llm.py 89.1 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import itertools
chenzk's avatar
chenzk committed
5
import os
chenzk's avatar
chenzk committed
6
from copy import deepcopy
7
from collections.abc import Callable, Iterable, Sequence
8
from pathlib import Path
9
from typing import TYPE_CHECKING, Any
10

11
import cloudpickle
12
import torch.nn as nn
13
from pydantic import ValidationError
14
from tqdm.auto import tqdm
15
from typing_extensions import TypeVar, overload
16

17
18
19
20
21
22
23
from vllm.beam_search import (
    BeamSearchInstance,
    BeamSearchOutput,
    BeamSearchSequence,
    create_sort_beams_key_function,
)
from vllm.config import (
24
    AttentionConfig,
25
    CompilationConfig,
26
    PoolerConfig,
27
    ProfilerConfig,
28
29
30
    StructuredOutputsConfig,
    is_init_field,
)
31
from vllm.config.compilation import CompilationMode
32
from vllm.config.model import (
33
34
    ConvertOption,
    HfOverrides,
35
    ModelDType,
36
    RunnerOption,
37
    TokenizerMode,
38
)
39
40
41
42
from vllm.distributed.weight_transfer.base import (
    WeightTransferInitRequest,
    WeightTransferUpdateRequest,
)
43
from vllm.engine.arg_utils import EngineArgs
44
45
from vllm.entrypoints.chat_utils import (
    ChatCompletionMessageParam,
46
    ChatTemplateConfig,
47
    ChatTemplateContentFormatOption,
48
    load_chat_template,
49
)
50
from vllm.entrypoints.pooling.io_processor_factories import init_pooling_io_processors
51
from vllm.entrypoints.pooling.score.utils import (
52
    ScoreData,
53
54
55
    ScoreMultiModalParam,
    _cosine_similarity,
    compress_token_type_ids,
56
    compute_maxsim_score,
57
    get_score_prompt,
58
    score_data_to_prompts,
59
    validate_score_input,
60
)
61
from vllm.entrypoints.utils import log_non_default_args
62
from vllm.inputs.data import (
63
    DataPrompt,
64
    ProcessorInputs,
65
66
67
68
69
    PromptType,
    SingletonPrompt,
    TextPrompt,
    TokensPrompt,
)
70
from vllm.logger import init_logger
71
from vllm.lora.request import LoRARequest
72
from vllm.model_executor.layers.quantization import QuantizationMethods
73
74
75
76
77
78
79
from vllm.outputs import (
    ClassificationRequestOutput,
    EmbeddingRequestOutput,
    PoolingRequestOutput,
    RequestOutput,
    ScoringRequestOutput,
)
80
from vllm.platforms import current_platform
81
from vllm.pooling_params import PoolingParams
82
from vllm.renderers import ChatParams, merge_kwargs
83
84
85
86
87
from vllm.renderers.inputs.preprocess import (
    conversation_to_seq,
    parse_model_prompt,
    prompt_to_seq,
)
88
from vllm.sampling_params import BeamSearchParams, RequestOutputKind, SamplingParams
89
from vllm.tasks import PoolingTask
90
from vllm.tokenizers import TokenizerLike
yhu422's avatar
yhu422 committed
91
from vllm.usage.usage_lib import UsageContext
92
from vllm.utils.counter import Counter
93
from vllm.utils.mistral import is_mistral_tokenizer
94
from vllm.utils.tqdm_utils import maybe_tqdm
95
from vllm.v1.engine import PauseMode
96
from vllm.v1.engine.llm_engine import LLMEngine
97
from vllm.v1.sample.logits_processor import LogitsProcessor
98

99
if TYPE_CHECKING:
chenzk's avatar
chenzk committed
100
    from vllm.kvprune.integration.compression_params import CompressionParams
101
102
    from vllm.v1.metrics.reader import Metric

103
104
logger = init_logger(__name__)

105
106
107
108
109
_O = TypeVar(
    "_O",
    bound=RequestOutput | PoolingRequestOutput,
    default=RequestOutput | PoolingRequestOutput,
)
110
_P = TypeVar("_P", bound=SamplingParams | PoolingParams | None)
111
112
_R = TypeVar("_R", default=Any)

113
114

class LLM:
Woosuk Kwon's avatar
Woosuk Kwon committed
115
116
117
118
119
120
121
122
123
124
    """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.
125
        tokenizer: The name or path of a HuggingFace Transformers tokenizer.
126
127
        tokenizer_mode: The tokenizer mode. "auto" will use the fast tokenizer
            if available, and "slow" will always use the slow tokenizer.
128
129
130
        skip_tokenizer_init: If true, skip initialization of tokenizer and
            detokenizer. Expect valid prompt_token_ids and None for prompt
            from the input.
131
132
        trust_remote_code: Trust remote code (e.g., from HuggingFace) when
            downloading the model and tokenizer.
133
134
135
136
        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.
137
        allowed_media_domains: If set, only media URLs that belong to this
138
            domain can be used for multi-modal inputs.
Woosuk Kwon's avatar
Woosuk Kwon committed
139
140
141
        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
142
            we support `float32`, `float16`, and `bfloat16`. If `auto`, we use
143
144
            the `dtype` attribute of the Transformers model's config. However,
            if the `dtype` in the config is `float32`, we will use `float16` instead.
145
        quantization: The method used to quantize the model weights. Currently,
146
            we support "awq", "gptq", and "fp8" (experimental).
147
148
149
150
            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
151
152
        revision: The specific model version to use. It can be a branch name,
            a tag name, or a commit id.
153
154
        tokenizer_revision: The specific tokenizer version to use. It can be a
            branch name, a tag name, or a commit id.
155
        chat_template: The chat template to apply.
156
157
158
159
160
161
        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.
162
163
164
165
166
        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
167
            compared with using gpu_memory_utilization. Note that
168
169
            kv_cache_memory_bytes (when not-None) ignores
            gpu_memory_utilization
170
171
172
173
        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.
174
175
176
177
178
179
180
181
182
183
184
185
186
        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.
187
188
189
        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.
chenzk's avatar
chenzk committed
190
191
192
193
194
195
196
        kvprune_compression: Compatibility flag for the integrated kvprune path.
            If ``None`` (default), read ``VLLM_KVPRUNE_COMPRESSION_DEFAULT``.
            When enabled, requests with ``compression_ratio < 1.0`` automatically
            rebuild the internal v1 engine into a kvprune-friendly mode
            (``enforce_eager=True`` and ``num_gpu_blocks_override=1``). Requests
            with ``compression_ratio >= 1.0`` use the caller-provided normal v1
            engine configuration.
197
        enable_return_routed_experts: Whether to return routed experts.
198
199
        disable_custom_all_reduce: See
            [ParallelConfig][vllm.config.ParallelConfig].
200
        hf_token: The token to use as HTTP bearer authorization for remote files
201
            . If `True`, will use the token generated when running
202
            `hf auth login` (stored in `~/.cache/huggingface/token`).
203
204
205
        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.
206
207
208
209
210
        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}`.
211
212
        pooler_config: Initialize non-default pooling config for the pooling model,
            e.g., `PoolerConfig(seq_pooling_type="MEAN", use_activation=False)`.
213
        compilation_config: Either an integer or a dictionary. If it is an
214
            integer, it is used as the mode of compilation optimization. If it
215
            is a dictionary, it can specify the full compilation configuration.
216
217
218
219
        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.
220
        **kwargs: Arguments for [`EngineArgs`][vllm.EngineArgs].
nunjunj's avatar
nunjunj committed
221

222
223
    Note:
        This class is intended to be used for offline inference. For online
224
        serving, use the [AsyncLLMEngine][vllm.AsyncLLMEngine] class instead.
225
    """
226
227
228
229

    def __init__(
        self,
        model: str,
230
        *,
231
232
        runner: RunnerOption = "auto",
        convert: ConvertOption = "auto",
233
        tokenizer: str | None = None,
234
        tokenizer_mode: TokenizerMode | str = "auto",
235
        skip_tokenizer_init: bool = False,
236
        trust_remote_code: bool = False,
237
        allowed_local_media_path: str = "",
238
        allowed_media_domains: list[str] | None = None,
239
        tensor_parallel_size: int = 1,
240
        dtype: ModelDType = "auto",
241
242
243
        quantization: QuantizationMethods | None = None,
        revision: str | None = None,
        tokenizer_revision: str | None = None,
244
        chat_template: Path | str | None = None,
245
        seed: int = 0,
246
        gpu_memory_utilization: float = 0.9,
247
        cpu_offload_gb: float = 0,
248
249
250
251
        offload_group_size: int = 0,
        offload_num_in_group: int = 1,
        offload_prefetch_step: int = 1,
        offload_params: set[str] | None = None,
252
        enforce_eager: bool = False,
chenzk's avatar
chenzk committed
253
        kvprune_compression: bool | None = None,
254
        enable_return_routed_experts: bool = False,
255
        disable_custom_all_reduce: bool = False,
256
257
258
259
260
261
262
        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,
263
        profiler_config: dict[str, Any] | ProfilerConfig | None = None,
264
        attention_config: dict[str, Any] | AttentionConfig | None = None,
265
266
267
        kv_cache_memory_bytes: int | None = None,
        compilation_config: int | dict[str, Any] | CompilationConfig | None = None,
        logits_processors: list[str | type[LogitsProcessor]] | None = None,
268
        **kwargs: Any,
269
    ) -> None:
270
        """LLM constructor."""
271

272
273
274
275
276
277
278
279
280
281
282
        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,
            )

283
284
        if "disable_log_stats" not in kwargs:
            kwargs["disable_log_stats"] = True
285

286
287
288
289
290
291
292
        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)

293
        if "kv_transfer_config" in kwargs and isinstance(
294
295
            kwargs["kv_transfer_config"], dict
        ):
296
            from vllm.config.kv_transfer import KVTransferConfig
297

298
299
            raw_config_dict = kwargs["kv_transfer_config"]
            try:
300
                kwargs["kv_transfer_config"] = KVTransferConfig(**raw_config_dict)
301
302
303
304
            except ValidationError as e:
                logger.error(
                    "Failed to convert 'kv_transfer_config' dict to "
                    "KVTransferConfig object. Dict: %s. Error: %s",
305
306
307
                    raw_config_dict,
                    e,
                )
308
309
                # Consider re-raising a more specific vLLM error or ValueError
                # to provide better context to the user.
310
                raise ValueError(f"Invalid 'kv_transfer_config' provided: {e}") from e
311

312
313
314
        if hf_overrides is None:
            hf_overrides = {}

315
316
317
318
319
320
321
        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
322

323
324
325
326
        if isinstance(compilation_config, int):
            compilation_config_instance = CompilationConfig(
                mode=CompilationMode(compilation_config)
            )
327
        else:
328
329
330
            compilation_config_instance = _make_config(
                compilation_config, CompilationConfig
            )
331

332
333
334
335
336
        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)
337

338
        # warn about single-process data parallel usage.
339
340
        _dp_size = int(kwargs.get("data_parallel_size", 1))
        _distributed_executor_backend = kwargs.get("distributed_executor_backend")
341
342
343
344
345
        if (
            _dp_size > 1
            and not _distributed_executor_backend == "external_launcher"
            and not current_platform.is_tpu()
        ):
346
            raise ValueError(
347
                f"LLM(data_parallel_size={_dp_size}) is not supported for single-"
348
349
350
351
352
                "process usage and may hang. Please use "
                "the explicit multi-process data-parallel example at "
                "'examples/offline_inference/data_parallel.py'."
            )

chenzk's avatar
chenzk committed
353
354
355
        # v1 ``enforce_eager`` is independent of kvprune compactor
        # ``LLMConfig.enforce_eager``. ``kvprune_compression`` enables automatic
        # switching between normal and compressed v1 engine init modes per request.
chenzk's avatar
chenzk committed
356
357
358
        if kvprune_compression is None:
            _kvd = os.environ.get("VLLM_KVPRUNE_COMPRESSION_DEFAULT", "0").strip().lower()
            kvprune_compression = _kvd in ("1", "true", "yes")
Zhuohan Li's avatar
Zhuohan Li committed
359
        engine_args = EngineArgs(
360
            model=model,
361
362
            runner=runner,
            convert=convert,
363
            tokenizer=tokenizer,
364
            tokenizer_mode=tokenizer_mode,
365
            skip_tokenizer_init=skip_tokenizer_init,
366
            trust_remote_code=trust_remote_code,
367
            allowed_local_media_path=allowed_local_media_path,
368
            allowed_media_domains=allowed_media_domains,
369
370
            tensor_parallel_size=tensor_parallel_size,
            dtype=dtype,
371
            quantization=quantization,
372
            revision=revision,
373
            tokenizer_revision=tokenizer_revision,
374
375
            seed=seed,
            gpu_memory_utilization=gpu_memory_utilization,
376
            kv_cache_memory_bytes=kv_cache_memory_bytes,
377
            cpu_offload_gb=cpu_offload_gb,
378
379
380
381
            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(),
382
            enforce_eager=enforce_eager,
383
            enable_return_routed_experts=enable_return_routed_experts,
384
            disable_custom_all_reduce=disable_custom_all_reduce,
385
            hf_token=hf_token,
386
            hf_overrides=hf_overrides,
387
            mm_processor_kwargs=mm_processor_kwargs,
388
            pooler_config=pooler_config,
389
            structured_outputs_config=structured_outputs_instance,
390
            profiler_config=profiler_config_instance,
391
            attention_config=attention_config_instance,
392
            compilation_config=compilation_config_instance,
393
            logits_processors=logits_processors,
394
395
            **kwargs,
        )
396

397
398
        log_non_default_args(engine_args)

399
        self.request_counter = Counter()
400
        self.default_sampling_params: dict[str, Any] | None = None
chenzk's avatar
chenzk committed
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
        # Cache for __repr__ to avoid repeated collective_rpc calls
        self._cached_repr: str | None = None
        # Lazy compactor engine (``vllm.kvprune``) when :meth:`generate` uses compression.
        self._kvprune_compactor_engine: Any = None
        self._kvprune_compression_enabled = bool(kvprune_compression)
        self._engine_args_base = deepcopy(engine_args)
        self._kvprune_v1_mode = "normal"
        self.chat_template = load_chat_template(chat_template)
        self.chat_template_config = ChatTemplateConfig(chat_template=self.chat_template)
        if not self._kvprune_compression_enabled:
            self._rebuild_llm_engine_for_kvprune_mode("normal")

    def _ensure_llm_engine_initialized(self, mode: str = "normal") -> None:
        if not hasattr(self, "llm_engine"):
            self._rebuild_llm_engine_for_kvprune_mode(mode)

    def _shutdown_llm_engine(self) -> None:
        old_engine = getattr(self, "llm_engine", None)
        if old_engine is None:
            return
        try:
            old_engine.engine_core.shutdown()
        except Exception:
            logger.warning("Failed to shutdown previous LLMEngine cleanly.", exc_info=True)
        try:
            dp_group = getattr(old_engine, "dp_group", None)
            if dp_group is not None and not getattr(old_engine, "external_launcher_dp", False):
                from vllm.distributed import (
                    stateless_destroy_torch_distributed_process_group,
                )

                stateless_destroy_torch_distributed_process_group(dp_group)
        except Exception:
            logger.warning(
                "Failed to destroy previous LLMEngine DP group cleanly.",
                exc_info=True,
            )

    def _attach_llm_engine(self, llm_engine: LLMEngine) -> None:
        self.llm_engine = llm_engine
        self.engine_class = type(self.llm_engine)
        self.default_sampling_params = None
        self._cached_repr = None
        self._kvprune_compactor_engine = None
445

446
447
        supported_tasks = self.llm_engine.get_supported_tasks()
        logger.info("Supported tasks: %s", supported_tasks)
448
449
        self.supported_tasks = supported_tasks

450
        self.model_config = self.llm_engine.model_config
451
        self.renderer = self.llm_engine.renderer
452
        self.io_processor = self.llm_engine.io_processor
453
        self.input_processor = self.llm_engine.input_processor
454
        self.pooling_io_processors = init_pooling_io_processors(
455
456
457
458
459
            supported_tasks=supported_tasks,
            model_config=self.model_config,
            renderer=self.renderer,
            chat_template_config=self.chat_template_config,
        )
chenzk's avatar
chenzk committed
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486

    def _rebuild_llm_engine_for_kvprune_mode(self, mode: str) -> None:
        if mode not in ("normal", "compressed"):
            raise ValueError(f"Unknown kvprune v1 mode: {mode!r}")
        if getattr(self, "_kvprune_v1_mode", None) == mode and hasattr(self, "llm_engine"):
            return

        engine_args = deepcopy(self._engine_args_base)
        if mode == "compressed":
            engine_args.enforce_eager = True
            engine_args.num_gpu_blocks_override = 1

        self._shutdown_llm_engine()
        llm_engine = LLMEngine.from_engine_args(
            engine_args=engine_args, usage_context=UsageContext.LLM_CLASS
        )
        self._attach_llm_engine(llm_engine)
        self._kvprune_v1_mode = mode

    def _compression_needs_kvprune(self, compression: Any) -> bool:
        if compression is None:
            return False
        from vllm.kvprune.integration.compression_params import CompressionParams

        if isinstance(compression, CompressionParams):
            return compression.compression_ratio < 1.0
        return any(cp.compression_ratio < 1.0 for cp in compression)
487

488
    def get_tokenizer(self) -> TokenizerLike:
chenzk's avatar
chenzk committed
489
        self._ensure_llm_engine_initialized("normal")
490
        return self.llm_engine.get_tokenizer()
491

492
493
494
495
496
497
498
499
500
501
502
503
    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.
        """
chenzk's avatar
chenzk committed
504
        self._ensure_llm_engine_initialized("normal")
505
506
507
508
509
        parallel_config = self.llm_engine.vllm_config.parallel_config
        if include_dp:
            return parallel_config.world_size_across_dp
        return parallel_config.world_size

510
    def reset_mm_cache(self) -> None:
chenzk's avatar
chenzk committed
511
        self._ensure_llm_engine_initialized("normal")
512
        self.renderer.clear_mm_cache()
513
514
        self.llm_engine.reset_mm_cache()

515
    def get_default_sampling_params(self) -> SamplingParams:
chenzk's avatar
chenzk committed
516
        self._ensure_llm_engine_initialized("normal")
517
        if self.default_sampling_params is None:
518
            self.default_sampling_params = self.model_config.get_diff_sampling_param()
519
520
        if self.default_sampling_params:
            return SamplingParams.from_optional(**self.default_sampling_params)
521
522
        return SamplingParams()

523
524
    def generate(
        self,
525
526
        prompts: PromptType | Sequence[PromptType],
        sampling_params: SamplingParams | Sequence[SamplingParams] | None = None,
527
        *,
528
        use_tqdm: bool | Callable[..., tqdm] = True,
529
        lora_request: Sequence[LoRARequest] | LoRARequest | None = None,
530
        priority: list[int] | None = None,
531
        tokenization_kwargs: dict[str, Any] | None = None,
chenzk's avatar
chenzk committed
532
        compression: "CompressionParams | Sequence[CompressionParams] | None" = None,
533
    ) -> list[RequestOutput]:
Woosuk Kwon's avatar
Woosuk Kwon committed
534
535
        """Generates the completions for the input prompts.

536
        This class automatically batches the given prompts, considering
Woosuk Kwon's avatar
Woosuk Kwon committed
537
538
539
540
        the memory constraint. For the best performance, put all of your prompts
        into a single list and pass it to this method.

        Args:
541
            prompts: The prompts to the LLM. You may pass a sequence of prompts
542
                for batch inference. See [PromptType][vllm.inputs.PromptType]
543
                for more details about the format of each prompt.
Woosuk Kwon's avatar
Woosuk Kwon committed
544
            sampling_params: The sampling parameters for text generation. If
nunjunj's avatar
nunjunj committed
545
546
547
                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
548
                prompts and it is paired one by one with the prompt.
549
550
551
552
            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.
553
            lora_request: LoRA request to use for generation, if any.
554
555
            priority: The priority of the requests, if any.
                Only applicable when priority scheduling policy is enabled.
556
557
558
                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.
559
            tokenization_kwargs: Overrides for `tokenizer.encode`.
chenzk's avatar
chenzk committed
560
561
562
563
            compression: Optional per-prompt KV compression (``vllm.kvprune``). If any
                prompt has ``compression_ratio < 1.0``, the batch is run on the integrated
                compactor engine with weights shared from this ``LLM``. Omit or use all
                ``compression_ratio >= 1`` to use the standard v1 engine only.
chenzk's avatar
chenzk committed
564
565
566
567
568
569
570
571
                If ``kvprune_compression=True``, requests with
                ``compression_ratio < 1.0`` automatically rebuild the internal v1
                engine into eager + 1-block mode before entering kvprune.
                Compactor decode graphs default on
                (``VLLM_KVPRUNE_COMPACTOR_CUDA_GRAPH`` default ``1``) with eager
                fallback if capture fails; set
                ``VLLM_KVPRUNE_COMPACTOR_ENFORCE_EAGER=1`` to skip compactor graph
                capture entirely.
Woosuk Kwon's avatar
Woosuk Kwon committed
572
573

        Returns:
574
            A list of `RequestOutput` objects containing the
575
576
            generated completions in the same order as the input prompts.
        """
chenzk's avatar
chenzk committed
577
578
579
580
581
582
583
584
585
586
587
588
        compression_eff = compression

        if self._kvprune_compression_enabled:
            target_mode = (
                "compressed"
                if self._compression_needs_kvprune(compression_eff)
                else "normal"
            )
            self._rebuild_llm_engine_for_kvprune_mode(target_mode)
        else:
            self._ensure_llm_engine_initialized("normal")

589
        runner_type = self.model_config.runner_type
590
        if runner_type != "generate":
591
592
593
            raise ValueError(
                "LLM.generate() is only supported for generative models. "
                "Try passing `--runner generate` to use the model as a "
594
595
                "generative model."
            )
chenzk's avatar
chenzk committed
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613

        if compression_eff is not None:
            from vllm.kvprune.integration.compressed_generate import (
                try_compressed_generate,
            )

            compressed_out = try_compressed_generate(
                self,
                prompts,
                sampling_params,
                compression=compression_eff,
                use_tqdm=use_tqdm,
                lora_request=lora_request,
                priority=priority,
                tokenization_kwargs=tokenization_kwargs,
            )
            if compressed_out is not None:
                return compressed_out
614

615
        if sampling_params is None:
616
            sampling_params = self.get_default_sampling_params()
617

618
        return self._run_completion(
619
            prompts=prompts,
620
            params=sampling_params,
621
            output_type=RequestOutput,
622
            use_tqdm=use_tqdm,
623
            lora_request=lora_request,
624
            tokenization_kwargs=tokenization_kwargs,
625
626
            priority=priority,
        )
627

628
629
630
631
    def enqueue(
        self,
        prompts: PromptType | Sequence[PromptType],
        sampling_params: SamplingParams | Sequence[SamplingParams] | None = None,
632
        lora_request: Sequence[LoRARequest] | LoRARequest | None = None,
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
        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.
        """
chenzk's avatar
chenzk committed
654
        self._ensure_llm_engine_initialized("normal")
655
        runner_type = self.model_config.runner_type
656
657
658
659
660
661
        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()

662
663
664
665
666
667
668
        return self._add_completion_requests(
            prompts=prompts,
            params=sampling_params,
            use_tqdm=use_tqdm,
            lora_request=lora_request,
            priority=priority,
            tokenization_kwargs=tokenization_kwargs,
669
670
        )

671
    @overload
672
673
    def wait_for_completion(
        self,
674
        *,
675
        use_tqdm: bool | Callable[..., tqdm] = True,
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
    ) -> 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]:
692
693
694
695
696
697
        """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:
698
            output_type: The expected output type, defaults to RequestOutput.
699
700
701
            use_tqdm: If True, shows a tqdm progress bar.

        Returns:
702
            A list of output objects for all completed requests.
703
        """
704
705
706
707
        if output_type is None:
            output_type = (RequestOutput, PoolingRequestOutput)

        return self._run_engine(output_type, use_tqdm=use_tqdm)
708

Cyrus Leung's avatar
Cyrus Leung committed
709
    def _resolve_mm_lora(
710
        self,
711
        prompt: ProcessorInputs,
712
        lora_request: LoRARequest | None,
Cyrus Leung's avatar
Cyrus Leung committed
713
714
715
716
717
718
719
    ) -> 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:
720
721
            return lora_request

722
723
        prompt_modalities = prompt["mm_placeholders"].keys()
        intersection = set(prompt_modalities).intersection(default_mm_loras.keys())
724
725
        if not intersection:
            return lora_request
Cyrus Leung's avatar
Cyrus Leung committed
726

727
728
729
        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
730
731
732
733
                "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",
734
735
                intersection,
            )
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
            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 "
751
752
                    "lora_request as we only apply one LoRARequest per prompt"
                )
753
754
755
756
757
758
759
760
            return lora_request

        return LoRARequest(
            modality_name,
            modality_lora_id,
            modality_lora_path,
        )

761
762
    def collective_rpc(
        self,
763
764
        method: str | Callable[..., _R],
        timeout: float | None = None,
765
        args: tuple = (),
766
        kwargs: dict[str, Any] | None = None,
767
    ) -> list[_R]:
768
769
770
771
772
773
774
775
776
777
778
        """
        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
779
                [`TimeoutError`][] on timeout. `None` means wait indefinitely.
780
781
782
783
784
            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.
785

786
787
788
        Note:
            It is recommended to use this API to only pass control messages,
            and set up data-plane communication to pass data.
789
        """
790

chenzk's avatar
chenzk committed
791
        self._ensure_llm_engine_initialized("normal")
792
        return self.llm_engine.collective_rpc(method, timeout, args, kwargs)
793
794

    def apply_model(self, func: Callable[[nn.Module], _R]) -> list[_R]:
795
        """
796
797
        Run a function directly on the model inside each worker,
        returning the result for each of them.
798
799
800
801
802
803

        !!! 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!
804
        """
chenzk's avatar
chenzk committed
805
        self._ensure_llm_engine_initialized("normal")
806
        return self.llm_engine.apply_model(func)
807

808
809
    def beam_search(
        self,
810
        prompts: list[TokensPrompt | TextPrompt],
811
        params: BeamSearchParams,
812
        lora_request: list[LoRARequest] | LoRARequest | None = None,
813
        use_tqdm: bool = False,
814
        concurrency_limit: int | None = None,
815
    ) -> list[BeamSearchOutput]:
816
817
818
819
820
821
        """
        Generate sequences using beam search.

        Args:
            prompts: A list of prompts. Each prompt can be a string or a list
                of token IDs.
822
            params: The beam search parameters.
823
            lora_request: LoRA request to use for generation, if any.
824
            use_tqdm: Whether to use tqdm to display the progress bar.
825
826
            concurrency_limit: The maximum number of concurrent requests.
                If None, the number of concurrent requests is unlimited.
827
        """
828
829
        # TODO: how does beam search work together with length penalty,
        # frequency, penalty, and stopping criteria, etc.?
830
831
832
833
        beam_width = params.beam_width
        max_tokens = params.max_tokens
        temperature = params.temperature
        ignore_eos = params.ignore_eos
834
835
        length_penalty = params.length_penalty

836
837
838
        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)
839

840
841
        engine_prompts = self._preprocess_cmpl(prompts)
        lora_requests = self._lora_request_to_seq(lora_request, len(engine_prompts))
842

843
844
845
        if use_tqdm and concurrency_limit is not None:
            logger.warning(
                "Progress bar is not supported when using concurrency_limit. "
846
847
                "Disabling progress bar."
            )
848
849
850
            use_tqdm = False

        if concurrency_limit is None:
851
            concurrency_limit = len(engine_prompts)
852

853
854
855
        # 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
856
        sampling_params = SamplingParams(
857
858
859
860
            logprobs=2 * beam_width,
            max_tokens=1,
            temperature=temperature,
            skip_clone=True,  # Internal beam search, safe to skip clone
861
        )
862
        instances: list[BeamSearchInstance] = []
863

864
865
866
867
868
        for lora_req, prompt in zip(lora_requests, engine_prompts):
            if prompt["type"] == "embeds":
                raise NotImplementedError(
                    "Embedding prompt not supported for beam search"
                )
869

870
            instances.append(
871
                BeamSearchInstance(
872
                    prompt,
873
874
                    lora_request=lora_req,
                    logprobs=None,
875
876
                ),
            )
877

878
        for prompt_start in range(0, len(instances), concurrency_limit):
879
            instances_batch = instances[prompt_start : prompt_start + concurrency_limit]
880
881
882

            token_iter = range(max_tokens)
            if use_tqdm:
883
884
885
                token_iter = tqdm(
                    token_iter, desc="Beam search", unit="token", unit_scale=False
                )
886
887
888
                logger.warning(
                    "The progress bar shows the upper bound on token steps and "
                    "may finish early due to stopping conditions. It does not "
889
890
                    "reflect instance-level progress."
                )
891
892
            for _ in token_iter:
                all_beams: list[BeamSearchSequence] = list(
893
894
                    sum((instance.beams for instance in instances_batch), [])
                )
895
896
                pos = [0] + list(
                    itertools.accumulate(
897
898
899
                        len(instance.beams) for instance in instances_batch
                    )
                )
900
                instance_start_and_end: list[tuple[int, int]] = list(
901
902
                    zip(pos[:-1], pos[1:])
                )
903
904
905
906
907
908

                if len(all_beams) == 0:
                    break

                # only runs for one step
                # we don't need to use tqdm here
909
                output = self._render_and_run_requests(
910
911
                    prompts=(beam.get_prompt() for beam in all_beams),
                    params=self._params_to_seq(sampling_params, len(all_beams)),
912
                    output_type=RequestOutput,
913
                    lora_requests=[beam.lora_request for beam in all_beams],
914
915
                    use_tqdm=False,
                )
916

917
918
919
                for (start, end), instance in zip(
                    instance_start_and_end, instances_batch
                ):
920
921
922
923
924
925
926
927
928
929
930
931
932
                    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(
933
                                    current_beam.orig_prompt,
934
                                    tokens=current_beam.tokens + [token_id],
935
                                    logprobs=current_beam.logprobs + [logprobs],
936
                                    lora_request=current_beam.lora_request,
937
938
939
940
                                    cum_logprob=current_beam.cum_logprob
                                    + logprob_obj.logprob,
                                )

941
                                if token_id == eos_token_id and not ignore_eos:
942
943
944
                                    instance.completed.append(new_beam)
                                else:
                                    instance_new_beams.append(new_beam)
945
946
947
                    sorted_beams = sorted(
                        instance_new_beams, key=sort_beams_key, reverse=True
                    )
948
                    instance.beams = sorted_beams[:beam_width]
949
950
951
952

        outputs = []
        for instance in instances:
            instance.completed.extend(instance.beams)
953
954
955
            sorted_completed = sorted(
                instance.completed, key=sort_beams_key, reverse=True
            )
956
957
958
959
            best_beams = sorted_completed[:beam_width]

            for beam in best_beams:
                beam.text = tokenizer.decode(beam.tokens)
960

961
962
963
964
            outputs.append(BeamSearchOutput(sequences=best_beams))

        return outputs

965
    def _preprocess_cmpl(
966
        self,
967
        prompts: Sequence[PromptType],
968
        tokenization_kwargs: dict[str, Any] | None = None,
969
    ) -> Sequence[ProcessorInputs]:
970
971
972
973
974
975
976
        """
        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:
977
            A list of `ProcessorInputs` objects ready to be passed into LLMEngine.
978
        """
979
        renderer = self.renderer
980
981
        model_config = self.model_config

982
983
984
        parsed_prompts = [
            parse_model_prompt(model_config, prompt) for prompt in prompts
        ]
985
986
987
        tok_params = renderer.default_cmpl_tok_params.with_kwargs(
            **(tokenization_kwargs or {})
        )
988

989
        return renderer.render_cmpl(parsed_prompts, tok_params)
990

991
992
993
994
995
996
997
998
    def _preprocess_cmpl_one(
        self,
        prompt: PromptType,
        tokenization_kwargs: dict[str, Any] | None = None,
    ) -> ProcessorInputs:
        (engine_prompt,) = self._preprocess_cmpl([prompt], tokenization_kwargs)
        return engine_prompt

999
1000
    def _preprocess_chat(
        self,
1001
        conversations: Sequence[list[ChatCompletionMessageParam]],
1002
        chat_template: str | None = None,
1003
        chat_template_content_format: ChatTemplateContentFormatOption = "auto",
1004
        chat_template_kwargs: dict[str, Any] | None = None,
1005
        add_generation_prompt: bool = True,
1006
        continue_final_message: bool = False,
1007
        tools: list[dict[str, Any]] | None = None,
1008
        tokenization_kwargs: dict[str, Any] | None = None,
1009
        mm_processor_kwargs: dict[str, Any] | None = None,
1010
    ) -> Sequence[ProcessorInputs]:
nunjunj's avatar
nunjunj committed
1011
        """
1012
1013
1014
1015
        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
1016
1017

        Returns:
1018
            A list of `ProcessorInputs` objects ready to be passed into LLMEngine.
nunjunj's avatar
nunjunj committed
1019
        """
1020
        renderer = self.renderer
1021

1022
1023
1024
1025
1026
1027
1028
1029
1030
        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,
1031
                    tokenize=is_mistral_tokenizer(renderer.tokenizer),
1032
1033
1034
                ),
            ),
        )
1035
1036
1037
        tok_params = renderer.default_chat_tok_params.with_kwargs(
            **(tokenization_kwargs or {})
        )
1038

1039
1040
1041
1042
1043
1044
        _, engine_prompts = renderer.render_chat(
            conversations,
            chat_params,
            tok_params,
            prompt_extras={"mm_processor_kwargs": mm_processor_kwargs},
        )
1045

1046
        return engine_prompts
1047

1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
    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,
    ) -> ProcessorInputs:
        (engine_prompt,) = self._preprocess_chat(
            [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,
        )

        return engine_prompt

1074
1075
    def chat(
        self,
1076
        messages: list[ChatCompletionMessageParam]
1077
1078
        | Sequence[list[ChatCompletionMessageParam]],
        sampling_params: SamplingParams | Sequence[SamplingParams] | None = None,
1079
        use_tqdm: bool | Callable[..., tqdm] = True,
1080
        lora_request: Sequence[LoRARequest] | LoRARequest | None = None,
1081
        chat_template: str | None = None,
1082
1083
1084
        chat_template_content_format: ChatTemplateContentFormatOption = "auto",
        add_generation_prompt: bool = True,
        continue_final_message: bool = False,
1085
1086
        tools: list[dict[str, Any]] | None = None,
        chat_template_kwargs: dict[str, Any] | None = None,
1087
        tokenization_kwargs: dict[str, Any] | None = None,
1088
        mm_processor_kwargs: dict[str, Any] | None = None,
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
    ) -> 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:
1101
            messages: A sequence of conversations or a single conversation.
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132

                - 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.
1133
1134
            tokenization_kwargs: Overrides for `tokenizer.encode`.
            mm_processor_kwargs: Overrides for `processor.__call__`.
1135
1136
1137
1138
1139

        Returns:
            A list of `RequestOutput` objects containing the generated
            responses in the same order as the input messages.
        """
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
        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()

1152
        return self._run_chat(
1153
1154
            messages=messages,
            params=sampling_params,
1155
            output_type=RequestOutput,
1156
1157
            use_tqdm=use_tqdm,
            lora_request=lora_request,
1158
1159
            chat_template=chat_template,
            chat_template_content_format=chat_template_content_format,
1160
            chat_template_kwargs=chat_template_kwargs,
1161
1162
1163
            add_generation_prompt=add_generation_prompt,
            continue_final_message=continue_final_message,
            tools=tools,
1164
            tokenization_kwargs=tokenization_kwargs,
1165
1166
1167
            mm_processor_kwargs=mm_processor_kwargs,
        )

1168
1169
    def encode(
        self,
1170
1171
        prompts: PromptType | Sequence[PromptType] | DataPrompt,
        pooling_params: PoolingParams | Sequence[PoolingParams] | None = None,
1172
        *,
1173
1174
        use_tqdm: bool | Callable[..., tqdm] = True,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
1175
        pooling_task: PoolingTask | None = None,
1176
        tokenization_kwargs: dict[str, Any] | None = None,
1177
    ) -> list[PoolingRequestOutput]:
1178
1179
        """Apply pooling to the hidden states corresponding to the input
        prompts.
1180

1181
        This class automatically batches the given prompts, considering
1182
1183
1184
1185
        the memory constraint. For the best performance, put all of your prompts
        into a single list and pass it to this method.

        Args:
1186
            prompts: The prompts to the LLM. You may pass a sequence of prompts
1187
                for batch inference. See [PromptType][vllm.inputs.PromptType]
1188
                for more details about the format of each prompt.
1189
1190
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1191
1192
1193
1194
            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.
1195
            lora_request: LoRA request to use for generation, if any.
1196
            pooling_task: Override the pooling task to use.
1197
            tokenization_kwargs: Overrides for `tokenizer.encode`.
1198
1199

        Returns:
1200
            A list of `PoolingRequestOutput` objects containing the
1201
            pooled hidden states in the same order as the input prompts.
1202
        """
1203

1204
        if pooling_task is None:
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
            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"`'
            )
1220

1221
        model_config = self.model_config
1222
        runner_type = model_config.runner_type
1223
        if runner_type != "pooling":
1224
1225
1226
            raise ValueError(
                "LLM.encode() is only supported for pooling models. "
                "Try passing `--runner pooling` to use the model as a "
1227
1228
                "pooling model."
            )
1229

1230
        if isinstance(prompts, dict) and "data" in prompts:
1231
1232
1233
1234
1235
            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' "
1236
1237
                    "offline inference example for more details."
                )
1238
1239

            # Validate the request data is valid for the loaded plugin
1240
1241
1242
1243
1244
1245
1246
1247
1248
            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)
1249
1250
1251

            # obtain the actual model prompts from the pre-processor
            prompts = self.io_processor.pre_process(prompt=validated_prompt)
1252
            prompts_seq = prompt_to_seq(prompts)
1253

1254
1255
1256
1257
1258
            params_seq: Sequence[PoolingParams] = [
                self.io_processor.merge_pooling_params(param)
                for param in self._params_to_seq(
                    pooling_params,
                    len(prompts_seq),
1259
                )
1260
1261
1262
1263
            ]
            for p in params_seq:
                if p.task is None:
                    p.task = "plugin"
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288

            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,
                )
            ]
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
        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)
1305

1306
1307
            if pooling_task in self.pooling_io_processors:
                io_processor = self.pooling_io_processors[pooling_task]
1308
1309
1310
1311
1312
1313
1314
                processor_inputs = io_processor.pre_process_offline(
                    prompts_seq, tokenization_kwargs
                )
                seq_lora_requests = self._lora_request_to_seq(
                    lora_request, len(prompts_seq)
                )
                seq_priority = self._priority_to_seq(None, len(prompts))
1315

1316
1317
1318
1319
1320
                self._render_and_add_requests(
                    prompts=processor_inputs,
                    params=params_seq,
                    lora_requests=seq_lora_requests,
                    priorities=seq_priority,
1321
                )
1322

1323
1324
1325
                outputs = self._run_engine(
                    use_tqdm=use_tqdm, output_type=PoolingRequestOutput
                )
1326
                outputs = io_processor.post_process_offline(outputs)
1327
1328
1329
1330
1331
1332
1333
1334
1335
            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,
                )
1336
        return outputs
1337

1338
1339
    def embed(
        self,
1340
        prompts: PromptType | Sequence[PromptType],
1341
        *,
1342
1343
1344
        use_tqdm: bool | Callable[..., tqdm] = True,
        pooling_params: PoolingParams | Sequence[PoolingParams] | None = None,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
1345
        tokenization_kwargs: dict[str, Any] | None = None,
1346
    ) -> list[EmbeddingRequestOutput]:
1347
1348
1349
1350
1351
1352
1353
1354
1355
        """
        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
1356
                for batch inference. See [PromptType][vllm.inputs.PromptType]
1357
                for more details about the format of each prompt.
1358
1359
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1360
1361
1362
1363
            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.
1364
            lora_request: LoRA request to use for generation, if any.
1365
            tokenization_kwargs: Overrides for `tokenizer.encode`.
1366
1367

        Returns:
1368
            A list of `EmbeddingRequestOutput` objects containing the
1369
1370
            embedding vectors in the same order as the input prompts.
        """
1371
        if "embed" not in self.supported_tasks:
1372
1373
            raise ValueError(
                "Embedding API is not supported by this model. "
1374
1375
                "Try converting the model using `--convert embed`."
            )
1376

1377
1378
1379
1380
1381
1382
        items = self.encode(
            prompts,
            use_tqdm=use_tqdm,
            pooling_params=pooling_params,
            lora_request=lora_request,
            pooling_task="embed",
1383
            tokenization_kwargs=tokenization_kwargs,
1384
        )
1385
1386
1387
1388
1389

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

    def classify(
        self,
1390
        prompts: PromptType | Sequence[PromptType],
1391
        *,
1392
        pooling_params: PoolingParams | Sequence[PoolingParams] | None = None,
1393
        use_tqdm: bool | Callable[..., tqdm] = True,
1394
        lora_request: list[LoRARequest] | LoRARequest | None = None,
1395
        tokenization_kwargs: dict[str, Any] | None = None,
1396
    ) -> list[ClassificationRequestOutput]:
1397
1398
1399
1400
1401
1402
1403
1404
1405
        """
        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
1406
                for batch inference. See [PromptType][vllm.inputs.PromptType]
1407
                for more details about the format of each prompt.
1408
1409
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1410
1411
1412
1413
            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.
1414
            lora_request: LoRA request to use for generation, if any.
1415
1416
            tokenization_kwargs: Overrides for `tokenizer.encode`.

1417
        Returns:
1418
            A list of `ClassificationRequestOutput` objects containing the
1419
1420
            embedding vectors in the same order as the input prompts.
        """
1421
        if "classify" not in self.supported_tasks:
1422
            raise ValueError(
1423
                "Classification API is not supported by this model. "
1424
1425
                "Try converting the model using `--convert classify`."
            )
1426

1427
1428
1429
        items = self.encode(
            prompts,
            use_tqdm=use_tqdm,
1430
            pooling_params=pooling_params,
1431
1432
            lora_request=lora_request,
            pooling_task="classify",
1433
            tokenization_kwargs=tokenization_kwargs,
1434
        )
1435
1436
1437

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

1438
1439
    def reward(
        self,
1440
        prompts: PromptType | Sequence[PromptType],
1441
1442
        /,
        *,
1443
        pooling_params: PoolingParams | Sequence[PoolingParams] | None = None,
1444
1445
        use_tqdm: bool | Callable[..., tqdm] = True,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
1446
        tokenization_kwargs: dict[str, Any] | None = None,
1447
1448
1449
1450
1451
1452
1453
    ) -> 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]
1454
                for more details about the format of each prompt.
1455
1456
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1457
1458
1459
1460
1461
            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.
1462
1463
            tokenization_kwargs: Overrides for `tokenizer.encode`.

1464
1465
1466
1467
1468
1469
1470
1471
1472
        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,
1473
            pooling_task="token_classify",
1474
            tokenization_kwargs=tokenization_kwargs,
1475
1476
        )

1477
1478
    def _embedding_score(
        self,
1479
1480
        data_1: list[ScoreData],
        data_2: list[ScoreData],
1481
1482
1483
1484
1485
        *,
        use_tqdm: bool | Callable[..., tqdm],
        pooling_params: PoolingParams | None,
        lora_request: list[LoRARequest] | LoRARequest | None,
        tokenization_kwargs: dict[str, Any],
1486
    ) -> list[ScoringRequestOutput]:
1487
1488
        tokenizer = self.get_tokenizer()

1489
1490
1491
1492
1493
1494
1495
1496
        input_texts: list[str] = []
        for text in data_1 + data_2:
            if not isinstance(text, str):
                raise NotImplementedError(
                    "Embedding scores currently do not support multimodal input."
                )
            input_texts.append(text)

1497
        encoded_output = self.encode(
1498
            input_texts,
1499
1500
            use_tqdm=use_tqdm,
            lora_request=lora_request,
1501
            pooling_params=pooling_params,
1502
            pooling_task="embed",
1503
            tokenization_kwargs=tokenization_kwargs,
1504
        )
1505

1506
1507
        encoded_output_1 = encoded_output[0 : len(data_1)]
        encoded_output_2 = encoded_output[len(data_1) :]
1508
1509
1510
1511

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

1512
        scores = _cosine_similarity(
1513
1514
1515
            tokenizer=tokenizer,
            embed_1=encoded_output_1,
            embed_2=encoded_output_2,
1516
        )
1517

1518
        return [ScoringRequestOutput.from_base(item) for item in scores]
1519

1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
    def _late_interaction_score(
        self,
        data_1: list[ScoreData],
        data_2: list[ScoreData],
        *,
        use_tqdm: bool | Callable[..., tqdm],
        pooling_params: PoolingParams | None,
        lora_request: list[LoRARequest] | LoRARequest | None,
        tokenization_kwargs: dict[str, Any],
    ) -> list[ScoringRequestOutput]:
        """
        Late interaction scoring (ColBERT MaxSim).

        Encodes queries and documents into per-token embeddings, then computes
        MaxSim: sum over query tokens of max similarity to any document token.
        """
        from vllm.outputs import PoolingOutput

        tokenizer = self.get_tokenizer()

1540
1541
1542
1543
        # Convert ScoreData to PromptType (handles both text and multimodal)
        model_config = self.model_config
        prompts_1 = score_data_to_prompts(data_1, "query", model_config)
        prompts_2 = score_data_to_prompts(data_2, "document", model_config)
1544

1545
1546
        encoded_output: list[PoolingRequestOutput] = self.encode(
            prompts_1 + prompts_2,
1547
1548
1549
1550
1551
1552
1553
            use_tqdm=use_tqdm,
            lora_request=lora_request,
            pooling_params=pooling_params,
            pooling_task="token_embed",
            tokenization_kwargs=tokenization_kwargs,
        )

1554
1555
        encoded_output_1: list[PoolingRequestOutput] = encoded_output[: len(prompts_1)]
        encoded_output_2: list[PoolingRequestOutput] = encoded_output[len(prompts_1) :]
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585

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

        # Compute MaxSim scores
        scores: list[PoolingRequestOutput] = []
        padding: list[int] = []
        if (pad_token_id := tokenizer.pad_token_id) is not None:
            padding = [pad_token_id]

        for emb_1, emb_2 in zip(encoded_output_1, encoded_output_2):
            # emb_1.outputs.data: [query_len, dim]
            # emb_2.outputs.data: [doc_len, dim]
            q_emb = emb_1.outputs.data
            d_emb = emb_2.outputs.data

            maxsim_score = compute_maxsim_score(q_emb, d_emb)

            tokens = emb_1.prompt_token_ids + padding + emb_2.prompt_token_ids

            scores.append(
                PoolingRequestOutput(
                    request_id=f"{emb_1.request_id}_{emb_2.request_id}",
                    outputs=PoolingOutput(data=maxsim_score),
                    prompt_token_ids=tokens,
                    num_cached_tokens=emb_1.num_cached_tokens + emb_2.num_cached_tokens,
                    finished=True,
                )
            )

1586
        return [ScoringRequestOutput.from_base(item) for item in scores]
1587

1588
1589
    def _cross_encoding_score(
        self,
1590
1591
        data_1: list[ScoreData],
        data_2: list[ScoreData],
1592
1593
1594
1595
1596
1597
        *,
        use_tqdm: bool | Callable[..., tqdm],
        pooling_params: PoolingParams | None,
        lora_request: list[LoRARequest] | LoRARequest | None,
        tokenization_kwargs: dict[str, Any],
        score_template: str | None,
1598
    ) -> list[ScoringRequestOutput]:
1599
        model_config = self.model_config
1600
        tokenizer = self.get_tokenizer()
1601

1602
        if is_mistral_tokenizer(tokenizer):
1603
            raise ValueError("Score API is not supported for Mistral tokenizer")
1604

1605
1606
        if len(data_1) == 1:
            data_1 = data_1 * len(data_2)
1607

1608
1609
        if pooling_params is None:
            pooling_params = PoolingParams(task="score")
1610
1611
        elif pooling_params.task is None:
            pooling_params.task = "score"
1612

1613
        pooling_params_list = list[PoolingParams]()
1614

1615
        prompts = list[PromptType]()
1616

1617
1618
        input_pairs = [(t1, t2) for t1, t2 in zip(data_1, data_2)]

1619
1620
        for q, d in input_pairs:
            _, engine_prompt = get_score_prompt(
1621
                model_config=model_config,
1622
1623
1624
1625
                data_1=q,
                data_2=d,
                tokenizer=tokenizer,
                tokenization_kwargs=tokenization_kwargs,
1626
                score_template=score_template,
1627
1628
            )

1629
            if token_type_ids := engine_prompt.pop("token_type_ids", None):
1630
1631
1632
1633
1634
1635
1636
                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)

1637
            prompts.append(engine_prompt)
1638

1639
        outputs = self._run_completion(
1640
            prompts=prompts,
1641
            params=pooling_params_list,
1642
            output_type=PoolingRequestOutput,
1643
            use_tqdm=use_tqdm,
1644
1645
1646
            lora_request=lora_request,
        )

1647
        return [ScoringRequestOutput.from_base(item) for item in outputs]
1648

1649
1650
    def score(
        self,
1651
1652
1653
1654
1655
1656
1657
1658
        data_1: SingletonPrompt
        | Sequence[SingletonPrompt]
        | ScoreMultiModalParam
        | list[ScoreMultiModalParam],
        data_2: SingletonPrompt
        | Sequence[SingletonPrompt]
        | ScoreMultiModalParam
        | list[ScoreMultiModalParam],
1659
        /,
1660
        *,
1661
1662
1663
        use_tqdm: bool | Callable[..., tqdm] = True,
        pooling_params: PoolingParams | None = None,
        lora_request: list[LoRARequest] | LoRARequest | None = None,
1664
        tokenization_kwargs: dict[str, Any] | None = None,
1665
        chat_template: str | None = None,
1666
    ) -> list[ScoringRequestOutput]:
1667
1668
        """Generate similarity scores for all pairs `<text,text_pair>` or
          `<multi-modal data, multi-modal data pair>`.
1669

1670
        The inputs can be `1 -> 1`, `1 -> N` or `N -> N`.
1671
1672
        In the `1 - N` case the `data_1` input will be replicated `N`
        times to pair with the `data_2` inputs.
1673
        The input pairs are used to build a list of prompts for the
1674
1675
        cross encoder model. This class automatically batches the prompts,
        considering the memory constraint. For the best performance, put all
1676
1677
1678
        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
1679
        appropriate multi-modal models. For multi-modal inputs, ensure the
1680
        prompt structure matches the model's expected input format.
1681
1682

        Args:
1683
1684
1685
            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
1686
                the `data_2` list.
1687
            data_2: The data to pair with the query to form the input to
1688
                the LLM. Can be text or multi-modal data. See [PromptType]
1689
                [vllm.inputs.PromptType] for more details about the format of
1690
                each prompt.
1691
1692
            pooling_params: The pooling parameters for pooling. If None, we
                use the default pooling parameters.
1693
1694
1695
1696
            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.
1697
            lora_request: LoRA request to use for generation, if any.
1698
1699
            chat_template: The chat template to use for the scoring. If None, we
                use the model's default chat template.
1700
            tokenization_kwargs: Overrides for `tokenizer.encode`.
1701
        Returns:
1702
            A list of `ScoringRequestOutput` objects containing the
1703
1704
            generated scores in the same order as the input prompts.
        """
1705
        model_config = self.model_config
1706

1707
        runner_type = model_config.runner_type
1708
        if runner_type != "pooling":
1709
1710
1711
            raise ValueError(
                "LLM.score() is only supported for pooling models. "
                "Try passing `--runner pooling` to use the model as a "
1712
1713
                "pooling model."
            )
1714

1715
        supported_tasks = self.supported_tasks
1716
1717
1718
1719
        score_type = self.model_config.score_type
        is_late_interaction = score_type == "late-interaction"
        is_cross_encoder = score_type == "cross-encoder"

1720
1721
1722
1723
        # Late interaction models (e.g., ColBERT) use token_embed for scoring
        if not is_late_interaction and all(
            t not in supported_tasks for t in ("embed", "classify")
        ):
1724
1725
1726
1727
1728
            raise ValueError(
                "Score API is not supported by this model. "
                "Try converting the model using "
                "`--convert embed` or `--convert classify`."
            )
1729

1730
        if is_cross_encoder and getattr(model_config.hf_config, "num_labels", 0) != 1:
1731
            raise ValueError("Score API is only enabled for num_labels == 1.")
1732

1733
        if not is_cross_encoder and chat_template is not None:
1734
1735
1736
1737
            raise ValueError(
                "chat_template is only supported for cross-encoder models."
            )

1738
1739
        is_multimodal_model = model_config.is_multimodal_model
        architecture = model_config.architecture
1740

1741
1742
1743
1744
1745
1746
        score_data_1, score_data_2 = validate_score_input(
            data_1,  # type: ignore[arg-type]
            data_2,  # type: ignore[arg-type]
            is_multimodal_model=is_multimodal_model,
            architecture=architecture,
        )
1747

1748
1749
1750
1751
        renderer = self.renderer
        tok_params = renderer.default_cmpl_tok_params.with_kwargs(
            **(tokenization_kwargs or {})
        )
1752
1753
        encode_kwargs = tok_params.get_encode_kwargs()

1754
        if is_cross_encoder:
1755
            return self._cross_encoding_score(
1756
1757
                score_data_1,
                score_data_2,
1758
1759
1760
1761
                use_tqdm=use_tqdm,
                pooling_params=pooling_params,
                lora_request=lora_request,
                tokenization_kwargs=encode_kwargs,
1762
                score_template=chat_template,
1763
            )
1764
1765
1766
1767
1768
1769
1770
1771
1772
        elif is_late_interaction:
            return self._late_interaction_score(
                score_data_1,
                score_data_2,
                use_tqdm=use_tqdm,
                pooling_params=pooling_params,
                lora_request=lora_request,
                tokenization_kwargs=encode_kwargs,
            )
1773
        else:
1774
            return self._embedding_score(
1775
1776
                score_data_1,
                score_data_2,
1777
1778
1779
1780
                use_tqdm=use_tqdm,
                pooling_params=pooling_params,
                lora_request=lora_request,
                tokenization_kwargs=encode_kwargs,
1781
            )
1782

1783
1784
1785
1786
1787
1788
1789
1790
1791
    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)
1792
1793
1794
1795

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

1796
1797
1798
1799
1800
1801
    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
        )
1802

1803
    def sleep(self, level: int = 1, mode: PauseMode = "abort"):
1804
1805
1806
1807
1808
        """
        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.

1809
        Args:
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
            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.
1823
1824
            mode: How to handle any existing requests, can be "abort", "wait",
                or "keep".
1825
        """
1826
        self.llm_engine.sleep(level=level, mode=mode)
1827

1828
    def wake_up(self, tags: list[str] | None = None):
1829
        """
1830
1831
        Wake up the engine from sleep mode. See the [sleep][vllm.LLM.sleep]
        method for more details.
1832

1833
        Args:
1834
1835
            tags: An optional list of tags to reallocate the engine memory
                for specific memory allocations. Values must be in
1836
1837
1838
1839
                `("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.
1840
1841
        """
        self.llm_engine.wake_up(tags)
1842

1843
1844
1845
1846
    def get_metrics(self) -> list["Metric"]:
        """Return a snapshot of aggregated metrics from Prometheus.

        Returns:
1847
            A `MetricSnapshot` instance capturing the current state
1848
1849
1850
1851
1852
1853
1854
            of all aggregated metrics from Prometheus.

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

1855
    def _params_to_seq(
1856
        self,
1857
        params: _P | Sequence[_P],
1858
        num_requests: int,
1859
    ) -> Sequence[_P]:
1860
1861
1862
1863
        if isinstance(params, Sequence):
            if len(params) != num_requests:
                raise ValueError(
                    f"The lengths of prompts ({params}) "
1864
                    f"and params ({len(params)}) must be the same."
1865
1866
                )

1867
            return params
1868

1869
1870
1871
1872
1873
1874
1875
        return [params] * num_requests

    def _lora_request_to_seq(
        self,
        lora_request: LoRARequest | None | Sequence[LoRARequest | None],
        num_requests: int,
    ) -> Sequence[LoRARequest | None]:
1876
1877
1878
1879
1880
1881
1882
        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."
                )

1883
1884
1885
            return lora_request

        return [lora_request] * num_requests
1886

1887
1888
1889
1890
1891
    def _priority_to_seq(
        self,
        priority: list[int] | None,
        num_requests: int,
    ) -> Sequence[int]:
1892
1893
1894
1895
1896
1897
1898
        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."
                )

1899
1900
1901
1902
            return priority

        return [0] * num_requests

1903
    def _add_completion_requests(
1904
1905
1906
1907
1908
1909
1910
        self,
        prompts: PromptType | Sequence[PromptType],
        params: SamplingParams
        | PoolingParams
        | Sequence[SamplingParams | PoolingParams],
        *,
        use_tqdm: bool | Callable[..., tqdm] = True,
1911
        lora_request: Sequence[LoRARequest] | LoRARequest | None = None,
1912
1913
        priority: list[int] | None = None,
        tokenization_kwargs: dict[str, Any] | None = None,
1914
    ) -> list[str]:
1915
1916
        seq_prompts = prompt_to_seq(prompts)
        seq_params = self._params_to_seq(params, len(seq_prompts))
1917
1918
1919
        seq_lora_requests = self._lora_request_to_seq(lora_request, len(seq_prompts))
        seq_priority = self._priority_to_seq(priority, len(prompts))

1920
        return self._render_and_add_requests(
1921
            prompts=(
1922
1923
1924
1925
1926
                self._preprocess_cmpl_one(prompt, tokenization_kwargs)
                for prompt in maybe_tqdm(
                    seq_prompts,
                    use_tqdm=use_tqdm,
                    desc="Rendering prompts",
1927
                )
1928
            ),
1929
            params=seq_params,
1930
1931
            lora_requests=seq_lora_requests,
            priorities=seq_priority,
1932
1933
        )

1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
    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)

1957
1958
1959
1960
1961
1962
1963
    def _run_chat(
        self,
        messages: list[ChatCompletionMessageParam]
        | Sequence[list[ChatCompletionMessageParam]],
        params: SamplingParams
        | PoolingParams
        | Sequence[SamplingParams | PoolingParams],
1964
        output_type: type[_O],
1965
1966
        *,
        use_tqdm: bool | Callable[..., tqdm] = True,
1967
        lora_request: Sequence[LoRARequest] | LoRARequest | None = None,
1968
1969
1970
1971
1972
1973
1974
1975
1976
        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,
    ):
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
        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,
1991
                    tokenization_kwargs=tokenization_kwargs,
1992
1993
                    mm_processor_kwargs=mm_processor_kwargs,
                )
1994
1995
1996
1997
                for conversation in maybe_tqdm(
                    seq_convs,
                    use_tqdm=use_tqdm,
                    desc="Rendering conversations",
1998
1999
2000
                )
            ),
            params=seq_params,
2001
            output_type=output_type,
2002
2003
            lora_requests=seq_lora_requests,
            use_tqdm=use_tqdm,
2004
2005
        )

2006
2007
2008
2009
    def _render_and_run_requests(
        self,
        prompts: Iterable[ProcessorInputs],
        params: Sequence[SamplingParams | PoolingParams],
2010
        output_type: type[_O],
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
        *,
        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,
2028
            params=params,
2029
2030
            lora_requests=lora_requests,
            priorities=priorities,
2031
2032
        )

2033
        return self._run_engine(output_type, use_tqdm=use_tqdm)
2034

2035
    def _render_and_add_requests(
2036
        self,
2037
2038
        prompts: Iterable[ProcessorInputs],
        params: Sequence[SamplingParams | PoolingParams],
2039
        *,
2040
2041
        lora_requests: Sequence[LoRARequest | None] | None = None,
        priorities: Sequence[int] | None = None,
2042
    ) -> list[str]:
2043
        added_request_ids: list[str] = []
2044

2045
        try:
2046
            for i, prompt in enumerate(prompts):
2047
2048
                request_id = self._add_request(
                    prompt,
2049
                    params[i],
Cyrus Leung's avatar
Cyrus Leung committed
2050
2051
2052
2053
                    lora_request=self._resolve_mm_lora(
                        prompt,
                        None if lora_requests is None else lora_requests[i],
                    ),
2054
                    priority=0 if priorities is None else priorities[i],
2055
2056
2057
2058
                )
                added_request_ids.append(request_id)
        except Exception as e:
            if added_request_ids:
2059
                self.llm_engine.abort_request(added_request_ids, internal=True)
2060
            raise e
2061

2062
2063
        return added_request_ids

2064
    def _add_request(
nunjunj's avatar
nunjunj committed
2065
        self,
2066
        prompt: ProcessorInputs,
2067
2068
        params: SamplingParams | PoolingParams,
        lora_request: LoRARequest | None = None,
2069
        priority: int = 0,
2070
    ) -> str:
2071
2072
2073
2074
        if isinstance(params, SamplingParams):
            # We only care about the final output
            params.output_kind = RequestOutputKind.FINAL_ONLY

2075
        request_id = str(next(self.request_counter))
2076

2077
        return self.llm_engine.add_request(
2078
            request_id,
2079
            prompt,
2080
2081
            params,
            lora_request=lora_request,
2082
            priority=priority,
nunjunj's avatar
nunjunj committed
2083
        )
2084

2085
    def _run_engine(
2086
        self,
2087
        output_type: type[_O] | tuple[type[_O], ...],
2088
2089
        *,
        use_tqdm: bool | Callable[..., tqdm] = True,
2090
    ) -> list[_O]:
2091
2092
        # Initialize tqdm.
        if use_tqdm:
Zhuohan Li's avatar
Zhuohan Li committed
2093
            num_requests = self.llm_engine.get_num_unfinished_requests()
2094
2095
            tqdm_func = use_tqdm if callable(use_tqdm) else tqdm
            pbar = tqdm_func(
2096
2097
2098
                total=num_requests,
                desc="Processed prompts",
                dynamic_ncols=True,
2099
                postfix=(f"est. speed input: {0:.2f} toks/s, output: {0:.2f} toks/s"),
2100
            )
2101

Zhuohan Li's avatar
Zhuohan Li committed
2102
        # Run the engine.
2103
        outputs: list[_O] = []
2104
2105
        total_in_toks = 0
        total_out_toks = 0
Zhuohan Li's avatar
Zhuohan Li committed
2106
2107
        while self.llm_engine.has_unfinished_requests():
            step_outputs = self.llm_engine.step()
2108
            for output in step_outputs:
2109
                assert isinstance(output, output_type)
2110
                if output.finished:
2111
                    outputs.append(output)  # type: ignore[arg-type]
2112
                    if use_tqdm:
2113
2114
                        if isinstance(output, RequestOutput):
                            # Calculate tokens only for RequestOutput
2115
                            n = len(output.outputs)
2116
                            assert output.prompt_token_ids is not None
2117
                            total_in_toks += len(output.prompt_token_ids) * n
2118
2119
                            in_spd = total_in_toks / pbar.format_dict["elapsed"]
                            total_out_toks += sum(
2120
2121
2122
                                len(stp.token_ids) for stp in output.outputs
                            )
                            out_spd = total_out_toks / pbar.format_dict["elapsed"]
2123
2124
                            pbar.postfix = (
                                f"est. speed input: {in_spd:.2f} toks/s, "
2125
2126
                                f"output: {out_spd:.2f} toks/s"
                            )
2127
                            pbar.update(n)
2128
2129
                        else:
                            pbar.update(1)
2130
2131
                        if pbar.n == num_requests:
                            pbar.refresh()
2132

2133
2134
        if use_tqdm:
            pbar.close()
2135
2136
2137
        # Sort the outputs by request ID.
        # This is necessary because some requests may be finished earlier than
        # its previous requests.
2138
        return sorted(outputs, key=lambda x: int(x.request_id))
2139

2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
    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}
        )

2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
    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