base.py 34.9 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
import asyncio
4
import copy
5
import time
6
from abc import ABC, abstractmethod
7
from collections.abc import Mapping, Sequence
8
from concurrent.futures import Executor, ThreadPoolExecutor
9
10
11
12
from functools import cached_property
from typing import TYPE_CHECKING, Any, Generic, overload

from typing_extensions import TypeVar
13

14
from vllm.inputs import (
15
    EmbedsInput,
16
    EmbedsPrompt,
17
18
19
20
21
22
    EncoderDecoderInput,
    EngineInput,
    MultiModalDataDict,
    MultiModalInput,
    MultiModalUUIDDict,
    SingletonInput,
23
    TextPrompt,
24
    TokensInput,
25
    TokensPrompt,
26
27
28
    build_enc_dec_input,
    embeds_input,
    tokens_input,
29
)
30
from vllm.logger import init_logger
31
32
33
34
35
36
37
38
39
40
from vllm.multimodal import MULTIMODAL_REGISTRY as mm_registry
from vllm.multimodal.cache import BaseMultiModalProcessorCache
from vllm.multimodal.parse import (
    MultiModalDataItems,
    MultiModalUUIDItems,
    parse_mm_uuids,
)
from vllm.multimodal.processing import BaseMultiModalProcessor
from vllm.multimodal.processing import ProcessorInputs as MMProcessorInputs
from vllm.multimodal.registry import MultiModalTimingRegistry
41
from vllm.tokenizers import TokenizerLike
42
43
44
45
from vllm.utils.async_utils import (
    AsyncMicrobatchTokenizer,
    make_async,
)
46
from vllm.utils.counter import AtomicCounter
47
48
from vllm.utils.torch_utils import set_default_torch_num_threads
from vllm.v1.metrics.stats import MultiModalCacheStats
49
50

from .embed_utils import safe_load_prompt_embeds
51
52
53
54
from .inputs import (
    DictPrompt,
    EncoderDecoderDictPrompt,
    EncoderDecoderTokPrompt,
55
56
    SingletonDictPrompt,
    SingletonTokPrompt,
57
58
    TokPrompt,
)
59
from .inputs.preprocess import extract_target_prompt
60
from .params import ChatParams, TokenizeParams
61
62

if TYPE_CHECKING:
63
    from vllm.config import VllmConfig
64
65
66
67
68
    from vllm.entrypoints.chat_utils import (
        ChatCompletionMessageParam,
        ConversationMessage,
    )

69
70
logger = init_logger(__name__)

71

72
73
74
75
76
_T = TypeVar("_T", bound=TokenizerLike, default=TokenizerLike)


class BaseRenderer(ABC, Generic[_T]):
    def __init__(self, config: "VllmConfig", tokenizer: _T | None) -> None:
77
78
        super().__init__()

79
        self.config = config
80
        self.model_config = config.model_config
81
        self.api_process_rank = config.parallel_config._api_process_rank
82

83
84
        self.tokenizer = tokenizer

85
86
87
88
89
90
91
92
93
94
95
        # Shared thread pool executor for blocking tokenizer and
        # multimodal preprocessing operations.  The multimodal processor
        # receives a deep-copied tokenizer (see #36557) so it is safe to
        # run tokenization and MM preprocessing concurrently.
        pool_workers = config.model_config.renderer_num_workers
        self._executor = ThreadPoolExecutor(max_workers=pool_workers)

        # Multimodal preprocessing is always offloaded to the thread pool
        # to keep the asyncio event loop responsive under concurrent load.
        self._mm_executor: Executor = self._executor

96
97
98
        # Lazy initialization since offline LLM doesn't use async
        self._async_tokenizer: AsyncMicrobatchTokenizer | None = None

99
        self.mm_processor: BaseMultiModalProcessor | None = None
100
        self._readonly_mm_processor: BaseMultiModalProcessor | None = None
101
        self._mm_cache_stats: MultiModalCacheStats | None = None
102
103
104
105
106
107
        self._clear_mm_cache_async = make_async(
            self.clear_mm_cache, executor=self._executor
        )
        self._process_multimodal_async = make_async(
            self._process_multimodal, executor=self._mm_executor
        )
108
109
110
        if config.model_config.is_multimodal_model:
            mm_processor_cache = mm_registry.processor_cache_from_config(config)

111
112
113
114
115
116
117
            # Deep-copy the tokenizer so the multimodal processor gets its
            # own Rust tokenizer backend.  Without this, concurrent access
            # from AsyncMicrobatchTokenizer and call_hf_processor causes
            # "RuntimeError: Already borrowed" from the Rust RefCell.
            # See: https://github.com/huggingface/tokenizers/issues/537
            mm_tokenizer = copy.deepcopy(tokenizer)

118
119
120
            with set_default_torch_num_threads():
                self.mm_processor = mm_registry.create_processor(
                    config.model_config,
121
                    tokenizer=mm_tokenizer,
122
123
124
125
126
                    cache=mm_processor_cache,
                )

            if mm_processor_cache:
                self._mm_cache_stats = MultiModalCacheStats()
127

128
129
130
131
132
133
134
135
136
137
138
139
140
            # A second processor with its own processor-only cache.
            # Used by the tokenize endpoint so that tokenize-only
            # requests don't pollute the sender cache.
            ro_cache = mm_registry.processor_only_cache_from_config(config)
            if ro_cache is not None:
                ro_tokenizer = copy.deepcopy(tokenizer)
                with set_default_torch_num_threads():
                    self._readonly_mm_processor = mm_registry.create_processor(
                        config.model_config,
                        tokenizer=ro_tokenizer,
                        cache=ro_cache,
                    )

141
142
143
            # This is used to generate internal request ID for MM processing
            # It has no relation to the request ID for engine core
            self._mm_req_counter = AtomicCounter()
144
145
146
            self._mm_timing_registry = MultiModalTimingRegistry(
                config.observability_config
            )
147

148
    def get_tokenizer(self) -> _T:
149
150
151
152
153
154
        tokenizer = self.tokenizer
        if tokenizer is None:
            raise ValueError("Tokenizer not available when `skip_tokenizer_init=True`")

        return tokenizer

155
    def get_async_tokenizer(self) -> AsyncMicrobatchTokenizer:
156
        if self._async_tokenizer is None:
157
158
159
            self._async_tokenizer = AsyncMicrobatchTokenizer(
                self.get_tokenizer(), executor=self._executor
            )
160
161
162

        return self._async_tokenizer

163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
    def get_mm_processor(self) -> "BaseMultiModalProcessor":
        if self.mm_processor is None:
            raise ValueError("Multi-modal processor not available for text-only models")

        return self.mm_processor

    @property
    def mm_processor_cache(self) -> "BaseMultiModalProcessorCache | None":
        if self.mm_processor is None:
            return None

        return self.mm_processor.cache

    def stat_mm_cache(self) -> MultiModalCacheStats | None:
        mm_cache_stats = self._mm_cache_stats
        if mm_cache_stats is None:
            return None

        self._mm_cache_stats = MultiModalCacheStats()

        return mm_cache_stats

    def update_mm_cache_stats(self) -> None:
        mm_processor_cache = self.mm_processor_cache
        mm_cache_stats = self._mm_cache_stats

        if mm_processor_cache and mm_cache_stats:
            delta = mm_processor_cache.make_stats(delta=True)
            mm_cache_stats.record(delta.total, delta.hits)

    def clear_mm_cache(self) -> None:
        mm_processor_cache = self.mm_processor_cache
        if mm_processor_cache is not None:
            mm_processor_cache.clear_cache()

        if self._mm_cache_stats is not None:
            self._mm_cache_stats.reset = True

201
202
203
204
205
206
207
    def warmup(self, chat_params: ChatParams) -> None:
        """
        Warm up this renderer to avoid first-request latency.

        For chat requests:
        - Jinja2 template compilation
        """
208
209
        from vllm.entrypoints.chat_utils import ChatTemplateResolutionError

210
        try:
211
            logger.debug("Warming up chat template processing...")
212
213
214
215
216
            start_time = time.perf_counter()

            self.render_chat([[{"role": "user", "content": "warmup"}]], chat_params)

            elapsed = time.perf_counter() - start_time
217
            logger.debug("Chat template warmup completed in %.3fs", elapsed)
218
        except ChatTemplateResolutionError:
219
            logger.debug("This model does not support chat template.")
220
        except Exception:
221
            logger.warning("Chat template warmup failed", exc_info=True)
222
223
224
225
226
227
228

        if self.mm_processor:
            from vllm.multimodal.processing import TimingContext

            model_config = self.model_config
            mm_config = model_config.get_multimodal_config()
            processor = self.mm_processor
229
230
231
            mm_limits = {
                k: v for k, v in processor.info.allowed_mm_limits.items() if v > 0
            }
232
233

            try:
234
                logger.debug("Warming up multi-modal processing...")
235
236
237
238
239
240
241
242
                start_time = time.perf_counter()

                processor_inputs = processor.dummy_inputs.get_dummy_processor_inputs(
                    seq_len=model_config.max_model_len,
                    mm_counts=dict.fromkeys(mm_limits, 1),
                    mm_options=mm_config.limit_per_prompt,
                )
                _ = processor.apply(
243
                    processor_inputs, timing_ctx=TimingContext(enabled=False)
244
245
246
247
248
                )

                elapsed = time.perf_counter() - start_time
                logger.info("Multi-modal warmup completed in %.3fs", elapsed)
            except Exception:
249
                logger.warning("Multi-modal warmup failed")
250
251
252
            finally:
                self.clear_mm_cache()

253
254
255
256
257
    async def clear_mm_cache_async(self) -> None:
        """Serialize clear_mm_cache through the shared executor to avoid
        races with concurrent process_inputs on the mm_processor_cache."""
        await self._clear_mm_cache_async()

258
259
260
261
262
    def shutdown(self) -> None:
        mm_processor_cache = self.mm_processor_cache
        if mm_processor_cache is not None:
            mm_processor_cache.close()

263
264
265
266
267
268
269
270
        if executor := getattr(self, "_executor", None):
            executor.shutdown(wait=False)

        if (
            mm_executor := getattr(self, "_mm_executor", None)
        ) is not None and mm_executor is not executor:
            mm_executor.shutdown(wait=False)

271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
    def get_bos_token_id(self) -> int | None:
        if self.tokenizer is None:
            logger.warning_once(
                "Using None for BOS token id because tokenizer is not initialized"
            )
            return None

        return self.tokenizer.bos_token_id

    def get_eos_token_id(self) -> int | None:
        if self.tokenizer is None:
            logger.warning_once(
                "Using None for EOS token id because tokenizer is not initialized"
            )
            return None

        return self.tokenizer.eos_token_id

289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
    def get_dec_start_token_id(self) -> int:
        """
        Obtain the decoder start token id employed by an encoder/decoder model,
        raising an error if it is not available.
        """
        dec_start_token_id = getattr(
            self.model_config.hf_config, "decoder_start_token_id", None
        )

        if dec_start_token_id is None:
            logger.warning_once(
                "Falling back on <BOS> for decoder start token id "
                "because decoder start token id is not available."
            )
            dec_start_token_id = self.get_bos_token_id()

        if dec_start_token_id is None:
            raise RuntimeError("Cannot find decoder start token id or <BOS>")

        return dec_start_token_id

310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
    @cached_property
    def default_cmpl_tok_params(self) -> TokenizeParams:
        mm_processor = self.mm_processor
        if mm_processor is not None:
            return mm_processor.info.default_tok_params

        model_config = self.model_config
        encoder_config = model_config.encoder_config or {}

        return TokenizeParams(
            max_total_tokens=model_config.max_model_len,
            do_lower_case=encoder_config.get("do_lower_case", False),
            add_special_tokens=True,
        )

    @cached_property
    def default_chat_tok_params(self) -> TokenizeParams:
        mm_processor = self.mm_processor
        if mm_processor is not None:
            return mm_processor.info.default_tok_params

        model_config = self.model_config
        encoder_config = model_config.encoder_config or {}

        return TokenizeParams(
            max_total_tokens=model_config.max_model_len,
            do_lower_case=encoder_config.get("do_lower_case", False),
            add_special_tokens=False,
        )

340
    # Step 1: Convert raw inputs to prompts
341
    def render_prompt(
342
        self,
343
344
345
        prompt: DictPrompt | bytes,
    ) -> DictPrompt:
        if isinstance(prompt, bytes):
346
            embeds = safe_load_prompt_embeds(self.model_config, prompt)
347
            prompt = EmbedsPrompt(prompt_embeds=embeds)
348

349
        return prompt
350

351
    def render_prompts(
352
        self,
353
354
355
        prompts: Sequence[DictPrompt | bytes],
    ) -> list[DictPrompt]:
        if len(prompts) == 0:
356
357
            raise ValueError("You must pass at least one prompt")

358
        return [self.render_prompt(prompt) for prompt in prompts]
359

360
    async def render_prompts_async(
361
        self,
362
363
364
        prompts: Sequence[DictPrompt | bytes],
    ) -> list[DictPrompt]:
        return self.render_prompts(prompts)
365

366
    @abstractmethod
367
368
369
    def render_messages(
        self,
        messages: list["ChatCompletionMessageParam"],
370
        params: ChatParams,
371
    ) -> tuple[list["ConversationMessage"], DictPrompt]:
372
373
374
375
376
        raise NotImplementedError

    async def render_messages_async(
        self,
        messages: list["ChatCompletionMessageParam"],
377
        params: ChatParams,
378
    ) -> tuple[list["ConversationMessage"], DictPrompt]:
379
380
381
        return self.render_messages(messages, params)

    # Step 2: Tokenize prompts if necessary
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
    def _tokenize_prompt(
        self,
        prompt: TextPrompt,
        params: TokenizeParams,
    ) -> TokensPrompt:
        tokenizer = self.get_tokenizer()
        prompt_token_ids = tokenizer.encode(
            prompt["prompt"],
            **params.get_encode_kwargs(),
        )

        return TokensPrompt(prompt_token_ids=prompt_token_ids, **prompt)

    async def _tokenize_prompt_async(
        self,
        prompt: TextPrompt,
        params: TokenizeParams,
    ) -> TokensPrompt:
        tokenizer = self.get_async_tokenizer()
        prompt_token_ids = await tokenizer.encode(
            prompt["prompt"],
            **params.get_encode_kwargs(),
        )

        return TokensPrompt(prompt_token_ids=prompt_token_ids, **prompt)

    def _detokenize_prompt(self, prompt: TokensPrompt) -> TokensPrompt:
        tokenizer = self.get_tokenizer()
        prompt["prompt"] = tokenizer.decode(prompt["prompt_token_ids"])

        return prompt

    async def _detokenize_prompt_async(self, prompt: TokensPrompt) -> TokensPrompt:
        tokenizer = self.get_async_tokenizer()
        prompt["prompt"] = await tokenizer.decode(prompt["prompt_token_ids"])

        return prompt

420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
    @overload
    def _tokenize_singleton_prompt(
        self,
        prompt: TextPrompt | TokensPrompt,
        params: TokenizeParams,
    ) -> TokensPrompt: ...

    @overload
    def _tokenize_singleton_prompt(  # type: ignore[misc]
        self,
        prompt: EmbedsPrompt,
        params: TokenizeParams,
    ) -> EmbedsPrompt: ...

    def _tokenize_singleton_prompt(
        self,
        prompt: SingletonDictPrompt,
        params: TokenizeParams,
    ) -> SingletonTokPrompt:
        if "prompt_token_ids" not in prompt and "prompt_embeds" not in prompt:
440
441
442
443
444
            if not isinstance(prompt.get("prompt"), str):
                raise TypeError(
                    "Expected prompt['prompt'] to be a string before tokenization; "
                    "use 'prompt_token_ids' for token ID inputs"
                )
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
            prompt = params.apply_pre_tokenization(self.tokenizer, prompt)  # type: ignore[arg-type]
            prompt = self._tokenize_prompt(prompt, params)

        if params.needs_detokenization and "prompt" not in prompt:
            if "prompt_token_ids" not in prompt:
                raise RuntimeError("Cannot run detokenization on embeddings")

            prompt = self._detokenize_prompt(prompt)  # type: ignore[arg-type]

        return params.apply_post_tokenization(self.tokenizer, prompt)  # type: ignore[arg-type]

    @overload
    async def _tokenize_singleton_prompt_async(
        self,
        prompt: TextPrompt | TokensPrompt,
        params: TokenizeParams,
    ) -> TokensPrompt: ...

    @overload
    async def _tokenize_singleton_prompt_async(  # type: ignore[misc]
        self,
        prompt: EmbedsPrompt,
        params: TokenizeParams,
    ) -> EmbedsPrompt: ...

    async def _tokenize_singleton_prompt_async(
        self,
        prompt: SingletonDictPrompt,
        params: TokenizeParams,
    ) -> SingletonTokPrompt:
        if "prompt_token_ids" not in prompt and "prompt_embeds" not in prompt:
476
477
478
479
480
            if not isinstance(prompt.get("prompt"), str):
                raise TypeError(
                    "Expected prompt['prompt'] to be a string before tokenization; "
                    "use 'prompt_token_ids' for token ID inputs"
                )
481
482
483
484
485
486
487
488
489
490
491
            prompt = params.apply_pre_tokenization(self.tokenizer, prompt)  # type: ignore[arg-type]
            prompt = await self._tokenize_prompt_async(prompt, params)

        if params.needs_detokenization and "prompt" not in prompt:
            if "prompt_token_ids" not in prompt:
                raise RuntimeError("Cannot run detokenization on embeddings")

            prompt = await self._detokenize_prompt_async(prompt)  # type: ignore[arg-type]

        return params.apply_post_tokenization(self.tokenizer, prompt)  # type: ignore[arg-type]

492
493
494
495
496
497
    def _tokenize_enc_dec_prompt(
        self,
        prompt: EncoderDecoderDictPrompt,
        params: TokenizeParams,
    ) -> EncoderDecoderTokPrompt:
        enc_prompt, dec_prompt = (
498
            self._tokenize_singleton_prompt(prompt["encoder_prompt"], params),
499
500
501
            (
                None
                if prompt["decoder_prompt"] is None
502
                else self._tokenize_singleton_prompt(prompt["decoder_prompt"], params)
503
504
505
506
507
508
509
510
511
512
513
514
515
516
            ),
        )

        return EncoderDecoderTokPrompt(
            encoder_prompt=enc_prompt,
            decoder_prompt=dec_prompt,
        )

    async def _tokenize_enc_dec_prompt_async(
        self,
        prompt: EncoderDecoderDictPrompt,
        params: TokenizeParams,
    ) -> EncoderDecoderTokPrompt:
        enc_prompt, dec_prompt = await asyncio.gather(
517
            self._tokenize_singleton_prompt_async(prompt["encoder_prompt"], params),
518
519
520
            (
                asyncio.sleep(0)
                if prompt["decoder_prompt"] is None
521
522
523
                else self._tokenize_singleton_prompt_async(
                    prompt["decoder_prompt"], params
                )
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
            ),
        )

        return EncoderDecoderTokPrompt(
            encoder_prompt=enc_prompt,
            decoder_prompt=dec_prompt,
        )

    def tokenize_prompt(
        self,
        prompt: DictPrompt,
        params: TokenizeParams,
    ) -> TokPrompt:
        if "encoder_prompt" in prompt:
            return self._tokenize_enc_dec_prompt(prompt, params)  # type: ignore[arg-type]

540
        return self._tokenize_singleton_prompt(prompt, params)
541
542
543

    def tokenize_prompts(
        self,
544
        prompts: Sequence[DictPrompt],
545
        params: TokenizeParams,
546
    ) -> list[TokPrompt]:
547
548
        return [self.tokenize_prompt(prompt, params) for prompt in prompts]

549
550
551
552
553
554
555
    async def tokenize_prompt_async(
        self,
        prompt: DictPrompt,
        params: TokenizeParams,
    ) -> TokPrompt:
        if "encoder_prompt" in prompt:
            return await self._tokenize_enc_dec_prompt_async(prompt, params)  # type: ignore[arg-type]
556

557
        return await self._tokenize_singleton_prompt_async(prompt, params)
558
559
560

    async def tokenize_prompts_async(
        self,
561
        prompts: Sequence[DictPrompt],
562
        params: TokenizeParams,
563
    ) -> list[TokPrompt]:
564
565
566
        return await asyncio.gather(
            *(self.tokenize_prompt_async(prompt, params) for prompt in prompts)
        )
567
568
569
570

    # Step 3: Add extra keys to the prompts
    def _apply_prompt_extras(
        self,
571
        prompts: Sequence[TokPrompt],
572
573
574
575
576
577
        prompt_extras: dict[str, Any] | None,
    ):
        if not prompt_extras:
            return

        for prompt in prompts:
578
            target_prompt = extract_target_prompt(self.model_config, prompt)
579
580
            target_prompt.update(prompt_extras)  # type: ignore[arg-type]

581
582
583
    # Step 4: Convert to engine inputs
    def _validate_mm_uuids(
        self,
584
585
586
        mm_data: MultiModalDataDict,
        mm_data_items: MultiModalDataItems,
        mm_uuid_items: MultiModalUUIDItems,
587
    ) -> None:
588
589
590
        # NOTE: Keys corresponding to `None` in `mm_data` don't appear in
        # `mm_data_items`
        modalities = mm_data.keys() | mm_uuid_items.keys()
591
592

        for modality in modalities:
593
594
            data_items = mm_data_items.get(modality)
            uuid_items = mm_uuid_items.get(modality)
595

596
597
598
599
600
601
            if data_items is None:
                if uuid_items is None:
                    raise ValueError(
                        f"multi_modal_data[{modality!r}] is empty but "
                        f"multi_modal_uuids[{modality!r}] is missing."
                    )
602

603
            elif uuid_items is not None:
604
                if len(data_items) != len(uuid_items):
605
606
607
608
609
610
611
                    raise ValueError(
                        f"If given, multi_modal_uuids[{modality!r}] must have "
                        f"same length as multi_modal_data[{modality!r}], but "
                        f"got {len(uuid_items)} vs {len(data_items)}."
                    )

                for i, item in enumerate(data_items):
612
613
614
615
616
                    if item is None and uuid_items[i] is None:
                        raise ValueError(
                            f"multi_modal_data[{modality!r}][{i}] is empty but "
                            f"multi_modal_uuids[{modality!r}][{i}] is missing."
                        )
617
618
619

    def _process_mm_uuids(
        self,
620
621
622
        mm_data: MultiModalDataDict,
        mm_data_items: MultiModalDataItems,
        mm_uuid_items: MultiModalUUIDItems,
623
        mm_req_id: str,
624
    ) -> MultiModalUUIDItems:
625
626
627
628
629
630
631
632
633
634
635
636
        model_config = self.model_config

        # NOTE: When users explicitly turn off BOTH prefix caching and input
        # processing caching, no multimodal features or embeddings will be
        # reused across requests, therefore identifying multimodal data items
        # by their content is no longer necessary, and we create uuids with
        # `<mm_req_id>-<modality>-<index>`, overriding even user-provided ones.
        if (
            model_config.multimodal_config
            and model_config.multimodal_config.mm_processor_cache_gb == 0
            and not self.config.cache_config.enable_prefix_caching
        ):
637
            mm_uuid_items = {
638
                modality: [f"{mm_req_id}-{modality}-{i}" for i in range(data_count)]
639
                for modality, data_count in mm_data_items.get_all_counts().items()
640
641
            }

642
        self._validate_mm_uuids(mm_data, mm_data_items, mm_uuid_items)
643

644
        return mm_uuid_items
645
646
647
648
649

    # TODO: Remove str and tokenization_kwargs after deprecating InputPreprocessor
    def _process_multimodal(
        self,
        prompt: list[int] | str,
650
651
        mm_data: MultiModalDataDict,
        mm_uuids: MultiModalUUIDDict | None,
652
653
        mm_processor_kwargs: Mapping[str, object] | None,
        tokenization_kwargs: dict[str, Any] | None,
654
655
        *,
        skip_mm_cache: bool = False,
656
    ) -> "MultiModalInput":
657
        mm_req_id = f"renderer{self.api_process_rank}-mm-{self._mm_req_counter.inc(1)}"
658

659
660
661
662
        if skip_mm_cache and self._readonly_mm_processor is not None:
            mm_processor = self._readonly_mm_processor
        else:
            mm_processor = self.get_mm_processor()
663

664
665
666
        mm_data_items = mm_processor.info.parse_mm_data(mm_data)
        mm_uuid_items = parse_mm_uuids(mm_uuids)

667
        mm_uuid_items = self._process_mm_uuids(
668
669
            mm_data, mm_data_items, mm_uuid_items, mm_req_id
        )
670

671
672
673
674
675
676
677
678
679
680
681
        mm_processor_inputs = MMProcessorInputs(
            prompt,
            mm_data_items,
            mm_uuid_items,
            hf_processor_mm_kwargs=mm_processor_kwargs or {},
            tokenization_kwargs=tokenization_kwargs or {},
        )
        mm_timing_ctx = self._mm_timing_registry.get(mm_req_id)

        with set_default_torch_num_threads():
            mm_inputs = mm_processor.apply(mm_processor_inputs, mm_timing_ctx)
682
683
684
685
686
687
688
689

        self.update_mm_cache_stats()

        return mm_inputs

    def _process_tokens(
        self,
        prompt: TokensPrompt,
690
691
        *,
        skip_mm_cache: bool = False,
692
    ) -> TokensInput | MultiModalInput:
693
694
695
        """Process token inputs, with multimodal preprocessing offloaded
        to the shared thread pool in the async variant.
        """
696
697
        prompt_token_ids = prompt["prompt_token_ids"]

698
        engine_input: TokensInput | MultiModalInput
699
        if multi_modal_data := prompt.get("multi_modal_data"):
700
            engine_input = self._process_multimodal(
701
702
703
704
705
                prompt_token_ids,
                multi_modal_data,
                mm_processor_kwargs=prompt.get("mm_processor_kwargs"),
                tokenization_kwargs=None,  # Tokenization already done in Step 2
                mm_uuids=prompt.get("multi_modal_uuids"),
706
                skip_mm_cache=skip_mm_cache,
707
708
            )
        else:
709
            engine_input = tokens_input(prompt_token_ids)
710
711

        if prompt_text := prompt.get("prompt"):
712
            engine_input["prompt"] = prompt_text
713
        if cache_salt := prompt.get("cache_salt"):
714
            engine_input["cache_salt"] = cache_salt
715

716
        return engine_input
717

718
    def _process_embeds(self, prompt: EmbedsPrompt) -> EmbedsInput:
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
        if not self.model_config.enable_prompt_embeds:
            raise ValueError(
                "You must set `--enable-prompt-embeds` to input `prompt_embeds`."
            )

        prompt_embeds = prompt["prompt_embeds"]

        # prompt_embeds must be (seq_len, hidden_size), but if the user
        # passes in a batch of size 1, i.e. (1, seq_len, hidden_size),
        # we can unambiguously process the intent by squeezing the batch
        # dimension.
        if prompt_embeds.ndim == 3:
            prompt_embeds = prompt_embeds.squeeze(dim=0)

        if prompt_embeds.ndim != 2:
            raise ValueError("prompt_embeds must be of shape (seq_len, hidden_size).")

        # Tensors must be on CPU for serialization between processes
        # in the MsgpackEncoder. Casting to CPU here ensures that there is no
        # hidden device transfer in the critical path of generation.
        prompt_embeds = prompt_embeds.cpu()

741
        return embeds_input(
742
743
744
745
            prompt_embeds=prompt_embeds,
            cache_salt=prompt.get("cache_salt"),
        )

746
747
748
    async def _process_tokens_async(
        self,
        prompt: TokensPrompt,
749
750
        *,
        skip_mm_cache: bool = False,
751
752
753
754
755
756
757
758
759
760
761
    ) -> TokensInput | MultiModalInput:
        prompt_token_ids = prompt["prompt_token_ids"]

        engine_input: TokensInput | MultiModalInput
        if multi_modal_data := prompt.get("multi_modal_data"):
            engine_input = await self._process_multimodal_async(
                prompt_token_ids,
                multi_modal_data,
                mm_processor_kwargs=prompt.get("mm_processor_kwargs"),
                tokenization_kwargs=None,
                mm_uuids=prompt.get("multi_modal_uuids"),
762
                skip_mm_cache=skip_mm_cache,
763
764
765
766
767
768
769
770
771
772
773
            )
        else:
            engine_input = tokens_input(prompt_token_ids)

        if prompt_text := prompt.get("prompt"):
            engine_input["prompt"] = prompt_text
        if cache_salt := prompt.get("cache_salt"):
            engine_input["cache_salt"] = cache_salt

        return engine_input

774
775
776
777
778
779
    def _process_singleton(
        self,
        prompt: SingletonTokPrompt,
        *,
        skip_mm_cache: bool = False,
    ) -> SingletonInput:
780
781
782
        if "prompt_embeds" in prompt:
            return self._process_embeds(prompt)  # type: ignore[arg-type]

783
        return self._process_tokens(prompt, skip_mm_cache=skip_mm_cache)  # type: ignore[arg-type]
784

785
786
787
    async def _process_singleton_async(
        self,
        prompt: SingletonTokPrompt,
788
789
        *,
        skip_mm_cache: bool = False,
790
791
792
793
    ) -> SingletonInput:
        if "prompt_embeds" in prompt:
            return self._process_embeds(prompt)  # type: ignore[arg-type]

794
        return await self._process_tokens_async(prompt, skip_mm_cache=skip_mm_cache)  # type: ignore[arg-type]
795

796
797
798
    def _process_enc_dec(
        self,
        prompt: EncoderDecoderTokPrompt,
799
800
        *,
        skip_mm_cache: bool = False,
801
    ) -> EncoderDecoderInput:
802
803
804
        enc_prompt = prompt["encoder_prompt"]
        dec_prompt = prompt["decoder_prompt"]

Ekagra Ranjan's avatar
Ekagra Ranjan committed
805
806
807
808
809
810
811
        skip_decoder_start_token = False
        if self.mm_processor is not None:
            from vllm.multimodal.processing import EncDecMultiModalProcessor

            if isinstance(self.mm_processor, EncDecMultiModalProcessor):
                skip_decoder_start_token = self.mm_processor.skip_decoder_start_token

812
        return build_enc_dec_input(
813
814
815
            encoder_input=self._process_singleton(
                enc_prompt, skip_mm_cache=skip_mm_cache
            ),
816
            decoder_input=(
817
818
819
                None
                if dec_prompt is None
                else self._process_singleton(dec_prompt, skip_mm_cache=skip_mm_cache)
820
821
            ),
            decoder_start_token_id=self.get_dec_start_token_id(),
Ekagra Ranjan's avatar
Ekagra Ranjan committed
822
            skip_decoder_start_token=skip_decoder_start_token,
823
824
        )

825
826
827
    async def _process_enc_dec_async(
        self,
        prompt: EncoderDecoderTokPrompt,
828
829
        *,
        skip_mm_cache: bool = False,
830
831
832
833
834
    ) -> EncoderDecoderInput:
        enc_prompt = prompt["encoder_prompt"]
        dec_prompt = prompt["decoder_prompt"]

        encoder_input, decoder_input = await asyncio.gather(
835
            self._process_singleton_async(enc_prompt, skip_mm_cache=skip_mm_cache),
836
837
838
            (
                asyncio.sleep(0)
                if dec_prompt is None
839
840
841
                else self._process_singleton_async(
                    dec_prompt, skip_mm_cache=skip_mm_cache
                )
842
843
844
845
846
847
848
849
850
            ),
        )

        return build_enc_dec_input(
            encoder_input=encoder_input,
            decoder_input=decoder_input,
            decoder_start_token_id=self.get_dec_start_token_id(),
        )

851
852
853
854
855
856
857
    def process_for_engine(
        self,
        prompt: TokPrompt,
        arrival_time: float,
        *,
        skip_mm_cache: bool = False,
    ) -> EngineInput:
858
        engine_input: EngineInput
859
        if "encoder_prompt" in prompt:
860
            engine_input = self._process_enc_dec(prompt, skip_mm_cache=skip_mm_cache)  # type: ignore[arg-type]
861
        else:
862
            engine_input = self._process_singleton(prompt, skip_mm_cache=skip_mm_cache)
863

864
        engine_input["arrival_time"] = arrival_time
865

866
        return engine_input
867

868
    async def process_for_engine_async(
869
870
871
872
873
        self,
        prompt: TokPrompt,
        arrival_time: float,
        *,
        skip_mm_cache: bool = False,
874
875
876
877
    ) -> EngineInput:
        engine_input: EngineInput
        if "encoder_prompt" in prompt:
            engine_input = await self._process_enc_dec_async(
878
879
                prompt,  # type: ignore[arg-type]
                skip_mm_cache=skip_mm_cache,
880
881
            )
        else:
882
883
884
            engine_input = await self._process_singleton_async(
                prompt, skip_mm_cache=skip_mm_cache
            )
885
886
887
888
889

        engine_input["arrival_time"] = arrival_time

        return engine_input

890
891
892
893
    # Top-level methods
    def render_cmpl(
        self,
        prompts: Sequence[DictPrompt | bytes],
894
        tok_params: TokenizeParams | None = None,
895
896
        *,
        prompt_extras: dict[str, Any] | None = None,
897
        skip_mm_cache: bool = False,
898
    ):
899
900
        arrival_time = time.time()

901
902
        if tok_params is None:
            tok_params = self.default_cmpl_tok_params
903

904
        dict_prompts = self.render_prompts(prompts)
905
906
907
908
        tok_prompts = self.tokenize_prompts(dict_prompts, tok_params)

        self._apply_prompt_extras(tok_prompts, prompt_extras)

909
910
911
912
        return [
            self.process_for_engine(prompt, arrival_time, skip_mm_cache=skip_mm_cache)
            for prompt in tok_prompts
        ]
913
914
915
916

    async def render_cmpl_async(
        self,
        prompts: Sequence[DictPrompt | bytes],
917
        tok_params: TokenizeParams | None = None,
918
919
        *,
        prompt_extras: dict[str, Any] | None = None,
920
        skip_mm_cache: bool = False,
921
    ):
922
923
        arrival_time = time.time()

924
925
        if tok_params is None:
            tok_params = self.default_cmpl_tok_params
926

927
        dict_prompts = await self.render_prompts_async(prompts)
928
929
930
931
        tok_prompts = await self.tokenize_prompts_async(dict_prompts, tok_params)

        self._apply_prompt_extras(tok_prompts, prompt_extras)

932
        return await asyncio.gather(
933
934
935
936
937
938
            *(
                self.process_for_engine_async(
                    p, arrival_time, skip_mm_cache=skip_mm_cache
                )
                for p in tok_prompts
            )
939
        )
940
941
942
943
944

    def render_chat(
        self,
        conversations: Sequence[list["ChatCompletionMessageParam"]],
        chat_params: ChatParams,
945
        tok_params: TokenizeParams | None = None,
946
947
        *,
        prompt_extras: dict[str, Any] | None = None,
948
        skip_mm_cache: bool = False,
949
    ):
950
951
        arrival_time = time.time()

952
953
954
        if tok_params is None:
            tok_params = self.default_chat_tok_params

955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
        rendered = [
            self.render_messages(conversation, chat_params)
            for conversation in conversations
        ]

        out_conversations = list[list["ConversationMessage"]]()
        dict_prompts = list[DictPrompt]()
        for conv, prompt in rendered:
            out_conversations.append(conv)
            dict_prompts.append(prompt)

        tok_prompts = self.tokenize_prompts(dict_prompts, tok_params)

        self._apply_prompt_extras(tok_prompts, prompt_extras)

970
        eng_prompts = [
971
972
            self.process_for_engine(prompt, arrival_time, skip_mm_cache=skip_mm_cache)
            for prompt in tok_prompts
973
974
975
        ]

        return out_conversations, eng_prompts
976
977
978
979
980

    async def render_chat_async(
        self,
        conversations: Sequence[list["ChatCompletionMessageParam"]],
        chat_params: ChatParams,
981
        tok_params: TokenizeParams | None = None,
982
983
        *,
        prompt_extras: dict[str, Any] | None = None,
984
        skip_mm_cache: bool = False,
985
    ):
986
987
        arrival_time = time.time()

988
989
990
        if tok_params is None:
            tok_params = self.default_chat_tok_params

991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
        rendered = [
            self.render_messages_async(conversation, chat_params)
            for conversation in conversations
        ]

        out_conversations = list[list["ConversationMessage"]]()
        dict_prompts = list[DictPrompt]()
        for conv, prompt in await asyncio.gather(*rendered):
            out_conversations.append(conv)
            dict_prompts.append(prompt)

        tok_prompts = await self.tokenize_prompts_async(dict_prompts, tok_params)

        self._apply_prompt_extras(tok_prompts, prompt_extras)

1006
        eng_prompts = await asyncio.gather(
1007
1008
1009
1010
1011
1012
            *(
                self.process_for_engine_async(
                    p, arrival_time, skip_mm_cache=skip_mm_cache
                )
                for p in tok_prompts
            )
1013
        )
1014
1015

        return out_conversations, eng_prompts