
nunjunj committed
208

209
210
    Note:
        This class is intended to be used for offline inference. For online
211
        serving, use the [AsyncLLMEngine][vllm.AsyncLLMEngine] class instead.
212
    """
213
214
215
216

    def __init__(
        self,
        model: str,
217
        *,
218
219
        runner: RunnerOption = "auto",
        convert: ConvertOption = "auto",
220
        tokenizer: str | None = None,
221
        tokenizer_mode: TokenizerMode | str = "auto",
222
        skip_tokenizer_init: bool = False,
223
        trust_remote_code: bool = False,
224
        allowed_local_media_path: str = "",
225
        allowed_media_domains: list[str] | None = None,
226
        tensor_parallel_size: int = 1,
227
        dtype: ModelDType = "auto",
228
229
230
        quantization: QuantizationMethods | None = None,
        revision: str | None = None,
        tokenizer_revision: str | None = None,
231
        chat_template: Path | str | None = None,
232
        seed: int = 0,
233
        gpu_memory_utilization: float = 0.9,
234
        cpu_offload_gb: float = 0,
235
236
237
238
        offload_group_size: int = 0,
        offload_num_in_group: int = 1,
        offload_prefetch_step: int = 1,
        offload_params: set[str] | None = None,
239
        enforce_eager: bool = False,
240
        enable_return_routed_experts: bool = False,
241
        disable_custom_all_reduce: bool = False,
242
243
244
245
246
247
248
        hf_token: bool | str | None = None,
        hf_overrides: HfOverrides | None = None,
        mm_processor_kwargs: dict[str, Any] | None = None,
        pooler_config: PoolerConfig | None = None,
        structured_outputs_config: dict[str, Any]
        | StructuredOutputsConfig
        | None = None,
249
        profiler_config: dict[str, Any] | ProfilerConfig | None = None,
250
        attention_config: dict[str, Any] | AttentionConfig | None = None,
251
252
        kv_cache_memory_bytes: int | None = None,
        compilation_config: int | dict[str, Any] | CompilationConfig | None = None,
253
254
255
        quantization_config: dict[str, Any]
        | OnlineQuantizationConfigArgs
        | None = None,
256
        logits_processors: list[str | type[LogitsProcessor]] | None = None,
257
        **kwargs: Any,
258
    ) -> None:
259
        """LLM constructor."""
260

261
262
263
264
265
266
267
268
269
270
271
        if "swap_space" in kwargs:
            kwargs.pop("swap_space")
            import warnings

            warnings.warn(
                "The 'swap_space' parameter is deprecated and ignored. "
                "It will be removed in a future version.",
                DeprecationWarning,
                stacklevel=2,
            )

272
273
        if "disable_log_stats" not in kwargs:
            kwargs["disable_log_stats"] = True
274

275
276
277
278
279
280
281
        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)

282
        if "kv_transfer_config" in kwargs and isinstance(
283
284
            kwargs["kv_transfer_config"], dict
        ):
285
            from vllm.config.kv_transfer import KVTransferConfig
286

287
288
            raw_config_dict = kwargs["kv_transfer_config"]
            try:
289
                kwargs["kv_transfer_config"] = KVTransferConfig(**raw_config_dict)
290
291
292
293
            except ValidationError as e:
                logger.error(
                    "Failed to convert 'kv_transfer_config' dict to "
                    "KVTransferConfig object. Dict: %s. Error: %s",
294
295
296
                    raw_config_dict,
                    e,
                )
297
298
                # Consider re-raising a more specific vLLM error or ValueError
                # to provide better context to the user.
299
                raise ValueError(f"Invalid 'kv_transfer_config' provided: {e}") from e
300

301
302
303
        if hf_overrides is None:
            hf_overrides = {}

304
305
306
307
308
309
310
        def _make_config(value: Any, cls: type[_R]) -> _R:
            """Convert dict/None/instance to a config instance."""
            if value is None:
                return cls()
            if isinstance(value, dict):
                return cls(**{k: v for k, v in value.items() if is_init_field(cls, k)})  # type: ignore[arg-type]
            return value
311

312
313
314
315
        if isinstance(compilation_config, int):
            compilation_config_instance = CompilationConfig(
                mode=CompilationMode(compilation_config)
            )
316
        else:
317
318
319
            compilation_config_instance = _make_config(
                compilation_config, CompilationConfig
            )
320

321
322
323
324
325
        structured_outputs_instance = _make_config(
            structured_outputs_config, StructuredOutputsConfig
        )
        profiler_config_instance = _make_config(profiler_config, ProfilerConfig)
        attention_config_instance = _make_config(attention_config, AttentionConfig)
326

327
        # warn about single-process data parallel usage.
328
329
        _dp_size = int(kwargs.get("data_parallel_size", 1))
        _distributed_executor_backend = kwargs.get("distributed_executor_backend")
330
331
332
333
334
        if (
            _dp_size > 1
            and not _distributed_executor_backend == "external_launcher"
            and not current_platform.is_tpu()
        ):
335
            raise ValueError(
336
                f"LLM(data_parallel_size={_dp_size}) is not supported for single-"
337
338
339
340
341
                "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
342
        engine_args = EngineArgs(
343
            model=model,
344
345
            runner=runner,
            convert=convert,
346
            tokenizer=tokenizer,
347
            tokenizer_mode=tokenizer_mode,
348
            skip_tokenizer_init=skip_tokenizer_init,
349
            trust_remote_code=trust_remote_code,
350
            allowed_local_media_path=allowed_local_media_path,
351
            allowed_media_domains=allowed_media_domains,
352
353
            tensor_parallel_size=tensor_parallel_size,
            dtype=dtype,
354
            quantization=quantization,
355
            revision=revision,
356
            tokenizer_revision=tokenizer_revision,
357
358
            seed=seed,
            gpu_memory_utilization=gpu_memory_utilization,
359
            kv_cache_memory_bytes=kv_cache_memory_bytes,
360
            cpu_offload_gb=cpu_offload_gb,
361
362
363
364
            offload_group_size=offload_group_size,
            offload_num_in_group=offload_num_in_group,
            offload_prefetch_step=offload_prefetch_step,
            offload_params=offload_params or set(),
365
            enforce_eager=enforce_eager,
366
            enable_return_routed_experts=enable_return_routed_experts,
367
            disable_custom_all_reduce=disable_custom_all_reduce,
368
            hf_token=hf_token,
369
            hf_overrides=hf_overrides,
370
            mm_processor_kwargs=mm_processor_kwargs,
371
            pooler_config=pooler_config,
372
            structured_outputs_config=structured_outputs_instance,
373
            profiler_config=profiler_config_instance,
374
            attention_config=attention_config_instance,
375
            compilation_config=compilation_config_instance,
376
            quantization_config=quantization_config,
377
            logits_processors=logits_processors,
378
379
            **kwargs,
        )
380

381
382
        log_non_default_args(engine_args)

383
        self.llm_engine = LLMEngine.from_engine_args(
384
385
            engine_args=engine_args, usage_context=UsageContext.LLM_CLASS
        )
386
        self.model_config = self.llm_engine.model_config
387
        self.engine_class = type(self.llm_engine)
388

389
        self.request_counter = Counter()
390
        self.default_sampling_params: dict[str, Any] | None = None
391

392
        supported_tasks = self.llm_engine.get_supported_tasks()
393
        self.supported_tasks = supported_tasks
394
395
396
        self.pooling_task = self.model_config.get_pooling_task(supported_tasks)
        if self.pooling_task is not None:
            logger.info("Supported pooling task: %s", self.pooling_task)
397

398
        self.runner_type = self.model_config.runner_type
399
        self.renderer = self.llm_engine.renderer
400
        self.chat_template = load_chat_template(chat_template)
401
        self.io_processor = self.llm_engine.io_processor
402
        self.input_processor = self.llm_engine.input_processor
403
        self.chat_template_config = ChatTemplateConfig(chat_template=self.chat_template)
404
        self.pooling_io_processors = init_pooling_io_processors(
405
406
407
408
409
            supported_tasks=supported_tasks,
            model_config=self.model_config,
            renderer=self.renderer,
            chat_template_config=self.chat_template_config,
        )
410
411
412
        # Cache for __repr__ to avoid repeated collective_rpc calls
        self._cached_repr: str | None = None

413
414
415
416
417
    @classmethod
    def from_engine_args(cls, engine_args: EngineArgs) -> "LLM":
        """Create an LLM instance from EngineArgs."""
        return cls(**vars(engine_args))

418
    def get_tokenizer(self) -> TokenizerLike:
419
        return self.llm_engine.get_tokenizer()
420

421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
    def get_world_size(self, include_dp: bool = True) -> int:
        """Get the world size from the parallel config.

        Args:
            include_dp: If True (default), returns the world size including
                data parallelism (TP * PP * DP). If False, returns the world
                size without data parallelism (TP * PP).

        Returns:
            The world size (tensor_parallel_size * pipeline_parallel_size),
            optionally multiplied by data_parallel_size if include_dp is True.
        """
        parallel_config = self.llm_engine.vllm_config.parallel_config
        if include_dp:
            return parallel_config.world_size_across_dp
        return parallel_config.world_size

438
    def reset_mm_cache(self) -> None:
439
        self.renderer.clear_mm_cache()
440
441
        self.llm_engine.reset_mm_cache()

442
    def get_default_sampling_params(self) -> SamplingParams:
443
        if self.default_sampling_params is None:
444
            self.default_sampling_params = self.model_config.get_diff_sampling_param()
445
446
        if self.default_sampling_params:
            return SamplingParams.from_optional(**self.default_sampling_params)
447
448
        return SamplingParams()

449
450
    def generate(
        self,
451
452
        prompts: PromptType | Sequence[PromptType],
        sampling_params: SamplingParams | Sequence[SamplingParams] | None = None,
453
        *,
454
        use_tqdm: bool | Callable[..., tqdm] = True,
455
        lora_request: Sequence[LoRARequest] | LoRARequest | None = None,
456
        priority: list[int] | None = None,
457
        tokenization_kwargs: dict[str, Any] | None = None,
458
    ) -> list[RequestOutput]:
Woosuk Kwon's avatar
Woosuk Kwon committed
459
460
        """Generates the completions for the input prompts.

461
        This class automatically batches the given prompts, considering
Woosuk Kwon's avatar
Woosuk Kwon committed
462
463
464
465
        the memory constraint. For the best performance, put all of your prompts
        into a single list and pass it to this method.

        Args:
466
            prompts: The prompts to the LLM. You may pass a sequence of prompts
467
                for batch inference. See [PromptType][vllm.inputs.PromptType]
468
                for more details about the format of each prompt.
Woosuk Kwon's avatar
Woosuk Kwon committed
469
            sampling_params: The sampling parameters for text generation. If
nunjunj's avatar
nunjunj committed
470
471
472
                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
473
                prompts and it is paired one by one with the prompt.
474
475
476
477
            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.
478
            lora_request: LoRA request to use for generation, if any.
479
480
            priority: The priority of the requests, if any.
                Only applicable when priority scheduling policy is enabled.
481
482
483
                If provided, must be a list of integers matching the length
                of `prompts`, where each priority value corresponds to the prompt
                at the same index.
484
            tokenization_kwargs: Overrides for `tokenizer.encode`.
Woosuk Kwon's avatar
Woosuk Kwon committed
485
486

        Returns:
487
            A list of `RequestOutput` objects containing the
488
489
            generated completions in the same order as the input prompts.
        """
490
        runner_type = self.model_config.runner_type
491
        if runner_type != "generate":
492
493
494
            raise ValueError(
                "LLM.generate() is only supported for generative models. "
                "Try passing `--runner generate` to use the model as a "
495
496
                "generative model."
            )
497

498
        if sampling_params is None:
499
            sampling_params = self.get_default_sampling_params()
500

501
        return self._run_completion(
502
            prompts=prompts,
503
            params=sampling_params,
504
            output_type=RequestOutput,
505
            use_tqdm=use_tqdm,
506
            lora_request=lora_request,
507
            tokenization_kwargs=tokenization_kwargs,
508
509
            priority=priority,
        )
510

511
512
513
514
    def enqueue(
        self,
        prompts: PromptType | Sequence[PromptType],
        sampling_params: SamplingParams | Sequence[SamplingParams] | None = None,
515
        lora_request: Sequence[LoRARequest] | LoRARequest | None = None,
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
        priority: list[int] | None = None,
        use_tqdm: bool | Callable[..., tqdm] = True,
        tokenization_kwargs: dict[str, Any] | None = None,
    ) -> list[str]:
        """Enqueue prompts for generation without waiting for completion.

        This method adds requests to the engine queue but does not start
        processing them. Use wait_for_completion() to process the queued
        requests and get results.

        Args:
            prompts: The prompts to the LLM. See generate() for details.
            sampling_params: The sampling parameters for text generation.
            lora_request: LoRA request to use for generation, if any.
            priority: The priority of the requests, if any.
            use_tqdm: If True, shows a tqdm progress bar while adding requests.
            tokenization_kwargs: Overrides for `tokenizer.encode`.

        Returns:
            A list of request IDs for the enqueued requests.
        """
537
        runner_type = self.model_config.runner_type
538
539
540
541
542
543
        if runner_type != "generate":
            raise ValueError("LLM.enqueue() is only supported for generative models.")

        if sampling_params is None:
            sampling_params = self.get_default_sampling_params()

544
545
546
547
548
549
550
        return self._add_completion_requests(
            prompts=prompts,
            params=sampling_params,
            use_tqdm=use_tqdm,
            lora_request=lora_request,
            priority=priority,
            tokenization_kwargs=tokenization_kwargs,
551
552
        )

553
    @overload
554
555
    def wait_for_completion(
        self,
556
        *,
557
        use_tqdm: bool | Callable[..., tqdm] = True,
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
    ) -> list[RequestOutput | PoolingRequestOutput]: ...

    @overload
    def wait_for_completion(
        self,
        output_type: type[_O] | tuple[type[_O], ...],
        *,
        use_tqdm: bool | Callable[..., tqdm] = True,
    ) -> list[_O]: ...

    def wait_for_completion(
        self,
        output_type: type[Any] | tuple[type[Any], ...] | None = None,
        *,
        use_tqdm: bool | Callable[..., tqdm] = True,
    ) -> list[Any]:
574
575
576
577
578
579
        """Wait for all enqueued requests to complete and return results.

        This method processes all requests currently in the engine queue
        and returns their outputs. Use after enqueue() to get results.

        Args:
580
            output_type: The expected output type, defaults to RequestOutput.
581
582
583
            use_tqdm: If True, shows a tqdm progress bar.

        Returns:
584
            A list of output objects for all completed requests.
585
        """
586
587
588
589
        if output_type is None:
            output_type = (RequestOutput, PoolingRequestOutput)

        return self._run_engine(output_type, use_tqdm=use_tqdm)
590

Cyrus Leung's avatar
Cyrus Leung committed
591
    def _resolve_mm_lora(
592
        self,
593
        prompt: EngineInput,
594
        lora_request: LoRARequest | None,
Cyrus Leung's avatar
Cyrus Leung committed
595
596
597
598
599
600
601
    ) -> LoRARequest | None:
        if prompt["type"] != "multimodal":
            return lora_request

        lora_config = self.llm_engine.vllm_config.lora_config
        default_mm_loras = None if lora_config is None else lora_config.default_mm_loras
        if not default_mm_loras:
602
603
            return lora_request

604
605
        prompt_modalities = prompt["mm_placeholders"].keys()
        intersection = set(prompt_modalities).intersection(default_mm_loras.keys())
606
607
        if not intersection:
            return lora_request
Cyrus Leung's avatar
Cyrus Leung committed
608

609
610
611
        if len(intersection) > 1:
            # TODO: Would be nice to be able to have multiple loras per prompt
            logger.warning(
Cyrus Leung's avatar
Cyrus Leung committed
612
613
614
615
                "Multiple modality specific loras were registered and would be "
                "used by a single prompt consuming several modalities; "
                "currently we only support one lora per request; as such, "
                "lora(s) registered with modalities: %s will be skipped",
616
617
                intersection,
            )
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
            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 "
633
634
                    "lora_request as we only apply one LoRARequest per prompt"
                )
635
636
637
638
639
640
641
642
            return lora_request

        return LoRARequest(
            modality_name,
            modality_lora_id,
            modality_lora_path,
        )

643
644
    def collective_rpc(
        self,
645
646
        method: str | Callable[..., _R],
        timeout: float | None = None,
647
        args: tuple = (),
648
        kwargs: dict[str, Any] | None = None,
649
    ) -> list[_R]:
650
651
652
653
654
655
656
657
658
659
660
        """
        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
661
                [`TimeoutError`][] on timeout. `None` means wait indefinitely.
662
663
664
665
666
            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.
667

668
669
670
        Note:
            It is recommended to use this API to only pass control messages,
            and set up data-plane communication to pass data.
671
        """
672
673

        return self.llm_engine.collective_rpc(method, timeout, args, kwargs)
674
675

    def apply_model(self, func: Callable[[nn.Module], _R]) -> list[_R]:
676
        """
677
678
        Run a function directly on the model inside each worker,
        returning the result for each of them.
679
680
681
682
683
684

        !!! 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!
685
        """
686
        return self.llm_engine.apply_model(func)
687

688
689
    def beam_search(
        self,
690
        prompts: list[TokensPrompt | TextPrompt],
691
        params: BeamSearchParams,
692
        lora_request: list[LoRARequest] | LoRARequest | None = None,
693
        use_tqdm: bool = False,
694
        concurrency_limit: int | None = None,
695
    ) -> list[BeamSearchOutput]:
696
697
698
699
700
701
        """
        Generate sequences using beam search.

        Args:
            prompts: A list of prompts. Each prompt can be a string or a list
                of token IDs.
702
            params: The beam search parameters.
703
            lora_request: LoRA request to use for generation, if any.
704
            use_tqdm: Whether to use tqdm to display the progress bar.
705
706
            concurrency_limit: The maximum number of concurrent requests.
                If None, the number of concurrent requests is unlimited.
707
        """
708
709
        # TODO: how does beam search work together with length penalty,
        # frequency, penalty, and stopping criteria, etc.?
710
711
712
713
        beam_width = params.beam_width
        max_tokens = params.max_tokens
        temperature = params.temperature
        ignore_eos = params.ignore_eos
714
715
        length_penalty = params.length_penalty

716
717
718
        tokenizer = self.renderer.get_tokenizer()
        eos_token_id = tokenizer.eos_token_id
        sort_beams_key = create_sort_beams_key_function(eos_token_id, length_penalty)
719

720
721
        engine_inputs = self._preprocess_cmpl(prompts)
        lora_requests = self._lora_request_to_seq(lora_request, len(engine_inputs))
722

723
724
725
        if use_tqdm and concurrency_limit is not None:
            logger.warning(
                "Progress bar is not supported when using concurrency_limit. "
726
727
                "Disabling progress bar."
            )
728
729
730
            use_tqdm = False

        if concurrency_limit is None:
731
            concurrency_limit = len(engine_inputs)
732

733
734
735
        # 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
736
        sampling_params = SamplingParams(
737
738
739
740
            logprobs=2 * beam_width,
            max_tokens=1,
            temperature=temperature,
            skip_clone=True,  # Internal beam search, safe to skip clone
741
        )
742
        instances: list[BeamSearchInstance] = []
743

744
        for lora_req, prompt in zip(lora_requests, engine_inputs):
745
746
747
748
            if prompt["type"] == "embeds":
                raise NotImplementedError(
                    "Embedding prompt not supported for beam search"
                )
749

750
            instances.append(
751
                BeamSearchInstance(
752
                    prompt,
753
754
                    lora_request=lora_req,
                    logprobs=None,
755
756
                ),
            )
757

758
        for prompt_start in range(0, len(instances), concurrency_limit):
759
            instances_batch = instances[prompt_start : prompt_start + concurrency_limit]
760
761
762

            token_iter = range(max_tokens)
            if use_tqdm:
763
764
765
                token_iter = tqdm(
                    token_iter, desc="Beam search", unit="token", unit_scale=False
                )
766
767
768
                logger.warning(
                    "The progress bar shows the upper bound on token steps and "
                    "may finish early due to stopping conditions. It does not "
769
770
                    "reflect instance-level progress."
                )
771
772
            for _ in token_iter:
                all_beams: list[BeamSearchSequence] = list(
773
774
                    sum((instance.beams for instance in instances_batch), [])
                )
775
776
                pos = [0] + list(
                    itertools.accumulate(
777
778
779
                        len(instance.beams) for instance in instances_batch
                    )
                )
780
                instance_start_and_end: list[tuple[int, int]] = list(
781
782
                    zip(pos[:-1], pos[1:])
                )
783
784
785
786
787
788

                if len(all_beams) == 0:
                    break

                # only runs for one step
                # we don't need to use tqdm here
789
                output = self._render_and_run_requests(
790
791
                    prompts=(beam.get_prompt() for beam in all_beams),
                    params=self._params_to_seq(sampling_params, len(all_beams)),
792
                    output_type=RequestOutput,
793
                    lora_requests=[beam.lora_request for beam in all_beams],
794
795
                    use_tqdm=False,
                )
796

797
798
799
                for (start, end), instance in zip(
                    instance_start_and_end, instances_batch
                ):
800
801
802
803
804
805
806
807
808
809
810
811
812
                    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(
813
                                    current_beam.orig_prompt,
814
                                    tokens=current_beam.tokens + [token_id],
815
                                    logprobs=current_beam.logprobs + [logprobs],
816
                                    lora_request=current_beam.lora_request,
817
818
819
820
                                    cum_logprob=current_beam.cum_logprob
                                    + logprob_obj.logprob,
                                )

821
                                if token_id == eos_token_id and not ignore_eos:
822
823
824
                                    instance.completed.append(new_beam)
                                else:
                                    instance_new_beams.append(new_beam)
825
826
827
                    sorted_beams = sorted(
                        instance_new_beams, key=sort_beams_key, reverse=True
                    )
828
                    instance.beams = sorted_beams[:beam_width]
829
830
831
832

        outputs = []
        for instance in instances:
            instance.completed.extend(instance.beams)
833
834
835
            sorted_completed = sorted(
                instance.completed, key=sort_beams_key, reverse=True
            )
836
837
838
839
            best_beams = sorted_completed[:beam_width]

            for beam in best_beams:
                beam.text = tokenizer.decode(beam.tokens)
840

841
842
843
844
            outputs.append(BeamSearchOutput(sequences=best_beams))

        return outputs

845
    def _preprocess_cmpl(
846
        self,
847
        prompts: Sequence[PromptType],
848
        tokenization_kwargs: dict[str, Any] | None = None,
849
    ) -> Sequence[EngineInput]:
850
851
852
853
854
855
856
        """
        Convert prompt inputs from LLM APIs (other than [LLM.chat][]) into
        a format that can be passed to `_add_request`.

        Refer to [LLM.generate][] for a complete description of the arguments.

        Returns:
857
            A list of `EngineInput` objects ready to be passed into LLMEngine.
858
        """
859
        renderer = self.renderer
860
861
        model_config = self.model_config

862
863
864
        parsed_prompts = [
            parse_model_prompt(model_config, prompt) for prompt in prompts
        ]
865
866
867
        tok_params = renderer.default_cmpl_tok_params.with_kwargs(
            **(tokenization_kwargs or {})
        )
868

869
        return renderer.render_cmpl(parsed_prompts, tok_params)
870

871
872
873
874
    def _preprocess_cmpl_one(
        self,
        prompt: PromptType,
        tokenization_kwargs: dict[str, Any] | None = None,
875
876
877
    ) -> EngineInput:
        (engine_input,) = self._preprocess_cmpl([prompt], tokenization_kwargs)
        return engine_input
878

879
880
    def _preprocess_chat(
        self,
881
        conversations: Sequence[list[ChatCompletionMessageParam]],
882
        chat_template: str | None = None,
883
        chat_template_content_format: ChatTemplateContentFormatOption = "auto",
884
        chat_template_kwargs: dict[str, Any] | None = None,
885
        add_generation_prompt: bool = True,
886
        continue_final_message: bool = False,
887
        tools: list[dict[str, Any]] | None = None,
888
        tokenization_kwargs: dict[str, Any] | None = None,
889
        mm_processor_kwargs: dict[str, Any] | None = None,
890
    ) -> Sequence[EngineInput]:
nunjunj's avatar
nunjunj committed
891
        """
892
893
894
895
        Convert a list of conversations into prompts so that they can then
        be used as input for other LLM APIs.

        Refer to [LLM.chat][] for a complete description of the arguments.
nunjunj's avatar
nunjunj committed
896
897

        Returns:
898
            A list of `EngineInput` objects ready to be passed into LLMEngine.
nunjunj's avatar
nunjunj committed
899
        """
900
        renderer = self.renderer
901

902
903
904
905
906
907
908
909
910
        chat_params = ChatParams(
            chat_template=chat_template,
            chat_template_content_format=chat_template_content_format,
            chat_template_kwargs=merge_kwargs(
                chat_template_kwargs,
                dict(
                    add_generation_prompt=add_generation_prompt,
                    continue_final_message=continue_final_message,
                    tools=tools,
911
                    tokenize=is_mistral_tokenizer(renderer.tokenizer),
912
913
914
                ),
            ),
        )
915
916
917
        tok_params = renderer.default_chat_tok_params.with_kwargs(
            **(tokenization_kwargs or {})
        )
918

919
        _, engine_inputs = renderer.render_chat(
920
921
922
923
924
            conversations,
            chat_params,
            tok_params,
            prompt_extras={"mm_processor_kwargs": mm_processor_kwargs},
        )
925

926
        return engine_inputs
927

928
929
930
931
932
933
934
935
936
937
938
    def _preprocess_chat_one(
        self,
        conversation: list[ChatCompletionMessageParam],
        chat_template: str | None = None,
        chat_template_content_format: ChatTemplateContentFormatOption = "auto",
        chat_template_kwargs: dict[str, Any] | None = None,
        add_generation_prompt: bool = True,
        continue_final_message: bool = False,
        tools: list[dict[str, Any]] | None = None,
        tokenization_kwargs: dict[str, Any] | None = None,
        mm_processor_kwargs: dict[str, Any] | None = None,
939
940
    ) -> EngineInput:
        (engine_input,) = self._preprocess_chat(
941
942
943
944
945
946
947
948
949
950
951
            [conversation],
            chat_template=chat_template,
            chat_template_content_format=chat_template_content_format,
            chat_template_kwargs=chat_template_kwargs,
            add_generation_prompt=add_generation_prompt,
            continue_final_message=continue_final_message,
            tools=tools,
            tokenization_kwargs=tokenization_kwargs,
            mm_processor_kwargs=mm_processor_kwargs,
        )

952
        return engine_input
953

954
955
    def chat(
        self,
956
        messages: list[ChatCompletionMessageParam]
957
958
        | Sequence[list[ChatCompletionMessageParam]],
        sampling_params: SamplingParams | Sequence[SamplingParams] | None = None,
959
        use_tqdm: bool | Callable[..., tqdm] = True,
960
        lora_request: Sequence[LoRARequest] | LoRARequest | None = None,
961
        chat_template: str | None = None,
962
963
964
        chat_template_content_format: ChatTemplateContentFormatOption = "auto",
        add_generation_prompt: bool = True,
        continue_final_message: bool = False,
965
966
        tools: list[dict[str, Any]] | None = None,
        chat_template_kwargs: dict[str, Any] | None = None,
967
        tokenization_kwargs: dict[str, Any] | None = None,
968
        mm_processor_kwargs: dict[str, Any] | None = None,
969
970
971
972
973
974
975
976
977
978
979
980
    ) -> 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:
981
            messages: A sequence of conversations or a single conversation.
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012

                - 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.
1013
1014
            tokenization_kwargs: Overrides for `tokenizer.encode`.
            mm_processor_kwargs: Overrides for `processor.__call__`.
1015
1016
1017
1018
1019

        Returns:
            A list of `RequestOutput` objects containing the generated
            responses in the same order as the input messages.
        """
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
        model_config = self.model_config
        runner_type = model_config.runner_type
        if runner_type != "generate":
            raise ValueError(
                "LLM.chat() is only supported for generative models. "
                "Try passing `--runner generate` to use the model as a "
                "generative model."
            )

        if sampling_params is None:
            sampling_params = self.get_default_sampling_params()

1032
        return self._run_chat(
1033
1034
            messages=messages,
            params=sampling_params,
1035
            output_type=RequestOutput,
1036
1037
            use_tqdm=use_tqdm,
            lora_request=lora_request,
1038
1039
            chat_template=chat_template,
            chat_template_content_format=chat_template_content_format,
1040
            chat_template_kwargs=chat_template_kwargs,
1041
1042
1043
            add_generation_prompt=add_generation_prompt,
            continue_final_message=continue_final_message,
            tools=tools,
1044
            tokenization_kwargs=tokenization_kwargs,
1045
1046
1047
            mm_processor_kwargs=mm_processor_kwargs,
        )

1048
1049
    def encode(
        self,
1050
1051
        prompts: PromptType | Sequence[PromptType] | DataPrompt,
        pooling_params: PoolingParams | Sequence[PoolingParams] | None = None,
1052
        *,
1053
1054
        use_tqdm: bool | Callable[..., tqdm] = True,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
1055
        pooling_task: PoolingTask | None = None,
1056
        tokenization_kwargs: dict[str, Any] | None = None,
1057
    ) -> list[PoolingRequestOutput]:
1058
1059
        """Apply pooling to the hidden states corresponding to the input
        prompts.
1060

1061
        This class automatically batches the given prompts, considering
1062
1063
1064
1065
        the memory constraint. For the best performance, put all of your prompts
        into a single list and pass it to this method.

        Args:
1066
            prompts: The prompts to the LLM. You may pass a sequence of prompts
1067
                for batch inference. See [PromptType][vllm.inputs.PromptType]
1068
                for more details about the format of each prompt.
1069
1070
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1071
1072
1073
1074
            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.
1075
            lora_request: LoRA request to use for generation, if any.
1076
            pooling_task: Override the pooling task to use.
1077
            tokenization_kwargs: Overrides for `tokenizer.encode`.
1078
1079

        Returns:
1080
            A list of `PoolingRequestOutput` objects containing the
1081
            pooled hidden states in the same order as the input prompts.
1082
        """
1083

1084
        self._verify_pooling_task(pooling_task)
1085

1086
        if isinstance(prompts, dict) and "data" in prompts:
1087
1088
1089
1090
1091
            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' "
1092
1093
                    "offline inference example for more details."
                )
1094
1095

            # Validate the request data is valid for the loaded plugin
1096
1097
1098
1099
1100
1101
1102
1103
1104
            prompt_data = prompts.get("data")
            if prompt_data is None:
                raise ValueError(
                    "The 'data' field of the prompt is expected to contain "
                    "the prompt data and it cannot be None. "
                    "Refer to the documentation of the IOProcessor "
                    "in use for more details."
                )
            validated_prompt = self.io_processor.parse_data(prompt_data)
1105
1106
1107

            # obtain the actual model prompts from the pre-processor
            prompts = self.io_processor.pre_process(prompt=validated_prompt)
1108
            prompts_seq = prompt_to_seq(prompts)
1109

1110
1111
1112
1113
1114
            params_seq: Sequence[PoolingParams] = [
                self.io_processor.merge_pooling_params(param)
                for param in self._params_to_seq(
                    pooling_params,
                    len(prompts_seq),
1115
                )
1116
1117
1118
1119
            ]
            for p in params_seq:
                if p.task is None:
                    p.task = "plugin"
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144

            outputs = self._run_completion(
                prompts=prompts_seq,
                params=params_seq,
                output_type=PoolingRequestOutput,
                use_tqdm=use_tqdm,
                lora_request=lora_request,
                tokenization_kwargs=tokenization_kwargs,
            )

            # get the post-processed model outputs
            assert self.io_processor is not None
            processed_outputs = self.io_processor.post_process(outputs)

            return [
                PoolingRequestOutput[Any](
                    request_id="",
                    outputs=processed_outputs,
                    num_cached_tokens=getattr(
                        processed_outputs, "num_cached_tokens", 0
                    ),
                    prompt_token_ids=[],
                    finished=True,
                )
            ]
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
        else:
            if pooling_params is None:
                # Use default pooling params.
                pooling_params = PoolingParams()

            prompts_seq = prompt_to_seq(prompts)
            params_seq = self._params_to_seq(pooling_params, len(prompts_seq))

            for param in params_seq:
                if param.task is None:
                    param.task = pooling_task
                elif param.task != pooling_task:
                    msg = (
                        f"You cannot overwrite {param.task=!r} with {pooling_task=!r}!"
                    )
                    raise ValueError(msg)
1161

1162
1163
            if pooling_task in self.pooling_io_processors:
                io_processor = self.pooling_io_processors[pooling_task]
1164
                processor_inputs = io_processor.pre_process_offline(
1165
1166
1167
                    ctx=OfflineInputsContext(
                        prompts=prompts_seq, tokenization_kwargs=tokenization_kwargs
                    )
1168
1169
1170
1171
1172
                )
                seq_lora_requests = self._lora_request_to_seq(
                    lora_request, len(prompts_seq)
                )
                seq_priority = self._priority_to_seq(None, len(prompts))
1173

1174
1175
1176
1177
1178
                self._render_and_add_requests(
                    prompts=processor_inputs,
                    params=params_seq,
                    lora_requests=seq_lora_requests,
                    priorities=seq_priority,
1179
                )
1180

1181
1182
1183
                outputs = self._run_engine(
                    use_tqdm=use_tqdm, output_type=PoolingRequestOutput
                )
1184
1185
1186
                outputs = io_processor.post_process_offline(
                    ctx=OfflineOutputsContext(outputs=outputs)
                )
1187
1188
1189
1190
1191
1192
1193
1194
1195
            else:
                outputs = self._run_completion(
                    prompts=prompts_seq,
                    params=params_seq,
                    output_type=PoolingRequestOutput,
                    use_tqdm=use_tqdm,
                    lora_request=lora_request,
                    tokenization_kwargs=tokenization_kwargs,
                )
1196
        return outputs
1197

1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
    def _verify_pooling_task(self, pooling_task: PoolingTask | None):
        if self.runner_type != "pooling":
            raise ValueError(
                "LLM.encode() is only supported for pooling models. "
                "Try passing `--runner pooling` to use the model as a "
                "pooling model."
            )

        if pooling_task is None:
            raise ValueError(
                "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"`'
            )

        if (
            pooling_task in ("embed", "token_embed")
            and pooling_task not in self.supported_tasks
        ):
            raise ValueError(
                "Embedding API is not supported by this model. "
                "Try converting the model using `--convert embed`."
            )

        if (
            pooling_task in ("classify", "token_classify")
            and pooling_task not in self.supported_tasks
        ):
            raise ValueError(
                "Classification API is not supported by this model. "
                "Try converting the model using `--convert classify`."
            )

        # plugin task uses io_processor.parse_request to verify inputs
        if pooling_task != "plugin" and pooling_task != self.pooling_task:
            if pooling_task not in self.supported_tasks:
                raise ValueError(
                    f"Unsupported task: {pooling_task!r} "
                    f"Supported tasks: {self.supported_tasks}"
                )
            else:
                logger.warning_once(
                    "Pooling multitask support is deprecated and will "
                    "be removed in v0.20. When the default pooling task is "
                    "not what you want, you need to manually specify it "
                    'via PoolerConfig(task="%s"). ',
                    pooling_task,
                )

1257
1258
    def embed(
        self,
1259
        prompts: PromptType | Sequence[PromptType],
1260
        *,
1261
1262
1263
        use_tqdm: bool | Callable[..., tqdm] = True,
        pooling_params: PoolingParams | Sequence[PoolingParams] | None = None,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
1264
        tokenization_kwargs: dict[str, Any] | None = None,
1265
    ) -> list[EmbeddingRequestOutput]:
1266
1267
1268
1269
1270
1271
1272
1273
1274
        """
        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
1275
                for batch inference. See [PromptType][vllm.inputs.PromptType]
1276
                for more details about the format of each prompt.
1277
1278
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1279
1280
1281
1282
            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.
1283
            lora_request: LoRA request to use for generation, if any.
1284
            tokenization_kwargs: Overrides for `tokenizer.encode`.
1285
1286

        Returns:
1287
            A list of `EmbeddingRequestOutput` objects containing the
1288
1289
1290
            embedding vectors in the same order as the input prompts.
        """

1291
1292
1293
1294
1295
1296
        items = self.encode(
            prompts,
            use_tqdm=use_tqdm,
            pooling_params=pooling_params,
            lora_request=lora_request,
            pooling_task="embed",
1297
            tokenization_kwargs=tokenization_kwargs,
1298
        )
1299
1300
1301
1302
1303

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

    def classify(
        self,
1304
        prompts: PromptType | Sequence[PromptType],
1305
        *,
1306
        pooling_params: PoolingParams | Sequence[PoolingParams] | None = None,
1307
        use_tqdm: bool | Callable[..., tqdm] = True,
1308
        lora_request: list[LoRARequest] | LoRARequest | None = None,
1309
        tokenization_kwargs: dict[str, Any] | None = None,
1310
    ) -> list[ClassificationRequestOutput]:
1311
1312
1313
1314
1315
1316
1317
1318
1319
        """
        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
1320
                for batch inference. See [PromptType][vllm.inputs.PromptType]
1321
                for more details about the format of each prompt.
1322
1323
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1324
1325
1326
1327
            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.
1328
            lora_request: LoRA request to use for generation, if any.
1329
1330
            tokenization_kwargs: Overrides for `tokenizer.encode`.

1331
        Returns:
1332
            A list of `ClassificationRequestOutput` objects containing the
1333
1334
1335
            embedding vectors in the same order as the input prompts.
        """

1336
1337
1338
        items = self.encode(
            prompts,
            use_tqdm=use_tqdm,
1339
            pooling_params=pooling_params,
1340
1341
            lora_request=lora_request,
            pooling_task="classify",
1342
            tokenization_kwargs=tokenization_kwargs,
1343
        )
1344
1345
1346

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

1347
1348
    def reward(
        self,
1349
        prompts: PromptType | Sequence[PromptType],
1350
1351
        /,
        *,
1352
        pooling_params: PoolingParams | Sequence[PoolingParams] | None = None,
1353
1354
        use_tqdm: bool | Callable[..., tqdm] = True,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
1355
        tokenization_kwargs: dict[str, Any] | None = None,
1356
1357
1358
1359
1360
1361
1362
    ) -> 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]
1363
                for more details about the format of each prompt.
1364
1365
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1366
1367
1368
1369
1370
            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.
1371
1372
            tokenization_kwargs: Overrides for `tokenizer.encode`.

1373
1374
1375
1376
1377
1378
1379
1380
1381
        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,
1382
            pooling_task="token_classify",
1383
            tokenization_kwargs=tokenization_kwargs,
1384
1385
        )

1386
1387
    def score(
        self,
1388
1389
        data_1: ScoreInput | list[ScoreInput],
        data_2: ScoreInput | list[ScoreInput],
1390
        /,
1391
        *,
1392
1393
1394
        use_tqdm: bool | Callable[..., tqdm] = True,
        pooling_params: PoolingParams | None = None,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
1395
        tokenization_kwargs: dict[str, Any] | None = None,
1396
        chat_template: str | None = None,
1397
    ) -> list[ScoringRequestOutput]:
1398
1399
        """Generate similarity scores for all pairs `<text,text_pair>` or
          `<multi-modal data, multi-modal data pair>`.
1400

1401
        The inputs can be `1 -> 1`, `1 -> N` or `N -> N`.
1402
1403
        In the `1 - N` case the `data_1` input will be replicated `N`
        times to pair with the `data_2` inputs.
1404
        The input pairs are used to build a list of prompts for the
1405
1406
        cross encoder model. This class automatically batches the prompts,
        considering the memory constraint. For the best performance, put all
1407
1408
1409
        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
1410
        appropriate multi-modal models. For multi-modal inputs, ensure the
1411
        prompt structure matches the model's expected input format.
1412
1413

        Args:
1414
1415
1416
            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
1417
                the `data_2` list.
1418
            data_2: The data to pair with the query to form the input to
1419
                the LLM. Can be text or multi-modal data. See [PromptType]
1420
                [vllm.inputs.PromptType] for more details about the format of
1421
                each prompt.
1422
1423
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1424
1425
1426
1427
            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.
1428
            lora_request: LoRA request to use for generation, if any.
1429
1430
            chat_template: The chat template to use for the scoring. If None, we
                use the model's default chat template.
1431
            tokenization_kwargs: Overrides for `tokenizer.encode`.
1432
        Returns:
1433
            A list of `ScoringRequestOutput` objects containing the
1434
1435
            generated scores in the same order as the input prompts.
        """
1436

1437
        if self.runner_type != "pooling":
1438
1439
1440
            raise ValueError(
                "LLM.score() is only supported for pooling models. "
                "Try passing `--runner pooling` to use the model as a "
1441
1442
                "pooling model."
            )
1443

1444
        score_type = self.model_config.score_type
1445
1446
1447
        if (
            score_type == "cross-encoder"
            and getattr(self.model_config.hf_config, "num_labels", 0) != 1
1448
        ):
1449
            raise ValueError("Scoring API is only enabled for num_labels == 1.")
1450

1451
1452
        if score_type is None or score_type not in self.pooling_io_processors:
            raise ValueError("This model does not support the Scoring API.")
1453

1454
1455
        io_processor = self.pooling_io_processors[score_type]
        assert isinstance(io_processor, ScoringIOProcessor)
1456

1457
1458
        pooling_task = io_processor.pooling_task
        scoring_data = io_processor.valid_inputs(data_1, data_2)
1459
        n_queries = len(scoring_data.data_1)
1460

1461
1462
1463
1464
1465
        ctx = OfflineInputsContext(
            prompts=scoring_data,
            pooling_params=pooling_params,
            tokenization_kwargs=tokenization_kwargs,
            chat_template=chat_template,
1466
            n_queries=n_queries,
1467
        )
1468

1469
1470
1471
1472
        processor_inputs = io_processor.pre_process_offline(ctx)

        seq_lora_requests = self._lora_request_to_seq(
            lora_request, len(processor_inputs)
1473
        )
1474

1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
        if ctx.pooling_params is None:
            ctx.pooling_params = PoolingParams()
        params_seq = self._params_to_seq(ctx.pooling_params, len(processor_inputs))

        for param in params_seq:
            if param.task is None:
                param.task = pooling_task
            elif param.task != pooling_task:
                msg = f"You cannot overwrite {param.task=!r} with {pooling_task=!r}!"
                raise ValueError(msg)

        seq_priority = self._priority_to_seq(None, len(processor_inputs))

        self._render_and_add_requests(
            prompts=processor_inputs,
            params=params_seq,
            lora_requests=seq_lora_requests,
            priorities=seq_priority,
        )

        outputs = self._run_engine(use_tqdm=use_tqdm, output_type=PoolingRequestOutput)
        outputs = io_processor.post_process_offline(
1497
            ctx=OfflineOutputsContext(outputs=outputs, n_queries=n_queries),
1498
1499
1500
        )

        return [ScoringRequestOutput.from_base(item) for item in outputs]
1501

1502
1503
1504
1505
1506
1507
1508
1509
1510
    def start_profile(self, profile_prefix: str | None = None) -> None:
        """Start profiling with optional custom trace prefix.

        Args:
            profile_prefix: Optional prefix for the trace file names. If provided,
                           trace files will be named as "<prefix>_dp<X>_pp<Y>_tp<Z>".
                           If not provided, default naming will be used.
        """
        self.llm_engine.start_profile(profile_prefix)
1511
1512
1513
1514

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

1515
1516
1517
1518
1519
1520
    def reset_prefix_cache(
        self, reset_running_requests: bool = False, reset_connector: bool = False
    ) -> bool:
        return self.llm_engine.reset_prefix_cache(
            reset_running_requests, reset_connector
        )
1521

1522
    def sleep(self, level: int = 1, mode: PauseMode = "abort"):
1523
1524
1525
1526
1527
        """
        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.

1528
        Args:
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
            level: The sleep level.
                - Level 0: Pause scheduling but continue accepting requests.
                           Requests are queued but not processed.
                - Level 1: Offload model weights to CPU, discard KV cache.
                           The content of kv cache is forgotten. Good for
                           sleeping and waking up the engine to run the same
                           model again. Please make sure there's enough CPU
                           memory to store the model weights.
                - Level 2: Discard all GPU memory (weights + KV cache).
                           Good for sleeping and waking up the engine to run
                           a different model or update the model, where
                           previous model weights are not needed. It reduces
                           CPU memory pressure.
1542
1543
            mode: How to handle any existing requests, can be "abort", "wait",
                or "keep".
1544
        """
1545
        self.llm_engine.sleep(level=level, mode=mode)
1546

1547
    def wake_up(self, tags: list[str] | None = None):
1548
        """
1549
1550
        Wake up the engine from sleep mode. See the [sleep][vllm.LLM.sleep]
        method for more details.
1551

1552
        Args:
1553
1554
            tags: An optional list of tags to reallocate the engine memory
                for specific memory allocations. Values must be in
1555
1556
1557
1558
                `("weights", "kv_cache", "scheduling")`. If None, all memory
                is reallocated. wake_up should be called with all tags
                (or None) before the engine is used again.
                Use tags=["scheduling"] to resume from level 0 sleep.
1559
1560
        """
        self.llm_engine.wake_up(tags)
1561

1562
1563
1564
1565
    def get_metrics(self) -> list["Metric"]:
        """Return a snapshot of aggregated metrics from Prometheus.

        Returns:
1566
            A `MetricSnapshot` instance capturing the current state
1567
1568
1569
1570
1571
1572
1573
            of all aggregated metrics from Prometheus.

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

1574
    def _params_to_seq(
1575
        self,
1576
        params: _P | Sequence[_P],
1577
        num_requests: int,
1578
    ) -> Sequence[_P]:
1579
1580
1581
1582
        if isinstance(params, Sequence):
            if len(params) != num_requests:
                raise ValueError(
                    f"The lengths of prompts ({params}) "
1583
                    f"and params ({len(params)}) must be the same."
1584
1585
                )

1586
            return params
1587

1588
1589
1590
1591
1592
1593
1594
        return [params] * num_requests

    def _lora_request_to_seq(
        self,
        lora_request: LoRARequest | None | Sequence[LoRARequest | None],
        num_requests: int,
    ) -> Sequence[LoRARequest | None]:
1595
1596
1597
1598
1599
1600
1601
        if isinstance(lora_request, Sequence):
            if len(lora_request) != num_requests:
                raise ValueError(
                    f"The lengths of prompts ({num_requests}) "
                    f"and lora_request ({len(lora_request)}) must be the same."
                )

1602
1603
1604
            return lora_request

        return [lora_request] * num_requests
1605

1606
1607
1608
1609
1610
    def _priority_to_seq(
        self,
        priority: list[int] | None,
        num_requests: int,
    ) -> Sequence[int]:
1611
1612
1613
1614
1615
1616
1617
        if priority is not None:
            if len(priority) != num_requests:
                raise ValueError(
                    f"The lengths of prompts ({num_requests}) "
                    f"and priority ({len(priority)}) must be the same."
                )

1618
1619
1620
1621
            return priority

        return [0] * num_requests

1622
    def _add_completion_requests(
1623
1624
1625
1626
1627
1628
1629
        self,
        prompts: PromptType | Sequence[PromptType],
        params: SamplingParams
        | PoolingParams
        | Sequence[SamplingParams | PoolingParams],
        *,
        use_tqdm: bool | Callable[..., tqdm] = True,
1630
        lora_request: Sequence[LoRARequest] | LoRARequest | None = None,
1631
1632
        priority: list[int] | None = None,
        tokenization_kwargs: dict[str, Any] | None = None,
1633
    ) -> list[str]:
1634
1635
        seq_prompts = prompt_to_seq(prompts)
        seq_params = self._params_to_seq(params, len(seq_prompts))
1636
1637
1638
        seq_lora_requests = self._lora_request_to_seq(lora_request, len(seq_prompts))
        seq_priority = self._priority_to_seq(priority, len(prompts))

1639
        return self._render_and_add_requests(
1640
            prompts=(
1641
1642
1643
1644
1645
                self._preprocess_cmpl_one(prompt, tokenization_kwargs)
                for prompt in maybe_tqdm(
                    seq_prompts,
                    use_tqdm=use_tqdm,
                    desc="Rendering prompts",
1646
                )
1647
            ),
1648
            params=seq_params,
1649
1650
            lora_requests=seq_lora_requests,
            priorities=seq_priority,
1651
1652
        )

1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
    def _run_completion(
        self,
        prompts: PromptType | Sequence[PromptType],
        params: SamplingParams
        | PoolingParams
        | Sequence[SamplingParams | PoolingParams],
        output_type: type[_O],
        *,
        use_tqdm: bool | Callable[..., tqdm] = True,
        lora_request: Sequence[LoRARequest] | LoRARequest | None = None,
        priority: list[int] | None = None,
        tokenization_kwargs: dict[str, Any] | None = None,
    ):
        self._add_completion_requests(
            prompts=prompts,
            params=params,
            use_tqdm=use_tqdm,
            lora_request=lora_request,
            priority=priority,
            tokenization_kwargs=tokenization_kwargs,
        )
        return self._run_engine(use_tqdm=use_tqdm, output_type=output_type)

1676
1677
1678
1679
1680
1681
1682
    def _run_chat(
        self,
        messages: list[ChatCompletionMessageParam]
        | Sequence[list[ChatCompletionMessageParam]],
        params: SamplingParams
        | PoolingParams
        | Sequence[SamplingParams | PoolingParams],
1683
        output_type: type[_O],
1684
1685
        *,
        use_tqdm: bool | Callable[..., tqdm] = True,
1686
        lora_request: Sequence[LoRARequest] | LoRARequest | None = None,
1687
1688
1689
1690
1691
1692
1693
1694
1695
        chat_template: str | None = None,
        chat_template_content_format: ChatTemplateContentFormatOption = "auto",
        add_generation_prompt: bool = True,
        continue_final_message: bool = False,
        tools: list[dict[str, Any]] | None = None,
        chat_template_kwargs: dict[str, Any] | None = None,
        tokenization_kwargs: dict[str, Any] | None = None,
        mm_processor_kwargs: dict[str, Any] | None = None,
    ):
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
        seq_convs = conversation_to_seq(messages)
        seq_params = self._params_to_seq(params, len(seq_convs))
        seq_lora_requests = self._lora_request_to_seq(lora_request, len(seq_convs))

        return self._render_and_run_requests(
            prompts=(
                self._preprocess_chat_one(
                    conversation,
                    chat_template=chat_template,
                    chat_template_content_format=chat_template_content_format,
                    chat_template_kwargs=chat_template_kwargs,
                    add_generation_prompt=add_generation_prompt,
                    continue_final_message=continue_final_message,
                    tools=tools,
1710
                    tokenization_kwargs=tokenization_kwargs,
1711
1712
                    mm_processor_kwargs=mm_processor_kwargs,
                )
1713
1714
1715
1716
                for conversation in maybe_tqdm(
                    seq_convs,
                    use_tqdm=use_tqdm,
                    desc="Rendering conversations",
1717
1718
1719
                )
            ),
            params=seq_params,
1720
            output_type=output_type,
1721
1722
            lora_requests=seq_lora_requests,
            use_tqdm=use_tqdm,
1723
1724
        )

1725
1726
    def _render_and_run_requests(
        self,
1727
        prompts: Iterable[EngineInput],
1728
        params: Sequence[SamplingParams | PoolingParams],
1729
        output_type: type[_O],
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
        *,
        lora_requests: Sequence[LoRARequest | None] | None = None,
        priorities: Sequence[int] | None = None,
        use_tqdm: bool | Callable[..., tqdm] = True,
    ):
        if isinstance(prompts, (list, tuple)):
            logger.warning_once(
                "Rendering all prompts before adding them to the engine "
                "is less efficient than performing both on the same prompt "
                "before processing the next prompt. You should instead pass "
                "a generator that renders one prompt per iteration, as that allows "
                "engine execution to begin for the first prompt while processing "
                "the next prompt."
            )

        self._render_and_add_requests(
            prompts=prompts,
1747
            params=params,
1748
1749
            lora_requests=lora_requests,
            priorities=priorities,
1750
1751
        )

1752
        return self._run_engine(output_type, use_tqdm=use_tqdm)
1753

1754
    def _render_and_add_requests(
1755
        self,
1756
        prompts: Iterable[EngineInput],
1757
        params: Sequence[SamplingParams | PoolingParams],
1758
        *,
1759
1760
        lora_requests: Sequence[LoRARequest | None] | None = None,
        priorities: Sequence[int] | None = None,
1761
    ) -> list[str]:
1762
        added_request_ids: list[str] = []
1763

1764
        try:
1765
            for i, prompt in enumerate(prompts):
1766
1767
                request_id = self._add_request(
                    prompt,
1768
                    params[i],
Cyrus Leung's avatar
Cyrus Leung committed
1769
1770
1771
1772
                    lora_request=self._resolve_mm_lora(
                        prompt,
                        None if lora_requests is None else lora_requests[i],
                    ),
1773
                    priority=0 if priorities is None else priorities[i],
1774
1775
1776
1777
                )
                added_request_ids.append(request_id)
        except Exception as e:
            if added_request_ids:
1778
                self.llm_engine.abort_request(added_request_ids, internal=True)
1779
            raise e
1780

1781
1782
        return added_request_ids

1783
    def _add_request(
nunjunj's avatar
nunjunj committed
1784
        self,
1785
        prompt: EngineInput,
1786
1787
        params: SamplingParams | PoolingParams,
        lora_request: LoRARequest | None = None,
1788
        priority: int = 0,
1789
    ) -> str:
1790
1791
1792
1793
        if isinstance(params, SamplingParams):
            # We only care about the final output
            params.output_kind = RequestOutputKind.FINAL_ONLY

1794
        request_id = str(next(self.request_counter))
1795

1796
        return self.llm_engine.add_request(
1797
            request_id,
1798
            prompt,
1799
1800
            params,
            lora_request=lora_request,
1801
            priority=priority,
nunjunj's avatar
nunjunj committed
1802
        )
1803

1804
    def _run_engine(
1805
        self,
1806
        output_type: type[_O] | tuple[type[_O], ...],
1807
1808
        *,
        use_tqdm: bool | Callable[..., tqdm] = True,
1809
    ) -> list[_O]:
1810
1811
        # Initialize tqdm.
        if use_tqdm:
Zhuohan Li's avatar
Zhuohan Li committed
1812
            num_requests = self.llm_engine.get_num_unfinished_requests()
1813
1814
            tqdm_func = use_tqdm if callable(use_tqdm) else tqdm
            pbar = tqdm_func(
1815
1816
1817
                total=num_requests,
                desc="Processed prompts",
                dynamic_ncols=True,
1818
                postfix=(f"est. speed input: {0:.2f} toks/s, output: {0:.2f} toks/s"),
1819
            )
1820

Zhuohan Li's avatar
Zhuohan Li committed
1821
        # Run the engine.
1822
        outputs: list[_O] = []
1823
1824
        total_in_toks = 0
        total_out_toks = 0
Zhuohan Li's avatar
Zhuohan Li committed
1825
1826
        while self.llm_engine.has_unfinished_requests():
            step_outputs = self.llm_engine.step()
1827
            for output in step_outputs:
1828
                assert isinstance(output, output_type)
1829
                if output.finished:
1830
                    outputs.append(output)  # type: ignore[arg-type]
1831
                    if use_tqdm:
1832
1833
                        if isinstance(output, RequestOutput):
                            # Calculate tokens only for RequestOutput
1834
                            n = len(output.outputs)
1835
                            assert output.prompt_token_ids is not None
1836
                            total_in_toks += len(output.prompt_token_ids) * n
1837
1838
                            in_spd = total_in_toks / pbar.format_dict["elapsed"]
                            total_out_toks += sum(
1839
1840
1841
                                len(stp.token_ids) for stp in output.outputs
                            )
                            out_spd = total_out_toks / pbar.format_dict["elapsed"]
1842
1843
                            pbar.postfix = (
                                f"est. speed input: {in_spd:.2f} toks/s, "
1844
1845
                                f"output: {out_spd:.2f} toks/s"
                            )
1846
                            pbar.update(n)
1847
1848
                        else:
                            pbar.update(1)
1849
1850
                        if pbar.n == num_requests:
                            pbar.refresh()
1851

1852
1853
        if use_tqdm:
            pbar.close()
1854
1855
1856
        # Sort the outputs by request ID.
        # This is necessary because some requests may be finished earlier than
        # its previous requests.
1857
        return sorted(outputs, key=lambda x: int(x.request_id))
1858

1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
    def init_weight_transfer_engine(
        self, request: WeightTransferInitRequest | dict
    ) -> None:
        """
        Initialize weight transfer for RL training.

        Args:
            request: Weight transfer initialization request with backend-specific info
        """
        init_info_dict = (
            request["init_info"] if isinstance(request, dict) else request.init_info
        )

        self.llm_engine.collective_rpc(
            "init_weight_transfer_engine", kwargs={"init_info": init_info_dict}
        )

    def update_weights(self, request: WeightTransferUpdateRequest | dict) -> None:
        """
        Update the weights of the model.

        Args:
            request: Weight update request with backend-specific update info
        """
        update_info_dict = (
            request["update_info"] if isinstance(request, dict) else request.update_info
        )

        self.llm_engine.collective_rpc(
            "update_weights", kwargs={"update_info": update_info_dict}
        )

1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
    def __repr__(self) -> str:
        """Return a transformers-style hierarchical view of the model."""
        # Cache the result to avoid repeated collective_rpc calls
        if self._cached_repr is None:
            results = self.llm_engine.collective_rpc("get_model_inspection")
            # In distributed settings, we get results from all workers
            # Just return the first one (they should all be the same)
            if results:
                self._cached_repr = results[0]
            else:
                self._cached_repr = f"LLM(model={self.model_config.model!r})"
        return self._cached_repr