serving.py 57 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
import asyncio
4
import json
5
import sys
6
import time
7
import traceback
8
from collections.abc import AsyncGenerator, Callable, Iterable, Mapping
9
from concurrent.futures import ThreadPoolExecutor
10
from dataclasses import dataclass, field
11
from http import HTTPStatus
12
from typing import Any, ClassVar, Generic, TypeAlias, TypeVar
13

14
import numpy as np
15
from fastapi import Request
16
17
18
from openai.types.responses import (
    ToolChoiceFunction,
)
19
20
from pydantic import ConfigDict, TypeAdapter
from starlette.datastructures import Headers
21

22
import vllm.envs as envs
23
from vllm.beam_search import BeamSearchSequence, create_sort_beams_key_function
24
from vllm.engine.protocol import EngineClient
25
26
27
28
29
30
31
32
33
from vllm.entrypoints.chat_utils import (
    ChatCompletionMessageParam,
    ChatTemplateContentFormatOption,
    ConversationMessage,
    apply_hf_chat_template,
    apply_mistral_chat_template,
    parse_chat_messages_futures,
    resolve_chat_template_content_format,
)
34
35
36
37
38
39
from vllm.entrypoints.context import (
    ConversationContext,
    HarmonyContext,
    ParsableContext,
    StreamingHarmonyContext,
)
40
from vllm.entrypoints.logger import RequestLogger
41
from vllm.entrypoints.openai.chat_completion.protocol import (
42
    ChatCompletionNamedToolChoiceParam,
43
44
    ChatCompletionRequest,
    ChatCompletionResponse,
45
46
)
from vllm.entrypoints.openai.engine.protocol import (
47
48
49
50
    CompletionRequest,
    CompletionResponse,
    ErrorInfo,
    ErrorResponse,
51
    FunctionCall,
52
    FunctionDefinition,
53
54
55
    TranscriptionRequest,
    TranscriptionResponse,
    TranslationRequest,
56
    VLLMValidationError,
57
)
58
59
60
61
from vllm.entrypoints.openai.responses.protocol import (
    ResponseInputOutputItem,
    ResponsesRequest,
)
62
from vllm.entrypoints.openai.serving_models import OpenAIServingModels
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
from vllm.entrypoints.pooling.classify.protocol import (
    ClassificationChatRequest,
    ClassificationCompletionRequest,
    ClassificationRequest,
    ClassificationResponse,
)
from vllm.entrypoints.pooling.embed.protocol import (
    EmbeddingChatRequest,
    EmbeddingCompletionRequest,
    EmbeddingRequest,
    EmbeddingResponse,
)
from vllm.entrypoints.pooling.pooling.protocol import (
    IOProcessorRequest,
    PoolingResponse,
)
from vllm.entrypoints.pooling.score.protocol import (
    RerankRequest,
    ScoreRequest,
    ScoreResponse,
)
84
from vllm.entrypoints.renderer import BaseRenderer, CompletionRenderer, RenderConfig
85
86
87
from vllm.entrypoints.responses_utils import (
    construct_input_messages,
)
88
from vllm.entrypoints.serve.disagg.protocol import GenerateRequest, GenerateResponse
89
90
91
92
93
94
from vllm.entrypoints.serve.tokenize.protocol import (
    DetokenizeRequest,
    TokenizeChatRequest,
    TokenizeCompletionRequest,
    TokenizeResponse,
)
95
from vllm.entrypoints.utils import _validate_truncation_size
96
from vllm.inputs.data import PromptType, TokensPrompt
97
98
99
100
101
from vllm.inputs.parse import (
    PromptComponents,
    get_prompt_components,
    is_explicit_encoder_decoder_prompt,
)
102
from vllm.logger import init_logger
103
from vllm.logprobs import Logprob, PromptLogprobs
104
from vllm.lora.request import LoRARequest
105
from vllm.multimodal import MultiModalDataDict
106
from vllm.outputs import CompletionOutput, PoolingRequestOutput, RequestOutput
107
from vllm.pooling_params import PoolingParams
108
from vllm.reasoning import ReasoningParser, ReasoningParserManager
109
from vllm.sampling_params import BeamSearchParams, SamplingParams
110
from vllm.tokenizers import TokenizerLike
111
from vllm.tokenizers.deepseek_v32 import DeepseekV32Tokenizer
112
from vllm.tokenizers.mistral import MistralTokenizer
113
from vllm.tool_parsers import ToolParser, ToolParserManager
114
115
116
117
118
from vllm.tracing import (
    contains_trace_headers,
    extract_trace_headers,
    log_tracing_disabled_warning,
)
119
from vllm.utils import random_uuid
120
from vllm.utils.async_utils import (
121
    AsyncMicrobatchTokenizer,
122
    collect_from_async_generator,
123
    make_async,
124
125
    merge_async_iterators,
)
126
from vllm.utils.collection_utils import is_list_of
127
from vllm.v1.engine import EngineCoreRequest
128

129
130
131
132
133
134
135
136
137

class GenerationError(Exception):
    """raised when finish_reason indicates internal server error (500)"""

    def __init__(self, message: str = "Internal server error"):
        super().__init__(message)
        self.status_code = HTTPStatus.INTERNAL_SERVER_ERROR


138
139
logger = init_logger(__name__)

140
141
142
143
144
CompletionLikeRequest: TypeAlias = (
    CompletionRequest
    | DetokenizeRequest
    | EmbeddingCompletionRequest
    | RerankRequest
145
    | ClassificationCompletionRequest
146
147
148
    | ScoreRequest
    | TokenizeCompletionRequest
)
149

150
ChatLikeRequest: TypeAlias = (
151
152
153
154
    ChatCompletionRequest
    | EmbeddingChatRequest
    | TokenizeChatRequest
    | ClassificationChatRequest
155
156
157
158
159
160
161
162
)
SpeechToTextRequest: TypeAlias = TranscriptionRequest | TranslationRequest
AnyRequest: TypeAlias = (
    CompletionLikeRequest
    | ChatLikeRequest
    | SpeechToTextRequest
    | ResponsesRequest
    | IOProcessorRequest
163
    | GenerateRequest
164
165
166
167
168
169
170
171
172
173
174
)

AnyResponse: TypeAlias = (
    CompletionResponse
    | ChatCompletionResponse
    | EmbeddingResponse
    | TranscriptionResponse
    | TokenizeResponse
    | PoolingResponse
    | ClassificationResponse
    | ScoreResponse
175
    | GenerateResponse
176
)
177

178

179
180
181
RequestT = TypeVar("RequestT", bound=AnyRequest)


182
183
@dataclass(kw_only=True)
class RequestProcessingMixin:
184
    """
185
    Mixin for request processing,
186
187
    handling prompt preparation and engine input.
    """
188

189
    engine_prompts: list[TokensPrompt] | None = field(default_factory=list)
190
191


192
193
@dataclass(kw_only=True)
class ResponseGenerationMixin:
194
    """
195
    Mixin for response generation,
196
197
    managing result generators and final batch results.
    """
198

199
200
201
    result_generator: (
        AsyncGenerator[tuple[int, RequestOutput | PoolingRequestOutput], None] | None
    ) = None
202
    final_res_batch: list[RequestOutput | PoolingRequestOutput] = field(
203
204
        default_factory=list
    )
205
206
207
208

    model_config = ConfigDict(arbitrary_types_allowed=True)


209
210
@dataclass(kw_only=True)
class ServeContext(RequestProcessingMixin, ResponseGenerationMixin, Generic[RequestT]):
211
212
    # Shared across all requests
    request: RequestT
213
    raw_request: Request | None = None
214
215
    model_name: str
    request_id: str
216
    created_time: int = field(default_factory=lambda: int(time.time()))
217
    lora_request: LoRARequest | None = None
218
219

    # Shared across most requests
220
    tokenizer: TokenizerLike | None = None
221
222


223
224
225
@dataclass(kw_only=True)
class ClassificationServeContext(ServeContext[ClassificationRequest]):
    pass
226
227


228
@dataclass(kw_only=True)
229
class EmbeddingServeContext(ServeContext[EmbeddingRequest]):
230
    chat_template: str | None = None
231
232
233
    chat_template_content_format: ChatTemplateContentFormatOption


234
class OpenAIServing:
235
236
237
238
    request_id_prefix: ClassVar[str] = """
    A short string prepended to every request’s ID (e.g. "embd", "classify")
    so you can easily tell “this ID came from Embedding vs Classification.”
    """
239

240
241
    def __init__(
        self,
242
        engine_client: EngineClient,
243
        models: OpenAIServingModels,
244
        *,
245
        request_logger: RequestLogger | None,
246
        return_tokens_as_token_ids: bool = False,
247
        log_error_stack: bool = False,
248
    ):
249
250
        super().__init__()

251
        self.engine_client = engine_client
252

253
        self.models = models
254

255
        self.request_logger = request_logger
256
        self.return_tokens_as_token_ids = return_tokens_as_token_ids
257
        self._tokenizer_executor = ThreadPoolExecutor(max_workers=1)
258
        self._apply_mistral_chat_template_async = make_async(
259
260
            apply_mistral_chat_template, executor=self._tokenizer_executor
        )
261

262
        self._async_tokenizer_pool: dict[TokenizerLike, AsyncMicrobatchTokenizer] = {}
263
        self.log_error_stack = log_error_stack
264

265
        self.input_processor = self.models.input_processor
266
267
268
269
        self.io_processor = self.models.io_processor
        self.model_config = self.models.model_config
        self.max_model_len = self.model_config.max_model_len

270
    def _get_tool_parser(
271
        self, tool_parser_name: str | None = None, enable_auto_tools: bool = False
272
    ) -> Callable[[TokenizerLike], ToolParser] | None:
273
274
275
276
        """Get the tool parser based on the name."""
        parser = None
        if not enable_auto_tools or tool_parser_name is None:
            return parser
277
        logger.info('"auto" tool choice has been enabled.')
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297

        try:
            if tool_parser_name == "pythonic" and self.model_config.model.startswith(
                "meta-llama/Llama-3.2"
            ):
                logger.warning(
                    "Llama3.2 models may struggle to emit valid pythonic tool calls"
                )
            parser = ToolParserManager.get_tool_parser(tool_parser_name)
        except Exception as e:
            raise TypeError(
                "Error: --enable-auto-tool-choice requires "
                f"tool_parser:'{tool_parser_name}' which has not "
                "been registered"
            ) from e
        return parser

    def _get_reasoning_parser(
        self,
        reasoning_parser_name: str,
298
    ) -> Callable[[TokenizerLike], ReasoningParser] | None:
299
300
301
302
303
304
305
306
307
308
309
        """Get the reasoning parser based on the name."""
        parser = None
        if not reasoning_parser_name:
            return None
        try:
            parser = ReasoningParserManager.get_reasoning_parser(reasoning_parser_name)
            assert parser is not None
        except Exception as e:
            raise TypeError(f"{reasoning_parser_name=} has not been registered") from e
        return parser

310
    async def reset_mm_cache(self) -> None:
311
        self.input_processor.clear_mm_cache()
312
313
        await self.engine_client.reset_mm_cache()

314
315
316
317
318
    async def beam_search(
        self,
        prompt: PromptType,
        request_id: str,
        params: BeamSearchParams,
319
        lora_request: LoRARequest | None = None,
320
        trace_headers: Mapping[str, str] | None = None,
321
322
323
324
325
326
327
328
    ) -> AsyncGenerator[RequestOutput, None]:
        beam_width = params.beam_width
        max_tokens = params.max_tokens
        ignore_eos = params.ignore_eos
        temperature = params.temperature
        length_penalty = params.length_penalty
        include_stop_str_in_output = params.include_stop_str_in_output

329
330
        input_processor = self.input_processor
        tokenizer = input_processor.tokenizer
331
        if tokenizer is None:
332
333
334
335
            raise VLLMValidationError(
                "You cannot use beam search when `skip_tokenizer_init=True`",
                parameter="skip_tokenizer_init",
                value=True,
336
337
338
339
340
341
342
            )

        eos_token_id: int = tokenizer.eos_token_id  # type: ignore

        if is_explicit_encoder_decoder_prompt(prompt):
            raise NotImplementedError

343
        prompt_text: str | None
344
        prompt_token_ids: list[int]
345
        multi_modal_data: MultiModalDataDict | None
346
347
348
349
350
351
352
353
354
        if isinstance(prompt, str):
            prompt_text = prompt
            prompt_token_ids = []
            multi_modal_data = None
        else:
            prompt_text = prompt.get("prompt")  # type: ignore
            prompt_token_ids = prompt.get("prompt_token_ids", [])  # type: ignore
            multi_modal_data = prompt.get("multi_modal_data")  # type: ignore

355
356
357
358
359
360
361
362
363
364
        mm_processor_kwargs: dict[str, Any] | None = None

        # This is a workaround to fix multimodal beam search; this is a
        # bandaid fix for 2 small problems:
        # 1. Multi_modal_data on the processed_inputs currently resolves to
        #    `None`.
        # 2. preprocessing above expands the multimodal placeholders. However,
        #    this happens again in generation, so the double expansion causes
        #    a mismatch.
        # TODO - would be ideal to handle this more gracefully.
365
366
367
368
369

        tokenized_length = len(prompt_token_ids)

        sort_beams_key = create_sort_beams_key_function(eos_token_id, length_penalty)

370
        logprobs_num = 2 * beam_width
371
        beam_search_params = SamplingParams(
372
            logprobs=logprobs_num,
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
            max_tokens=1,
            temperature=temperature,
        )
        all_beams = [
            BeamSearchSequence(
                tokens=prompt_token_ids,
                cum_logprob=0,
                logprobs=[],
                multi_modal_data=multi_modal_data,
                mm_processor_kwargs=mm_processor_kwargs,
                lora_request=lora_request,
            )
        ]
        completed = []

        for _ in range(max_tokens):
            prompts_batch, lora_req_batch = zip(
                *[
                    (
392
                        TokensPrompt(
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
                            prompt_token_ids=beam.tokens,
                            multi_modal_data=beam.multi_modal_data,
                            mm_processor_kwargs=beam.mm_processor_kwargs,
                        ),
                        beam.lora_request,
                    )
                    for beam in all_beams
                ]
            )

            tasks = []
            request_id_batch = f"{request_id}-{random_uuid()}"

            for i, (individual_prompt, lora_req) in enumerate(
                zip(prompts_batch, lora_req_batch)
            ):
                request_id_item = f"{request_id_batch}-beam-{i}"
                task = asyncio.create_task(
                    collect_from_async_generator(
                        self.engine_client.generate(
                            individual_prompt,
                            beam_search_params,
                            request_id_item,
                            lora_request=lora_req,
417
                            trace_headers=trace_headers,
418
419
420
421
422
423
424
425
                        )
                    )
                )
                tasks.append(task)

            output = [x[0] for x in await asyncio.gather(*tasks)]

            new_beams = []
426
427
428
429
430
431
432
433
            # Store all new tokens generated by beam
            all_beams_token_id = []
            # Store the cumulative probability of all tokens
            # generated by beam search
            all_beams_logprob = []
            # Iterate through all beam inference results
            for i, result in enumerate(output):
                current_beam = all_beams[i]
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456

                # check for error finish reason and abort beam search
                if result.outputs[0].finish_reason == "error":
                    # yield error output and terminate beam search
                    yield RequestOutput(
                        request_id=request_id,
                        prompt=prompt_text,
                        outputs=[
                            CompletionOutput(
                                index=0,
                                text="",
                                token_ids=[],
                                cumulative_logprob=None,
                                logprobs=None,
                                finish_reason="error",
                            )
                        ],
                        finished=True,
                        prompt_token_ids=prompt_token_ids,
                        prompt_logprobs=None,
                    )
                    return

457
458
                if result.outputs[0].logprobs is not None:
                    logprobs = result.outputs[0].logprobs[0]
459
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
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
                    all_beams_token_id.extend(list(logprobs.keys()))
                    all_beams_logprob.extend(
                        [
                            current_beam.cum_logprob + obj.logprob
                            for obj in logprobs.values()
                        ]
                    )

            # Handle the token for the end of sentence (EOS)
            all_beams_token_id = np.array(all_beams_token_id)
            all_beams_logprob = np.array(all_beams_logprob)

            if not ignore_eos:
                # Get the index position of eos token in all generated results
                eos_idx = np.where(all_beams_token_id == eos_token_id)[0]
                for idx in eos_idx:
                    current_beam = all_beams[idx // logprobs_num]
                    result = output[idx // logprobs_num]
                    assert result.outputs[0].logprobs is not None
                    logprobs_entry = result.outputs[0].logprobs[0]
                    completed.append(
                        BeamSearchSequence(
                            tokens=current_beam.tokens + [eos_token_id]
                            if include_stop_str_in_output
                            else current_beam.tokens,
                            logprobs=current_beam.logprobs + [logprobs_entry],
                            cum_logprob=float(all_beams_logprob[idx]),
                            finish_reason="stop",
                            stop_reason=eos_token_id,
                        )
                    )
                # After processing, set the log probability of the eos condition
                # to negative infinity.
                all_beams_logprob[eos_idx] = -np.inf

            # Processing non-EOS tokens
            # Get indices of the top beam_width probabilities
            topn_idx = np.argpartition(np.negative(all_beams_logprob), beam_width)[
                :beam_width
            ]

            for idx in topn_idx:
                current_beam = all_beams[idx // logprobs_num]
                result = output[idx // logprobs_num]
                token_id = int(all_beams_token_id[idx])
                assert result.outputs[0].logprobs is not None
                logprobs_entry = result.outputs[0].logprobs[0]
                new_beams.append(
                    BeamSearchSequence(
                        tokens=current_beam.tokens + [token_id],
                        logprobs=current_beam.logprobs + [logprobs_entry],
                        lora_request=current_beam.lora_request,
                        cum_logprob=float(all_beams_logprob[idx]),
                        multi_modal_data=current_beam.multi_modal_data,
                        mm_processor_kwargs=current_beam.mm_processor_kwargs,
                    )
                )

            all_beams = new_beams
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551

        completed.extend(all_beams)
        sorted_completed = sorted(completed, key=sort_beams_key, reverse=True)
        best_beams = sorted_completed[:beam_width]

        for beam in best_beams:
            if beam.tokens[-1] == eos_token_id and not ignore_eos:
                # Skip the eos token in the text.
                tokens = beam.tokens[tokenized_length:-1]
            else:
                tokens = beam.tokens[tokenized_length:]
            beam.text = tokenizer.decode(tokens)

        yield RequestOutput(
            request_id=request_id,
            prompt=prompt_text,
            outputs=[
                CompletionOutput(
                    text=beam.text,  # type: ignore
                    cumulative_logprob=beam.cum_logprob,
                    token_ids=beam.tokens[tokenized_length:],
                    index=i,
                    logprobs=beam.logprobs,
                    finish_reason=beam.finish_reason
                    if beam.finish_reason is not None
                    else "length",
                    stop_reason=beam.stop_reason,
                )
                for (i, beam) in enumerate(best_beams)
            ],
            finished=True,
            prompt_token_ids=prompt_token_ids,
            prompt_logprobs=None,
        )
552

553
    def _get_renderer(self, tokenizer: TokenizerLike | None) -> BaseRenderer:
554
555
556
557
558
559
560
        """
        Get a Renderer instance with the provided tokenizer.
        Uses shared async tokenizer pool for efficiency.
        """
        return CompletionRenderer(
            model_config=self.model_config,
            tokenizer=tokenizer,
561
562
            async_tokenizer_pool=self._async_tokenizer_pool,
        )
563

564
565
566
567
568
569
570
571
572
573
574
575
576
    def _build_render_config(
        self,
        request: Any,
    ) -> RenderConfig:
        """
        Build and return a `RenderConfig` for an endpoint.

        Used by the renderer to control how prompts are prepared
        (e.g., tokenization and length handling). Endpoints should
        implement this with logic appropriate to their request type.
        """
        raise NotImplementedError

577
578
    def _get_async_tokenizer(self, tokenizer) -> AsyncMicrobatchTokenizer:
        """
579
        Return (and cache) an `AsyncMicrobatchTokenizer` bound to the
580
581
582
583
584
585
586
        given tokenizer.
        """
        async_tokenizer = self._async_tokenizer_pool.get(tokenizer)
        if async_tokenizer is None:
            async_tokenizer = AsyncMicrobatchTokenizer(tokenizer)
            self._async_tokenizer_pool[tokenizer] = async_tokenizer
        return async_tokenizer
587

588
589
590
    async def _preprocess(
        self,
        ctx: ServeContext,
591
    ) -> ErrorResponse | None:
592
593
594
595
596
597
598
599
600
        """
        Default preprocessing hook. Subclasses may override
        to prepare `ctx` (classification, embedding, etc.).
        """
        return None

    def _build_response(
        self,
        ctx: ServeContext,
601
    ) -> AnyResponse | ErrorResponse:
602
603
604
605
606
607
608
609
610
        """
        Default response builder. Subclass may override this method
        to return the appropriate response object.
        """
        return self.create_error_response("unimplemented endpoint")

    async def handle(
        self,
        ctx: ServeContext,
611
612
    ) -> AnyResponse | ErrorResponse:
        generation: AsyncGenerator[AnyResponse | ErrorResponse, None]
613
614
615
616
617
618
619
620
621
622
        generation = self._pipeline(ctx)

        async for response in generation:
            return response

        return self.create_error_response("No response yielded from pipeline")

    async def _pipeline(
        self,
        ctx: ServeContext,
623
    ) -> AsyncGenerator[AnyResponse | ErrorResponse, None]:
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
        """Execute the request processing pipeline yielding responses."""
        if error := await self._check_model(ctx.request):
            yield error
        if error := self._validate_request(ctx):
            yield error

        preprocess_ret = await self._preprocess(ctx)
        if isinstance(preprocess_ret, ErrorResponse):
            yield preprocess_ret

        generators_ret = await self._prepare_generators(ctx)
        if isinstance(generators_ret, ErrorResponse):
            yield generators_ret

        collect_ret = await self._collect_batch(ctx)
        if isinstance(collect_ret, ErrorResponse):
            yield collect_ret

        yield self._build_response(ctx)

644
    def _validate_request(self, ctx: ServeContext) -> ErrorResponse | None:
645
        truncate_prompt_tokens = getattr(ctx.request, "truncate_prompt_tokens", None)
646

647
648
649
650
        if (
            truncate_prompt_tokens is not None
            and truncate_prompt_tokens > self.max_model_len
        ):
651
652
653
            return self.create_error_response(
                "truncate_prompt_tokens value is "
                "greater than max_model_len."
654
655
                " Please, select a smaller truncation size."
            )
656
657
        return None

658
659
660
    def _create_pooling_params(
        self,
        ctx: ServeContext,
661
    ) -> PoolingParams | ErrorResponse:
662
663
        if not hasattr(ctx.request, "to_pooling_params"):
            return self.create_error_response(
664
665
                "Request type does not support pooling parameters"
            )
666
667
668

        return ctx.request.to_pooling_params()

669
670
671
    async def _prepare_generators(
        self,
        ctx: ServeContext,
672
    ) -> ErrorResponse | None:
673
        """Schedule the request and get the result generator."""
674
        generators: list[
675
            AsyncGenerator[RequestOutput | PoolingRequestOutput, None]
676
        ] = []
677
678

        try:
679
680
681
682
683
            trace_headers = (
                None
                if ctx.raw_request is None
                else await self._get_trace_headers(ctx.raw_request.headers)
            )
684

685
686
687
            pooling_params = self._create_pooling_params(ctx)
            if isinstance(pooling_params, ErrorResponse):
                return pooling_params
688
689

            if ctx.engine_prompts is None:
690
                return self.create_error_response("Engine prompts not available")
691
692
693
694

            for i, engine_prompt in enumerate(ctx.engine_prompts):
                request_id_item = f"{ctx.request_id}-{i}"

695
696
                self._log_inputs(
                    request_id_item,
697
                    engine_prompt,
698
699
700
                    params=pooling_params,
                    lora_request=ctx.lora_request,
                )
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717

                generator = self.engine_client.encode(
                    engine_prompt,
                    pooling_params,
                    request_id_item,
                    lora_request=ctx.lora_request,
                    trace_headers=trace_headers,
                    priority=getattr(ctx.request, "priority", 0),
                )

                generators.append(generator)

            ctx.result_generator = merge_async_iterators(*generators)

            return None

        except Exception as e:
718
            return self.create_error_response(e)
719
720
721
722

    async def _collect_batch(
        self,
        ctx: ServeContext,
723
    ) -> ErrorResponse | None:
724
725
726
        """Collect batch results from the result generator."""
        try:
            if ctx.engine_prompts is None:
727
                return self.create_error_response("Engine prompts not available")
728
729

            num_prompts = len(ctx.engine_prompts)
730
            final_res_batch: list[RequestOutput | PoolingRequestOutput | None]
731
732
733
            final_res_batch = [None] * num_prompts

            if ctx.result_generator is None:
734
                return self.create_error_response("Result generator not available")
735
736
737
738
739
740

            async for i, res in ctx.result_generator:
                final_res_batch[i] = res

            if None in final_res_batch:
                return self.create_error_response(
741
742
                    "Failed to generate results for all prompts"
                )
743

744
            ctx.final_res_batch = [res for res in final_res_batch if res is not None]
745
746
747
748

            return None

        except Exception as e:
749
            return self.create_error_response(e)
750

751
    def create_error_response(
752
        self,
753
        message: str | Exception,
754
755
        err_type: str = "BadRequestError",
        status_code: HTTPStatus = HTTPStatus.BAD_REQUEST,
756
        param: str | None = None,
757
    ) -> ErrorResponse:
758
759
760
761
762
        exc: Exception | None = None

        if isinstance(message, Exception):
            exc = message

763
            from vllm.exceptions import VLLMValidationError
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785

            if isinstance(exc, VLLMValidationError):
                err_type = "BadRequestError"
                status_code = HTTPStatus.BAD_REQUEST
                param = exc.parameter
            elif isinstance(exc, (ValueError, TypeError, RuntimeError)):
                # Common validation errors from user input
                err_type = "BadRequestError"
                status_code = HTTPStatus.BAD_REQUEST
                param = None
            elif exc.__class__.__name__ == "TemplateError":
                # jinja2.TemplateError (avoid importing jinja2)
                err_type = "BadRequestError"
                status_code = HTTPStatus.BAD_REQUEST
                param = None
            else:
                err_type = "InternalServerError"
                status_code = HTTPStatus.INTERNAL_SERVER_ERROR
                param = None

            message = str(exc)

786
787
788
789
790
791
        if self.log_error_stack:
            exc_type, _, _ = sys.exc_info()
            if exc_type is not None:
                traceback.print_exc()
            else:
                traceback.print_stack()
792
        return ErrorResponse(
793
794
795
796
797
798
            error=ErrorInfo(
                message=message,
                type=err_type,
                code=status_code.value,
                param=param,
            )
799
        )
800

801
    def create_streaming_error_response(
802
        self,
803
        message: str | Exception,
804
805
        err_type: str = "BadRequestError",
        status_code: HTTPStatus = HTTPStatus.BAD_REQUEST,
806
        param: str | None = None,
807
    ) -> str:
808
        json_str = json.dumps(
809
            self.create_error_response(
810
811
812
813
                message=message,
                err_type=err_type,
                status_code=status_code,
                param=param,
814
815
            ).model_dump()
        )
816
817
        return json_str

818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
    def _raise_if_error(self, finish_reason: str | None, request_id: str) -> None:
        """Raise GenerationError if finish_reason indicates an error."""
        if finish_reason == "error":
            logger.error(
                "Request %s failed with an internal error during generation",
                request_id,
            )
            raise GenerationError("Internal server error")

    def _convert_generation_error_to_response(
        self, e: GenerationError
    ) -> ErrorResponse:
        """Convert GenerationError to ErrorResponse."""
        return self.create_error_response(
            str(e),
            err_type="InternalServerError",
            status_code=e.status_code,
        )

    def _convert_generation_error_to_streaming_response(
        self, e: GenerationError
    ) -> str:
        """Convert GenerationError to streaming error response."""
        return self.create_streaming_error_response(
            str(e),
            err_type="InternalServerError",
            status_code=e.status_code,
        )

847
    async def _check_model(
848
849
        self,
        request: AnyRequest,
850
    ) -> ErrorResponse | None:
851
852
        error_response = None

853
        if self._is_model_supported(request.model):
854
            return None
855
        if request.model in self.models.lora_requests:
856
            return None
857
858
859
860
861
        if (
            envs.VLLM_ALLOW_RUNTIME_LORA_UPDATING
            and request.model
            and (load_result := await self.models.resolve_lora(request.model))
        ):
862
863
            if isinstance(load_result, LoRARequest):
                return None
864
865
866
867
            if (
                isinstance(load_result, ErrorResponse)
                and load_result.error.code == HTTPStatus.BAD_REQUEST.value
            ):
868
869
870
                error_response = load_result

        return error_response or self.create_error_response(
871
872
            message=f"The model `{request.model}` does not exist.",
            err_type="NotFoundError",
873
            status_code=HTTPStatus.NOT_FOUND,
874
            param="model",
875
        )
876

877
    def _get_active_default_mm_loras(self, request: AnyRequest) -> LoRARequest | None:
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
        """Determine if there are any active default multimodal loras."""
        # TODO: Currently this is only enabled for chat completions
        # to be better aligned with only being enabled for .generate
        # when run offline. It would be nice to support additional
        # tasks types in the future.
        message_types = self._get_message_types(request)
        default_mm_loras = set()

        for lora in self.models.lora_requests.values():
            # Best effort match for default multimodal lora adapters;
            # There is probably a better way to do this, but currently
            # this matches against the set of 'types' in any content lists
            # up until '_', e.g., to match audio_url -> audio
            if lora.lora_name in message_types:
                default_mm_loras.add(lora)

        # Currently only support default modality specific loras if
        # we have exactly one lora matched on the request.
        if len(default_mm_loras) == 1:
            return default_mm_loras.pop()
        return None

900
    def _maybe_get_adapters(
901
902
903
        self,
        request: AnyRequest,
        supports_default_mm_loras: bool = False,
904
    ) -> LoRARequest | None:
905
        if request.model in self.models.lora_requests:
906
            return self.models.lora_requests[request.model]
907
908
909
910
911
912

        # Currently only support default modality specific loras
        # if we have exactly one lora matched on the request.
        if supports_default_mm_loras:
            default_mm_lora = self._get_active_default_mm_loras(request)
            if default_mm_lora is not None:
913
                return default_mm_lora
914
915

        if self._is_model_supported(request.model):
916
            return None
917

918
        # if _check_model has been called earlier, this will be unreachable
919
        raise ValueError(f"The model `{request.model}` does not exist.")
920

921
922
923
924
925
926
927
928
929
930
    def _get_message_types(self, request: AnyRequest) -> set[str]:
        """Retrieve the set of types from message content dicts up
        until `_`; we use this to match potential multimodal data
        with default per modality loras.
        """
        message_types: set[str] = set()

        if not hasattr(request, "messages"):
            return message_types

931
932
933
934
935
        messages = request.messages
        if messages is None or isinstance(messages, (str, bytes)):
            return message_types

        for message in messages:
936
937
938
939
940
            if (
                isinstance(message, dict)
                and "content" in message
                and isinstance(message["content"], list)
            ):
941
942
943
944
945
                for content_dict in message["content"]:
                    if "type" in content_dict:
                        message_types.add(content_dict["type"].split("_")[0])
        return message_types

946
    async def _normalize_prompt_text_to_input(
947
948
949
        self,
        request: AnyRequest,
        prompt: str,
950
        tokenizer: TokenizerLike,
951
        add_special_tokens: bool,
952
    ) -> TokensPrompt:
953
954
        async_tokenizer = self._get_async_tokenizer(tokenizer)

955
956
957
958
        if (
            self.model_config.encoder_config is not None
            and self.model_config.encoder_config.get("do_lower_case", False)
        ):
959
960
            prompt = prompt.lower()

961
        truncate_prompt_tokens = getattr(request, "truncate_prompt_tokens", None)
962

963
        if truncate_prompt_tokens is None:
964
            encoded = await async_tokenizer(
965
966
                prompt, add_special_tokens=add_special_tokens
            )
967
968
        elif truncate_prompt_tokens < 0:
            # Negative means we cap at the model's max length
969
970
971
972
            encoded = await async_tokenizer(
                prompt,
                add_special_tokens=add_special_tokens,
                truncation=True,
973
974
                max_length=self.max_model_len,
            )
975
        else:
976
977
978
979
            encoded = await async_tokenizer(
                prompt,
                add_special_tokens=add_special_tokens,
                truncation=True,
980
981
                max_length=truncate_prompt_tokens,
            )
982
983
984
985
986
987

        input_ids = encoded.input_ids
        input_text = prompt

        return self._validate_input(request, input_ids, input_text)

988
    async def _normalize_prompt_tokens_to_input(
989
990
        self,
        request: AnyRequest,
991
        prompt_ids: list[int],
992
        tokenizer: TokenizerLike | None,
993
    ) -> TokensPrompt:
994
        truncate_prompt_tokens = getattr(request, "truncate_prompt_tokens", None)
995

996
        if truncate_prompt_tokens is None:
997
            input_ids = prompt_ids
998
        elif truncate_prompt_tokens < 0:
999
            input_ids = prompt_ids[-self.max_model_len :]
1000
1001
1002
        else:
            input_ids = prompt_ids[-truncate_prompt_tokens:]

1003
1004
1005
1006
1007
        if tokenizer is None:
            input_text = ""
        else:
            async_tokenizer = self._get_async_tokenizer(tokenizer)
            input_text = await async_tokenizer.decode(input_ids)
1008

1009
1010
1011
1012
1013
        return self._validate_input(request, input_ids, input_text)

    def _validate_input(
        self,
        request: AnyRequest,
1014
        input_ids: list[int],
1015
        input_text: str,
1016
    ) -> TokensPrompt:
1017
1018
        token_num = len(input_ids)

1019
1020
        # Note: EmbeddingRequest, ClassificationRequest,
        # and ScoreRequest doesn't have max_tokens
1021
        if isinstance(
1022
            request,
1023
1024
1025
1026
1027
            (
                EmbeddingChatRequest,
                EmbeddingCompletionRequest,
                ScoreRequest,
                RerankRequest,
1028
1029
                ClassificationCompletionRequest,
                ClassificationChatRequest,
1030
1031
            ),
        ):
1032
1033
            # Note: input length can be up to the entire model context length
            # since these requests don't generate tokens.
1034
            if token_num > self.max_model_len:
1035
1036
                operations: dict[type[AnyRequest], str] = {
                    ScoreRequest: "score",
1037
1038
                    ClassificationCompletionRequest: "classification",
                    ClassificationChatRequest: "classification",
1039
                }
1040
                operation = operations.get(type(request), "embedding generation")
1041
                raise VLLMValidationError(
1042
1043
                    f"This model's maximum context length is "
                    f"{self.max_model_len} tokens. However, you requested "
1044
                    f"{token_num} tokens in the input for {operation}. "
1045
1046
1047
                    f"Please reduce the length of the input.",
                    parameter="input_tokens",
                    value=token_num,
1048
                )
1049
            return TokensPrompt(prompt=input_text, prompt_token_ids=input_ids)
1050

1051
1052
        # Note: TokenizeRequest and DetokenizeRequest doesn't have max_tokens
        # and does not require model context length validation
1053
        if isinstance(
1054
1055
            request,
            (TokenizeCompletionRequest, TokenizeChatRequest, DetokenizeRequest),
1056
        ):
1057
            return TokensPrompt(prompt=input_text, prompt_token_ids=input_ids)
1058

1059
1060
1061
1062
1063
        # chat completion endpoint supports max_completion_tokens
        if isinstance(request, ChatCompletionRequest):
            # TODO(#9845): remove max_tokens when field dropped from OpenAI API
            max_tokens = request.max_completion_tokens or request.max_tokens
        else:
1064
            max_tokens = getattr(request, "max_tokens", None)
1065
1066
1067
1068

        # Note: input length can be up to model context length - 1 for
        # completion-like requests.
        if token_num >= self.max_model_len:
1069
            raise VLLMValidationError(
1070
                f"This model's maximum context length is "
1071
1072
                f"{self.max_model_len} tokens. However, your request has "
                f"{token_num} input tokens. Please reduce the length of "
1073
1074
1075
                "the input messages.",
                parameter="input_tokens",
                value=token_num,
1076
            )
1077

1078
        if max_tokens is not None and token_num + max_tokens > self.max_model_len:
1079
            raise VLLMValidationError(
1080
1081
1082
1083
                "'max_tokens' or 'max_completion_tokens' is too large: "
                f"{max_tokens}. This model's maximum context length is "
                f"{self.max_model_len} tokens and your request has "
                f"{token_num} input tokens ({max_tokens} > {self.max_model_len}"
1084
1085
1086
                f" - {token_num}).",
                parameter="max_tokens",
                value=max_tokens,
1087
            )
1088

1089
        return TokensPrompt(prompt=input_text, prompt_token_ids=input_ids)
1090

1091
    async def _tokenize_prompt_input_async(
1092
1093
        self,
        request: AnyRequest,
1094
        tokenizer: TokenizerLike,
1095
        prompt_input: str | list[int],
1096
        add_special_tokens: bool = True,
1097
    ) -> TokensPrompt:
1098
        """
1099
        A simpler implementation that tokenizes a single prompt input.
1100
        """
1101
        async for result in self._tokenize_prompt_inputs_async(
1102
1103
            request,
            tokenizer,
1104
            [prompt_input],
1105
            add_special_tokens=add_special_tokens,
1106
1107
1108
        ):
            return result
        raise ValueError("No results yielded from tokenization")
1109

1110
    async def _tokenize_prompt_inputs_async(
1111
1112
        self,
        request: AnyRequest,
1113
        tokenizer: TokenizerLike,
1114
        prompt_inputs: Iterable[str | list[int]],
1115
        add_special_tokens: bool = True,
1116
    ) -> AsyncGenerator[TokensPrompt, None]:
1117
        """
1118
        A simpler implementation that tokenizes multiple prompt inputs.
1119
        """
1120
1121
        for prompt in prompt_inputs:
            if isinstance(prompt, str):
1122
                yield await self._normalize_prompt_text_to_input(
1123
                    request,
1124
1125
                    prompt=prompt,
                    tokenizer=tokenizer,
1126
1127
1128
                    add_special_tokens=add_special_tokens,
                )
            else:
1129
                yield await self._normalize_prompt_tokens_to_input(
1130
                    request,
1131
1132
                    prompt_ids=prompt,
                    tokenizer=tokenizer,
1133
1134
                )

1135
1136
    def _validate_chat_template(
        self,
1137
1138
        request_chat_template: str | None,
        chat_template_kwargs: dict[str, Any] | None,
1139
        trust_request_chat_template: bool,
1140
    ) -> ErrorResponse | None:
1141
        if not trust_request_chat_template and (
1142
1143
1144
1145
1146
1147
            request_chat_template is not None
            or (
                chat_template_kwargs
                and chat_template_kwargs.get("chat_template") is not None
            )
        ):
1148
1149
1150
            return self.create_error_response(
                "Chat template is passed with request, but "
                "--trust-request-chat-template is not set. "
1151
1152
                "Refused request with untrusted chat template."
            )
1153
1154
        return None

1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
    @staticmethod
    def _prepare_extra_chat_template_kwargs(
        request_chat_template_kwargs: dict[str, Any] | None = None,
        default_chat_template_kwargs: dict[str, Any] | None = None,
    ) -> dict[str, Any]:
        """Helper to merge server-default and request-specific chat template kwargs."""
        request_chat_template_kwargs = request_chat_template_kwargs or {}
        if default_chat_template_kwargs is None:
            return request_chat_template_kwargs
        # Apply server defaults first, then request kwargs override.
        return default_chat_template_kwargs | request_chat_template_kwargs

1167
1168
    async def _preprocess_chat(
        self,
1169
        request: ChatLikeRequest | ResponsesRequest,
1170
        tokenizer: TokenizerLike | None,
1171
        messages: list[ChatCompletionMessageParam],
1172
        chat_template: str | None,
1173
        chat_template_content_format: ChatTemplateContentFormatOption,
1174
1175
        add_generation_prompt: bool = True,
        continue_final_message: bool = False,
1176
1177
1178
        tool_dicts: list[dict[str, Any]] | None = None,
        documents: list[dict[str, str]] | None = None,
        chat_template_kwargs: dict[str, Any] | None = None,
1179
        default_chat_template_kwargs: dict[str, Any] | None = None,
1180
        tool_parser: Callable[[TokenizerLike], ToolParser] | None = None,
1181
        add_special_tokens: bool = False,
1182
    ) -> tuple[list[ConversationMessage], list[TokensPrompt]]:
1183
        model_config = self.model_config
1184

1185
1186
        resolved_content_format = resolve_chat_template_content_format(
            chat_template,
1187
            tool_dicts,
1188
1189
            chat_template_content_format,
            tokenizer,
1190
            model_config=model_config,
1191
        )
1192
        conversation, mm_data_future, mm_uuids = parse_chat_messages_futures(
1193
            messages,
1194
            model_config,
1195
            content_format=resolved_content_format,
1196
1197
        )

1198
        _chat_template_kwargs: dict[str, Any] = dict(
1199
1200
1201
1202
1203
1204
            chat_template=chat_template,
            add_generation_prompt=add_generation_prompt,
            continue_final_message=continue_final_message,
            tools=tool_dicts,
            documents=documents,
        )
1205
1206
1207
1208
        _chat_template_kwargs |= self._prepare_extra_chat_template_kwargs(
            chat_template_kwargs,
            default_chat_template_kwargs,
        )
1209

1210
        request_prompt: str | list[int]
1211
1212
1213
1214

        if tokenizer is None:
            request_prompt = "placeholder"
        elif isinstance(tokenizer, MistralTokenizer):
1215
            request_prompt = await self._apply_mistral_chat_template_async(
1216
1217
                tokenizer,
                messages=messages,
1218
                **_chat_template_kwargs,
1219
            )
1220
1221
1222
1223
        elif isinstance(tokenizer, DeepseekV32Tokenizer):
            request_prompt = tokenizer.apply_chat_template(
                conversation=conversation,
                messages=messages,
1224
                model_config=model_config,
1225
1226
                **_chat_template_kwargs,
            )
1227
1228
        else:
            request_prompt = apply_hf_chat_template(
1229
                tokenizer=tokenizer,
1230
                conversation=conversation,
1231
                model_config=model_config,
1232
                **_chat_template_kwargs,
1233
1234
1235
1236
            )

        mm_data = await mm_data_future

1237
1238
1239
        # tool parsing is done only if a tool_parser has been set and if
        # tool_choice is not "none" (if tool_choice is "none" but a tool_parser
        # is set, we want to prevent parsing a tool_call hallucinated by the LLM
1240
1241
1242
        should_parse_tools = tool_parser is not None and (
            hasattr(request, "tool_choice") and request.tool_choice != "none"
        )
1243
1244

        if should_parse_tools:
1245
1246
1247
1248
1249
            if not isinstance(request, ChatCompletionRequest | ResponsesRequest):
                msg = (
                    "Tool usage is only supported for Chat Completions API "
                    "or Responses API requests."
                )
1250
                raise NotImplementedError(msg)
1251
            request = tool_parser(tokenizer).adjust_request(request=request)  # type: ignore
1252

1253
1254
        if tokenizer is None:
            assert isinstance(request_prompt, str), (
1255
1256
                "Prompt has to be a string",
                "when the tokenizer is not initialised",
1257
            )
1258
            prompt_inputs = TokensPrompt(prompt=request_prompt, prompt_token_ids=[1])
1259
        elif isinstance(request_prompt, str):
1260
            prompt_inputs = await self._tokenize_prompt_input_async(
1261
1262
1263
1264
1265
1266
1267
1268
                request,
                tokenizer,
                request_prompt,
                add_special_tokens=add_special_tokens,
            )
        else:
            # For MistralTokenizer
            assert is_list_of(request_prompt, int), (
1269
1270
                "Prompt has to be either a string or a list of token ids"
            )
1271
            prompt_inputs = TokensPrompt(
1272
                prompt=tokenizer.decode(request_prompt),
1273
1274
                prompt_token_ids=request_prompt,
            )
1275

1276
1277
1278
1279
        engine_prompt = TokensPrompt(prompt_token_ids=prompt_inputs["prompt_token_ids"])
        if "prompt" in prompt_inputs:
            engine_prompt["prompt"] = prompt_inputs["prompt"]

1280
1281
        if mm_data is not None:
            engine_prompt["multi_modal_data"] = mm_data
1282
1283
1284
1285

        if mm_uuids is not None:
            engine_prompt["multi_modal_uuids"] = mm_uuids

1286
1287
        if request.mm_processor_kwargs is not None:
            engine_prompt["mm_processor_kwargs"] = request.mm_processor_kwargs
1288

1289
1290
1291
        if hasattr(request, "cache_salt") and request.cache_salt is not None:
            engine_prompt["cache_salt"] = request.cache_salt

1292
        return conversation, [engine_prompt]
1293

1294
1295
1296
1297
    async def _process_inputs(
        self,
        request_id: str,
        engine_prompt: PromptType,
1298
        params: SamplingParams | PoolingParams,
1299
        *,
1300
1301
        lora_request: LoRARequest | None,
        trace_headers: Mapping[str, str] | None,
1302
        priority: int,
1303
        data_parallel_rank: int | None = None,
1304
    ) -> tuple[EngineCoreRequest, dict[str, Any]]:
1305
        """Use the Processor to process inputs for AsyncLLM."""
1306
        tokenization_kwargs: dict[str, Any] = {}
1307
1308
1309
        _validate_truncation_size(
            self.max_model_len, params.truncate_prompt_tokens, tokenization_kwargs
        )
1310

1311
        engine_request = self.input_processor.process_inputs(
1312
1313
            request_id,
            engine_prompt,
1314
            params,
1315
1316
1317
1318
            lora_request=lora_request,
            tokenization_kwargs=tokenization_kwargs,
            trace_headers=trace_headers,
            priority=priority,
1319
            data_parallel_rank=data_parallel_rank,
1320
1321
1322
        )
        return engine_request, tokenization_kwargs

1323
1324
1325
    async def _render_next_turn(
        self,
        request: ResponsesRequest,
1326
        tokenizer: TokenizerLike | None,
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
        messages: list[ResponseInputOutputItem],
        tool_dicts: list[dict[str, Any]] | None,
        tool_parser,
        chat_template: str | None,
        chat_template_content_format: ChatTemplateContentFormatOption,
    ):
        new_messages = construct_input_messages(
            request_input=messages,
        )

1337
        _, engine_prompts = await self._preprocess_chat(
1338
1339
1340
1341
1342
1343
1344
1345
            request,
            tokenizer,
            new_messages,
            tool_dicts=tool_dicts,
            tool_parser=tool_parser,
            chat_template=chat_template,
            chat_template_content_format=chat_template_content_format,
        )
1346
        return engine_prompts
1347

1348
1349
1350
    async def _generate_with_builtin_tools(
        self,
        request_id: str,
1351
        engine_prompt: TokensPrompt,
1352
1353
        sampling_params: SamplingParams,
        context: ConversationContext,
1354
        lora_request: LoRARequest | None = None,
1355
1356
1357
        priority: int = 0,
        **kwargs,
    ):
1358
1359
        prompt_text, _, _ = self._get_prompt_components(engine_prompt)

1360
        orig_priority = priority
1361
        sub_request = 0
1362
        while True:
1363
1364
            # Ensure that each sub-request has a unique request id.
            sub_request_id = f"{request_id}_{sub_request}"
1365
            self._log_inputs(
1366
                sub_request_id,
1367
                engine_prompt,
1368
1369
1370
                params=sampling_params,
                lora_request=lora_request,
            )
1371
            trace_headers = kwargs.get("trace_headers")
1372
            engine_request, tokenization_kwargs = await self._process_inputs(
1373
                sub_request_id,
1374
1375
                engine_prompt,
                sampling_params,
1376
1377
1378
                lora_request=lora_request,
                trace_headers=trace_headers,
                priority=priority,
1379
            )
1380
1381
1382
1383

            generator = self.engine_client.generate(
                engine_request,
                sampling_params,
1384
                sub_request_id,
1385
1386
                lora_request=lora_request,
                priority=priority,
1387
1388
                prompt_text=prompt_text,
                tokenization_kwargs=tokenization_kwargs,
1389
1390
                **kwargs,
            )
1391

1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
            async for res in generator:
                context.append_output(res)
                # NOTE(woosuk): The stop condition is handled by the engine.
                yield context

            if not context.need_builtin_tool_call():
                # The model did not ask for a tool call, so we're done.
                break

            # Call the tool and update the context with the result.
            tool_output = await context.call_tool()
1403
            context.append_tool_output(tool_output)
1404
1405
1406
1407
1408
1409

            # TODO: uncomment this and enable tool output streaming
            # yield context

            # Create inputs for the next turn.
            # Render the next prompt token ids.
1410
1411
            if isinstance(context, (HarmonyContext, StreamingHarmonyContext)):
                prompt_token_ids = context.render_for_completion()
1412
                engine_prompt = TokensPrompt(prompt_token_ids=prompt_token_ids)
1413
            elif isinstance(context, ParsableContext):
1414
                engine_prompts = await self._render_next_turn(
1415
1416
1417
1418
1419
1420
1421
1422
1423
                    context.request,
                    context.tokenizer,
                    context.parser.response_messages,
                    context.tool_dicts,
                    context.tool_parser_cls,
                    context.chat_template,
                    context.chat_template_content_format,
                )
                engine_prompt = engine_prompts[0]
1424
                prompt_text, _, _ = self._get_prompt_components(engine_prompt)
1425

1426
            # Update the sampling params.
1427
1428
1429
            sampling_params.max_tokens = self.max_model_len - len(
                engine_prompt["prompt_token_ids"]
            )
1430
1431
            # OPTIMIZATION
            priority = orig_priority - 1
1432
            sub_request += 1
1433

1434
1435
    def _get_prompt_components(self, prompt: PromptType) -> PromptComponents:
        return get_prompt_components(prompt)
1436

1437
1438
1439
    def _log_inputs(
        self,
        request_id: str,
1440
        inputs: PromptType,
1441
1442
        params: SamplingParams | PoolingParams | BeamSearchParams | None,
        lora_request: LoRARequest | None,
1443
1444
1445
    ) -> None:
        if self.request_logger is None:
            return
1446

1447
        prompt, prompt_token_ids, prompt_embeds = self._get_prompt_components(inputs)
1448
1449
1450
1451
1452

        self.request_logger.log_inputs(
            request_id,
            prompt,
            prompt_token_ids,
1453
            prompt_embeds,
1454
1455
1456
            params=params,
            lora_request=lora_request,
        )
1457

1458
1459
1460
    async def _get_trace_headers(
        self,
        headers: Headers,
1461
    ) -> Mapping[str, str] | None:
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
        is_tracing_enabled = await self.engine_client.is_tracing_enabled()

        if is_tracing_enabled:
            return extract_trace_headers(headers)

        if contains_trace_headers(headers):
            log_tracing_disabled_warning()

        return None

1472
    @staticmethod
1473
    def _base_request_id(
1474
1475
        raw_request: Request | None, default: str | None = None
    ) -> str | None:
1476
        """Pulls the request id to use from a header, if provided"""
1477
1478
1479
1480
        if raw_request is not None and (
            (req_id := raw_request.headers.get("X-Request-Id")) is not None
        ):
            return req_id
1481

1482
        return random_uuid() if default is None else default
1483

1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
    @staticmethod
    def _get_data_parallel_rank(raw_request: Request | None) -> int | None:
        """Pulls the data parallel rank from a header, if provided"""
        if raw_request is None:
            return None

        rank_str = raw_request.headers.get("X-data-parallel-rank")
        if rank_str is None:
            return None

        try:
            return int(rank_str)
        except ValueError:
            return None

1499
1500
1501
    @staticmethod
    def _parse_tool_calls_from_content(
        request: ResponsesRequest | ChatCompletionRequest,
1502
        tokenizer: TokenizerLike | None,
1503
        enable_auto_tools: bool,
1504
        tool_parser_cls: Callable[[TokenizerLike], ToolParser] | None,
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
        content: str | None = None,
    ) -> tuple[list[FunctionCall] | None, str | None]:
        function_calls = list[FunctionCall]()
        if request.tool_choice and isinstance(request.tool_choice, ToolChoiceFunction):
            assert content is not None
            # Forced Function Call
            function_calls.append(
                FunctionCall(name=request.tool_choice.name, arguments=content)
            )
            content = None  # Clear content since tool is called.
        elif request.tool_choice and isinstance(
            request.tool_choice, ChatCompletionNamedToolChoiceParam
        ):
            assert content is not None
            # Forced Function Call
            function_calls.append(
                FunctionCall(name=request.tool_choice.function.name, arguments=content)
            )
            content = None  # Clear content since tool is called.
        elif request.tool_choice == "required":
            assert content is not None
            tool_calls = TypeAdapter(list[FunctionDefinition]).validate_json(content)
            function_calls.extend(
                [
                    FunctionCall(
                        name=tool_call.name,
                        arguments=json.dumps(tool_call.parameters, ensure_ascii=False),
                    )
                    for tool_call in tool_calls
                ]
            )
            content = None  # Clear content since tool is called.
        elif (
            tool_parser_cls
            and enable_auto_tools
            and (request.tool_choice == "auto" or request.tool_choice is None)
        ):
1542
1543
1544
1545
1546
            if tokenizer is None:
                raise ValueError(
                    "Tokenizer not available when `skip_tokenizer_init=True`"
                )

1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
            # Automatic Tool Call Parsing
            try:
                tool_parser = tool_parser_cls(tokenizer)
            except RuntimeError as e:
                logger.exception("Error in tool parser creation.")
                raise e
            tool_call_info = tool_parser.extract_tool_calls(
                content if content is not None else "",
                request=request,  # type: ignore
            )
            if tool_call_info is not None and tool_call_info.tools_called:
                # extract_tool_calls() returns a list of tool calls.
                function_calls.extend(
                    FunctionCall(
                        name=tool_call.function.name,
                        arguments=tool_call.function.arguments,
                    )
                    for tool_call in tool_call_info.tool_calls
                )
                content = tool_call_info.content
1567
1568
                if content and content.strip() == "":
                    content = None
1569
1570
1571
1572
1573
1574
            else:
                # No tool calls.
                return None, content

        return function_calls, content

1575
    @staticmethod
1576
1577
1578
    def _get_decoded_token(
        logprob: Logprob,
        token_id: int,
1579
        tokenizer: TokenizerLike | None,
1580
1581
        return_as_token_id: bool = False,
    ) -> str:
1582
1583
1584
        if return_as_token_id:
            return f"token_id:{token_id}"

1585
1586
        if logprob.decoded_token is not None:
            return logprob.decoded_token
1587
1588
1589
1590
1591
1592

        if tokenizer is None:
            raise ValueError(
                "Unable to get tokenizer because `skip_tokenizer_init=True`"
            )

1593
        return tokenizer.decode(token_id)
1594

1595
    def _is_model_supported(self, model_name: str | None) -> bool:
1596
1597
        if not model_name:
            return True
1598
        return self.models.is_base_model(model_name)
1599

1600
1601

def clamp_prompt_logprobs(
1602
1603
    prompt_logprobs: PromptLogprobs | None,
) -> PromptLogprobs | None:
1604
1605
1606
1607
1608
1609
1610
    if prompt_logprobs is None:
        return prompt_logprobs

    for logprob_dict in prompt_logprobs:
        if logprob_dict is None:
            continue
        for logprob_values in logprob_dict.values():
1611
            if logprob_values.logprob == float("-inf"):
1612
1613
                logprob_values.logprob = -9999.0
    return prompt_logprobs