serving_chat.py 49.3 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
import asyncio
4
import json
5
import time
6
7
8
from collections.abc import AsyncGenerator, AsyncIterator
from collections.abc import Sequence as GenericSequence
from typing import Callable, Final, Optional, Union
9

10
import jinja2
11
from fastapi import Request
12

13
from vllm.config import ModelConfig
14
from vllm.engine.protocol import EngineClient
15
16
from vllm.entrypoints.chat_utils import (ChatTemplateContentFormatOption,
                                         ConversationMessage)
17
from vllm.entrypoints.logger import RequestLogger
18
from vllm.entrypoints.openai.protocol import (
19
20
    ChatCompletionLogProb, ChatCompletionLogProbs,
    ChatCompletionLogProbsContent, ChatCompletionNamedToolChoiceParam,
21
    ChatCompletionRequest, ChatCompletionResponse,
22
    ChatCompletionResponseChoice, ChatCompletionResponseStreamChoice,
23
    ChatCompletionStreamResponse, ChatMessage, DeltaFunctionCall, DeltaMessage,
24
25
    DeltaToolCall, ErrorResponse, FunctionCall, PromptTokenUsageInfo,
    RequestResponseMetadata, ToolCall, UsageInfo)
26
27
from vllm.entrypoints.openai.reasoning_parsers import (ReasoningParser,
                                                       ReasoningParserManager)
28
29
from vllm.entrypoints.openai.serving_engine import (OpenAIServing,
                                                    clamp_prompt_logprobs)
30
from vllm.entrypoints.openai.serving_models import OpenAIServingModels
31
from vllm.entrypoints.openai.tool_parsers import ToolParser, ToolParserManager
32
33
from vllm.entrypoints.openai.tool_parsers.mistral_tool_parser import (
    MistralToolCall)
34
from vllm.logger import init_logger
35
from vllm.outputs import CompletionOutput, RequestOutput
36
from vllm.sampling_params import BeamSearchParams, SamplingParams
37
from vllm.sequence import Logprob
38
from vllm.transformers_utils.tokenizer import AnyTokenizer, MistralTokenizer
39
40
from vllm.transformers_utils.tokenizers import (maybe_serialize_tool_calls,
                                                truncate_tool_call_ids)
41
42
43
44
45
46

logger = init_logger(__name__)


class OpenAIServingChat(OpenAIServing):

47
48
49
50
    def __init__(
        self,
        engine_client: EngineClient,
        model_config: ModelConfig,
51
        models: OpenAIServingModels,
52
53
54
55
56
57
        response_role: str,
        *,
        request_logger: Optional[RequestLogger],
        chat_template: Optional[str],
        chat_template_content_format: ChatTemplateContentFormatOption,
        return_tokens_as_token_ids: bool = False,
58
59
        enable_reasoning: bool = False,
        reasoning_parser: Optional[str] = None,
60
61
62
63
        enable_auto_tools: bool = False,
        tool_parser: Optional[str] = None,
        enable_prompt_tokens_details: bool = False,
    ) -> None:
64
        super().__init__(engine_client=engine_client,
65
                         model_config=model_config,
66
                         models=models,
67
68
                         request_logger=request_logger,
                         return_tokens_as_token_ids=return_tokens_as_token_ids)
69

70
        self.response_role = response_role
71
72
        self.chat_template = chat_template
        self.chat_template_content_format: Final = chat_template_content_format
73

74
75
76
77
78
79
80
81
        # set up tool use
        self.enable_auto_tools: bool = enable_auto_tools
        if self.enable_auto_tools:
            logger.info(
                "\"auto\" tool choice has been enabled please note that while"
                " the parallel_tool_calls client option is preset for "
                "compatibility reasons, it will be ignored.")

82
83
84
85
86
87
88
89
90
91
92
93
        self.enable_reasoning: bool = enable_reasoning
        self.reasoning_parser: Optional[Callable[[AnyTokenizer],
                                                 ReasoningParser]] = None
        if self.enable_reasoning:
            try:
                self.reasoning_parser = (
                    ReasoningParserManager.get_reasoning_parser(
                        reasoning_parser))
            except Exception as e:
                raise TypeError("Error: --enable-reasoning requires "
                                f"reasoning_parser:'{reasoning_parser}' "
                                "which has not been registered") from e
94
95
        self.tool_parser: Optional[Callable[[AnyTokenizer], ToolParser]] = None
        if self.enable_auto_tools:
96
            try:
97
98
99
100
101
                if (tool_parser == "pythonic" and
                        model_config.model.startswith("meta-llama/Llama-3.2")):
                    logger.warning(
                        "Llama3.2 models may struggle to emit valid pythonic"
                        " tool calls")
102
103
104
                self.tool_parser = ToolParserManager.get_tool_parser(
                    tool_parser)
            except Exception as e:
105
                raise TypeError("Error: --enable-auto-tool-choice requires "
106
107
                                f"tool_parser:'{tool_parser}' which has not "
                                "been registered") from e
108

109
        self.enable_prompt_tokens_details = enable_prompt_tokens_details
110
111
112
        self.default_sampling_params = (
            self.model_config.get_diff_sampling_param())
        if self.default_sampling_params:
113
114
115
116
            source = self.model_config.generation_config
            source = "model" if source == "auto" else source
            logger.info("Using default chat sampling params from %s: %s",
                        source, self.default_sampling_params)
117

118
    async def create_chat_completion(
119
120
        self,
        request: ChatCompletionRequest,
121
122
123
        raw_request: Optional[Request] = None,
    ) -> Union[AsyncGenerator[str, None], ChatCompletionResponse,
               ErrorResponse]:
124
125
        """
        Chat Completion API similar to OpenAI's API.
126

127
128
        See https://platform.openai.com/docs/api-reference/chat/create
        for the API specification. This API mimics the OpenAI
129
        Chat Completion API.
130
131
132
        """
        error_check_ret = await self._check_model(request)
        if error_check_ret is not None:
133
            logger.error("Error with model %s", error_check_ret)
134
135
            return error_check_ret

136
137
138
139
140
141
        # If the engine is dead, raise the engine's DEAD_ERROR.
        # This is required for the streaming case, where we return a
        # success status before we actually start generating text :).
        if self.engine_client.errored:
            raise self.engine_client.dead_error

142
        try:
143
144
145
146
147
            (
                lora_request,
                prompt_adapter_request,
            ) = self._maybe_get_adapters(request)

148
            model_name = self._get_model_name(request.model, lora_request)
149

150
            tokenizer = await self.engine_client.get_tokenizer(lora_request)
151

152
153
154
155
156
157
158
159
            tool_parser = self.tool_parser

            # validation for OpenAI tools
            # tool_choice = "required" is not supported
            if request.tool_choice == "required":
                return self.create_error_response(
                    "tool_choice = \"required\" is not supported!")

160
            if isinstance(tokenizer, MistralTokenizer):
161
162
163
                # because of issues with pydantic we need to potentially
                # re-serialize the tool_calls field of the request
                # for more info: see comment in `maybe_serialize_tool_calls`
164
                maybe_serialize_tool_calls(request)
165
                truncate_tool_call_ids(request)
166

167
168
169
170
171
172
173
174
175
            if (request.tool_choice == "auto" and
                    not (self.enable_auto_tools and tool_parser is not None)
                    and not isinstance(tokenizer, MistralTokenizer)):
                # for hf tokenizers, "auto" tools requires
                # --enable-auto-tool-choice and --tool-call-parser
                return self.create_error_response(
                    "\"auto\" tool choice requires "
                    "--enable-auto-tool-choice and --tool-call-parser to be set"
                )
176

177
178
179
180
            tool_dicts = None if request.tools is None else [
                tool.model_dump() for tool in request.tools
            ]

181
182
183
184
185
186
187
188
189
            (
                conversation,
                request_prompts,
                engine_prompts,
            ) = await self._preprocess_chat(
                request,
                tokenizer,
                request.messages,
                chat_template=request.chat_template or self.chat_template,
190
                chat_template_content_format=self.chat_template_content_format,
191
192
193
194
195
196
197
198
199
200
201
                add_generation_prompt=request.add_generation_prompt,
                continue_final_message=request.continue_final_message,
                tool_dicts=tool_dicts,
                documents=request.documents,
                chat_template_kwargs=request.chat_template_kwargs,
                tool_parser=tool_parser,
                truncate_prompt_tokens=request.truncate_prompt_tokens,
                add_special_tokens=request.add_special_tokens,
            )
        except ValueError as e:
            logger.exception("Error in preprocessing prompt inputs")
202
            return self.create_error_response(str(e))
203
204
205
206
207
208
209
210
211
        except TypeError as e:
            logger.exception("Error in preprocessing prompt inputs")
            return self.create_error_response(str(e))
        except RuntimeError as e:
            logger.exception("Error in preprocessing prompt inputs")
            return self.create_error_response(str(e))
        except jinja2.TemplateError as e:
            logger.exception("Error in preprocessing prompt inputs")
            return self.create_error_response(str(e))
212

213
214
        request_id = "chatcmpl-" \
                     f"{self._base_request_id(raw_request, request.request_id)}"
215
216
217
218
219

        request_metadata = RequestResponseMetadata(request_id=request_id)
        if raw_request:
            raw_request.state.request_metadata = request_metadata

220
        # Schedule the request and get the result generator.
221
        generators: list[AsyncGenerator[RequestOutput, None]] = []
222
        try:
223
224
225
226
227
228
            for i, engine_prompt in enumerate(engine_prompts):
                sampling_params: Union[SamplingParams, BeamSearchParams]
                default_max_tokens = self.max_model_len - len(
                    engine_prompt["prompt_token_ids"])
                if request.use_beam_search:
                    sampling_params = request.to_beam_search_params(
229
                        default_max_tokens, self.default_sampling_params)
230
231
                else:
                    sampling_params = request.to_sampling_params(
232
                        default_max_tokens,
233
                        self.model_config.logits_processor_pattern,
234
                        self.default_sampling_params)
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262

                self._log_inputs(request_id,
                                 request_prompts[i],
                                 params=sampling_params,
                                 lora_request=lora_request,
                                 prompt_adapter_request=prompt_adapter_request)

                trace_headers = (None if raw_request is None else await
                                 self._get_trace_headers(raw_request.headers))

                if isinstance(sampling_params, BeamSearchParams):
                    generator = self.engine_client.beam_search(
                        prompt=engine_prompt,
                        request_id=request_id,
                        params=sampling_params,
                    )
                else:
                    generator = self.engine_client.generate(
                        engine_prompt,
                        sampling_params,
                        request_id,
                        lora_request=lora_request,
                        trace_headers=trace_headers,
                        prompt_adapter_request=prompt_adapter_request,
                        priority=request.priority,
                    )

                generators.append(generator)
263
        except ValueError as e:
264
            # TODO: Use a vllm-specific Validation Error
265
266
            return self.create_error_response(str(e))

267
268
269
        assert len(generators) == 1
        result_generator, = generators

270
271
272
        # Streaming response
        if request.stream:
            return self.chat_completion_stream_generator(
273
274
                request, result_generator, request_id, model_name,
                conversation, tokenizer, request_metadata)
275

276
277
        try:
            return await self.chat_completion_full_generator(
278
279
                request, result_generator, request_id, model_name,
                conversation, tokenizer, request_metadata)
280
281
282
        except ValueError as e:
            # TODO: Use a vllm-specific Validation Error
            return self.create_error_response(str(e))
283
284
285
286

    def get_chat_request_role(self, request: ChatCompletionRequest) -> str:
        if request.add_generation_prompt:
            return self.response_role
287
        return request.messages[-1]["role"]
288
289

    async def chat_completion_stream_generator(
290
291
292
293
        self,
        request: ChatCompletionRequest,
        result_generator: AsyncIterator[RequestOutput],
        request_id: str,
294
        model_name: str,
295
        conversation: list[ConversationMessage],
296
        tokenizer: AnyTokenizer,
297
        request_metadata: RequestResponseMetadata,
298
    ) -> AsyncGenerator[str, None]:
299
        created_time = int(time.time())
300
        chunk_object_type: Final = "chat.completion.chunk"
301
        first_iteration = True
302
303

        # Send response for each token for each request.n (index)
304
305
306
        num_choices = 1 if request.n is None else request.n
        previous_num_tokens = [0] * num_choices
        finish_reason_sent = [False] * num_choices
307
        num_prompt_tokens = 0
308
        num_cached_tokens = None
309
310
311
312
313
314
315
316
317
318
319

        if isinstance(request.tool_choice, ChatCompletionNamedToolChoiceParam):
            tool_choice_function_name = request.tool_choice.function.name
        else:
            tool_choice_function_name = None

        # Determine whether tools are in use with "auto" tool choice
        tool_choice_auto = (
            not tool_choice_function_name
            and self._should_stream_with_auto_tool_parsing(request))

320
321
322
        should_stream_with_reasoning_parsing = (
            self._should_stream_with_reasoning_parsing(request))

323
        all_previous_token_ids: Optional[list[list[int]]]
324
325
326
327

        # Only one of these will be used, thus previous_texts and
        # all_previous_token_ids will not be used twice in the same iteration.
        if tool_choice_auto or should_stream_with_reasoning_parsing:
328
329
330
            # These are only required in "auto" tool choice case
            previous_texts = [""] * num_choices
            all_previous_token_ids = [[]] * num_choices
331
332
333
            # For reasoning parser and tool call all enabled
            added_content_delta_arr = [False] * num_choices
            reasoning_end_arr = [False] * num_choices
334
335
336
        else:
            previous_texts, all_previous_token_ids = None, None

337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
        try:
            # There is no need to check if the reasoning_parser is None
            # because the should_stream_with_reasoning_parsing check
            # already ensures that the reasoning_parser is not None.
            # but the pre-commit hook requires it.
            if should_stream_with_reasoning_parsing and \
                self.reasoning_parser is not None:
                reasoning_parser = self.reasoning_parser(tokenizer)
        except RuntimeError as e:
            logger.exception("Error in reasoning parser creation.")
            data = self.create_streaming_error_response(str(e))
            yield f"data: {data}\n\n"
            yield "data: [DONE]\n\n"
            return

352
353
354
        # Prepare the tool parser if it's needed
        try:
            if tool_choice_auto and self.tool_parser:
355
                tool_parsers: list[Optional[ToolParser]] = [
356
357
358
359
                    self.tool_parser(tokenizer)
                ] * num_choices
            else:
                tool_parsers = [None] * num_choices
360
        except Exception as e:
361
            logger.exception("Error in tool parser creation.")
362
363
364
365
366
            data = self.create_streaming_error_response(str(e))
            yield f"data: {data}\n\n"
            yield "data: [DONE]\n\n"
            return

367
368
369
370
371
372
373
374
        stream_options = request.stream_options
        if stream_options:
            include_usage = stream_options.include_usage
            include_continuous_usage = include_usage and \
                                       stream_options.continuous_usage_stats
        else:
            include_usage, include_continuous_usage = False, False

375
376
        try:
            async for res in result_generator:
377
378
                if res.prompt_token_ids is not None:
                    num_prompt_tokens = len(res.prompt_token_ids)
379
380
                    if res.encoder_prompt_token_ids is not None:
                        num_prompt_tokens += len(res.encoder_prompt_token_ids)
381

382
383
384
385
                # We need to do it here, because if there are exceptions in
                # the result_generator, it needs to be sent as the FIRST
                # response (by the try...catch).
                if first_iteration:
386
                    num_cached_tokens = res.num_cached_tokens
387
388
                    # Send first response for each request.n (index) with
                    # the role
389
                    role = self.get_chat_request_role(request)
390
391
392

                    # NOTE num_choices defaults to 1 so this usually executes
                    # once per request
393
                    for i in range(num_choices):
394
395
                        choice_data = ChatCompletionResponseStreamChoice(
                            index=i,
396
397
398
399
                            delta=DeltaMessage(
                                role=role,
                                content="",
                            ),
400
401
402
403
404
405
406
407
                            logprobs=None,
                            finish_reason=None)
                        chunk = ChatCompletionStreamResponse(
                            id=request_id,
                            object=chunk_object_type,
                            created=created_time,
                            choices=[choice_data],
                            model=model_name)
408

409
410
411
412
413
414
                        # if continuous usage stats are requested, add it
                        if include_continuous_usage:
                            chunk.usage = UsageInfo(
                                prompt_tokens=num_prompt_tokens,
                                completion_tokens=0,
                                total_tokens=num_prompt_tokens)
415

416
417
418
                        data = chunk.model_dump_json(exclude_unset=True)
                        yield f"data: {data}\n\n"

419
420
                    # Send response to echo the input portion of the
                    # last message
421
                    if request.echo:
422
                        last_msg_content: Union[str, list[dict[str, str]]] = ""
423
424
425
                        if conversation and "content" in conversation[
                                -1] and conversation[-1].get("role") == role:
                            last_msg_content = conversation[-1]["content"] or ""
426
427

                        if last_msg_content:
428
                            for i in range(num_choices):
429
430
431
432
433
                                choice_data = (
                                    ChatCompletionResponseStreamChoice(
                                        index=i,
                                        delta=DeltaMessage(
                                            content=last_msg_content),
434
                                        logprobs=None,
435
                                        finish_reason=None))
436
437
438
439
440
441
                                chunk = ChatCompletionStreamResponse(
                                    id=request_id,
                                    object=chunk_object_type,
                                    created=created_time,
                                    choices=[choice_data],
                                    model=model_name)
442
443
444
445
446
                                if include_continuous_usage:
                                    chunk.usage = UsageInfo(
                                        prompt_tokens=num_prompt_tokens,
                                        completion_tokens=0,
                                        total_tokens=num_prompt_tokens)
447

448
449
450
451
452
453
454
                                data = chunk.model_dump_json(
                                    exclude_unset=True)
                                yield f"data: {data}\n\n"
                    first_iteration = False

                for output in res.outputs:
                    i = output.index
455
                    tool_parser = tool_parsers[i]
456
457
458
459

                    if finish_reason_sent[i]:
                        continue

460
                    if request.logprobs and request.top_logprobs is not None:
461
                        assert output.logprobs is not None, (
462
                            "Did not output logprobs")
463
                        logprobs = self._create_chat_logprobs(
464
465
                            token_ids=output.token_ids,
                            top_logprobs=output.logprobs,
466
                            tokenizer=tokenizer,
467
                            num_output_top_logprobs=request.top_logprobs,
468
469
                            return_as_token_id=request.
                            return_tokens_as_token_ids,
470
471
472
473
                        )
                    else:
                        logprobs = None

474
                    delta_text = output.text
475
476
477
478
479
480

                    if not delta_text and not output.token_ids and \
                        not previous_num_tokens[i]:
                        # Chunked prefill case, don't return empty chunks
                        continue

481
                    delta_message: Optional[DeltaMessage]
482

483
484
                    # just update previous_texts and previous_token_ids
                    if tool_choice_auto or should_stream_with_reasoning_parsing:
485
486
487
488
489
490
491
492
                        assert previous_texts is not None
                        assert all_previous_token_ids is not None
                        previous_text = previous_texts[i]
                        previous_token_ids = all_previous_token_ids[i]
                        current_text = previous_text + delta_text
                        current_token_ids = previous_token_ids + list(
                            output.token_ids)

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
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
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
                    # handle streaming deltas for tools with named tool_choice
                    if tool_choice_function_name:
                        if (self.enable_reasoning
                                and not reasoning_parser.is_reasoning_end(
                                    previous_token_ids)):
                            assert reasoning_parser is not None
                            delta_message = (
                                reasoning_parser.
                                extract_reasoning_content_streaming(
                                    previous_text,
                                    current_text,
                                    delta_text,
                                    previous_token_ids,
                                    current_token_ids,
                                    output.token_ids,
                                ))
                            # When encountering think end id in delta_token_ids,
                            # process the `content`. Only keep 'content',
                            # remove 'reasoning_content'
                            if reasoning_parser.is_reasoning_end(
                                    list(output.token_ids)):
                                if delta_message and delta_message.content:
                                    # This need to be added to next `delta_text`
                                    current_text = delta_message.content
                                    delta_message.content = None
                                else:
                                    current_text = ""
                        else:
                            # Just to add remaining `content`
                            if self.enable_reasoning:
                                delta_text = previous_text + delta_text
                                current_text = ""

                            delta_message = DeltaMessage(tool_calls=[
                                DeltaToolCall(function=DeltaFunctionCall(
                                    name=tool_choice_function_name,
                                    arguments=delta_text),
                                              index=i)
                            ])

                    # handle streaming deltas for tools with "auto" tool choice
                    # and reasoning parser
                    elif tool_choice_auto and self.enable_reasoning:
                        assert tool_parser is not None
                        assert reasoning_parser is not None
                        assert added_content_delta_arr is not None
                        assert reasoning_end_arr is not None
                        if not reasoning_end_arr[i]:
                            delta_message = (
                                reasoning_parser.
                                extract_reasoning_content_streaming(
                                    previous_text,
                                    current_text,
                                    delta_text,
                                    previous_token_ids,
                                    current_token_ids,
                                    output.token_ids,
                                ))

                            # When encountering think end id in delta_token_ids,
                            # set reasoning status to end.
                            # Remove the text and token ids related
                            # to 'reasoning_content'.
                            if reasoning_parser.is_reasoning_end(
                                    list(output.token_ids)):
                                reasoning_end_arr[i] = True
                                current_token_ids =  \
                                    reasoning_parser.extract_content_ids(
                                        list(output.token_ids))
                                if delta_message and delta_message.content:
                                    current_text = delta_message.content
                                    delta_message.content = None
                                else:
                                    current_text = ""

                        # handle tool calls only after reasoning is done,
                        else:
                            delta_token_ids = list(output.token_ids)
                            # First time to tool call,
                            # add the remaining text and token ids
                            # to delta from previous
                            if not added_content_delta_arr[i]:
                                added_content_delta_arr[i] = True
                                previous_text = ""
                                previous_token_ids = []
                                delta_text = current_text
                                delta_token_ids = current_token_ids

                            delta_message = (
                                tool_parser.extract_tool_calls_streaming(
                                    previous_text=previous_text,
                                    current_text=current_text,
                                    delta_text=delta_text,
                                    previous_token_ids=previous_token_ids,
                                    current_token_ids=current_token_ids,
                                    delta_token_ids=delta_token_ids,
                                    request=request))
                    # when only tool calls
                    elif tool_choice_auto:
                        assert tool_parser is not None
593
594
                        delta_message = (
                            tool_parser.extract_tool_calls_streaming(
595
596
                                previous_text=previous_text,
                                current_text=current_text,
597
                                delta_text=delta_text,
598
599
                                previous_token_ids=previous_token_ids,
                                current_token_ids=current_token_ids,
600
601
                                delta_token_ids=output.token_ids,
                                request=request))
602
                    # when only reasoning
603
604
605
606
607
608
609
610
611
612
613
                    elif self.enable_reasoning:
                        assert reasoning_parser is not None
                        delta_message = (reasoning_parser.
                                         extract_reasoning_content_streaming(
                                             previous_text,
                                             current_text,
                                             delta_text,
                                             previous_token_ids,
                                             current_token_ids,
                                             output.token_ids,
                                         ))
614
                    # handle streaming just a content delta
615
616
617
                    else:
                        delta_message = DeltaMessage(content=delta_text)

618
619
620
621
622
623
624
                    # update the previous values for the next iteration
                    if tool_choice_auto or should_stream_with_reasoning_parsing:
                        assert previous_texts is not None
                        assert all_previous_token_ids is not None
                        previous_texts[i] = current_text
                        all_previous_token_ids[i] = current_token_ids

625
                    # set the previous values for the next iteration
626
                    previous_num_tokens[i] += len(output.token_ids)
627
628
629
630
631
632
633
634

                    # if the message delta is None (e.g. because it was a
                    # "control token" for tool calls or the parser otherwise
                    # wasn't ready to send a token, then
                    #   get the next token without streaming a chunk
                    if delta_message is None:
                        continue

635
636
637
638
                    if output.finish_reason is None:
                        # Send token-by-token response for each request.n
                        choice_data = ChatCompletionResponseStreamChoice(
                            index=i,
639
                            delta=delta_message,
640
641
                            logprobs=logprobs,
                            finish_reason=None)
642
643

                    # if the model is finished generating
644
                    else:
645
646
647
648
                        # check to make sure we haven't "forgotten" to stream
                        #   any tokens that were generated but previously
                        #   matched by partial json parsing
                        # only happens if we are NOT using guided decoding
649
                        auto_tools_called = False
650
                        if tool_parser:
651
652
653
654
                            auto_tools_called = len(
                                tool_parser.prev_tool_call_arr) > 0
                            index = len(tool_parser.prev_tool_call_arr
                                        ) - 1 if auto_tools_called else 0
655
656
657
658
659
                        else:
                            index = 0

                        if self._should_check_for_unstreamed_tool_arg_tokens(
                                delta_message, output) and tool_parser:
660
661
662
663
664
665
666
667
668
669
                            latest_delta_len = 0
                            if ((isinstance(
                                    delta_message.tool_calls[0].function,
                                    DeltaFunctionCall)) and isinstance(
                                        delta_message.tool_calls[0].function.
                                        arguments, str)):
                                latest_delta_len = len(
                                    delta_message.tool_calls[0].function.
                                    arguments)

670
671
672
673
                            # get the expected call based on partial JSON
                            # parsing which "autocompletes" the JSON
                            expected_call = json.dumps(
                                tool_parser.prev_tool_call_arr[index].get(
674
675
                                    "arguments", {}),
                                ensure_ascii=False)
676

677
                            # get what we've streamed so far for arguments
678
679
680
                            # for the current tool
                            actual_call = tool_parser.streamed_args_for_tool[
                                index]
681
682
                            if (latest_delta_len > 0):
                                actual_call = actual_call[:-latest_delta_len]
683
684
685
686
687
688
689
690
691
692
693
694

                            # check to see if there's anything left to stream
                            remaining_call = expected_call.replace(
                                actual_call, "", 1)
                            # set that as a delta message
                            delta_message = DeltaMessage(tool_calls=[
                                DeltaToolCall(index=index,
                                              function=DeltaFunctionCall(
                                                  arguments=remaining_call).
                                              model_dump(exclude_none=True))
                            ])

695
696
697
                        # Send the finish response for each request.n only once
                        choice_data = ChatCompletionResponseStreamChoice(
                            index=i,
698
                            delta=delta_message,
699
                            logprobs=logprobs,
700
                            finish_reason=output.finish_reason
701
                            if not auto_tools_called else "tool_calls",
702
                            stop_reason=output.stop_reason)
703

704
                        finish_reason_sent[i] = True
705

706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
                    chunk = ChatCompletionStreamResponse(
                        id=request_id,
                        object=chunk_object_type,
                        created=created_time,
                        choices=[choice_data],
                        model=model_name)

                    # handle usage stats if requested & if continuous
                    if include_continuous_usage:
                        completion_tokens = previous_num_tokens[i]
                        chunk.usage = UsageInfo(
                            prompt_tokens=num_prompt_tokens,
                            completion_tokens=completion_tokens,
                            total_tokens=num_prompt_tokens + completion_tokens,
                        )

                    data = chunk.model_dump_json(exclude_unset=True)
                    yield f"data: {data}\n\n"

725
726
            # once the final token is handled, if stream_options.include_usage
            # is sent, send the usage
727
728
            if include_usage:
                completion_tokens = sum(previous_num_tokens)
729
730
731
732
733
734
735
                final_usage = UsageInfo(prompt_tokens=num_prompt_tokens,
                                        completion_tokens=completion_tokens,
                                        total_tokens=num_prompt_tokens +
                                        completion_tokens)
                if self.enable_prompt_tokens_details and num_cached_tokens:
                    final_usage.prompt_tokens_details = PromptTokenUsageInfo(
                        cached_tokens=num_cached_tokens)
736
737
738
739
740
741
742
743
744
745
746

                final_usage_chunk = ChatCompletionStreamResponse(
                    id=request_id,
                    object=chunk_object_type,
                    created=created_time,
                    choices=[],
                    model=model_name,
                    usage=final_usage)
                final_usage_data = (final_usage_chunk.model_dump_json(
                    exclude_unset=True, exclude_none=True))
                yield f"data: {final_usage_data}\n\n"
747

748
749
750
751
752
753
754
            # report to FastAPI middleware aggregate usage across all choices
            num_completion_tokens = sum(previous_num_tokens)
            request_metadata.final_usage_info = UsageInfo(
                prompt_tokens=num_prompt_tokens,
                completion_tokens=num_completion_tokens,
                total_tokens=num_prompt_tokens + num_completion_tokens)

755
        except Exception as e:
756
            # TODO: Use a vllm-specific Validation Error
757
            logger.exception("Error in chat completion stream generator.")
758
759
            data = self.create_streaming_error_response(str(e))
            yield f"data: {data}\n\n"
760
761
762
763
        # Send the final done message after all response.n are finished
        yield "data: [DONE]\n\n"

    async def chat_completion_full_generator(
764
765
766
767
        self,
        request: ChatCompletionRequest,
        result_generator: AsyncIterator[RequestOutput],
        request_id: str,
768
        model_name: str,
769
        conversation: list[ConversationMessage],
770
        tokenizer: AnyTokenizer,
771
        request_metadata: RequestResponseMetadata,
772
    ) -> Union[ErrorResponse, ChatCompletionResponse]:
773

774
        created_time = int(time.time())
775
        final_res: Optional[RequestOutput] = None
776

777
778
779
780
781
        try:
            async for res in result_generator:
                final_res = res
        except asyncio.CancelledError:
            return self.create_error_response("Client disconnected")
782
783
784
        except ValueError as e:
            # TODO: Use a vllm-specific Validation Error
            return self.create_error_response(str(e))
785

786
787
        assert final_res is not None

788
        choices: list[ChatCompletionResponseChoice] = []
789

790
791
        role = self.get_chat_request_role(request)
        for output in final_res.outputs:
792
            token_ids = output.token_ids
793
            out_logprobs = output.logprobs
794

795
796
            if request.logprobs and request.top_logprobs is not None:
                assert out_logprobs is not None, "Did not output logprobs"
797
                logprobs = self._create_chat_logprobs(
798
                    token_ids=token_ids,
799
                    top_logprobs=out_logprobs,
800
                    num_output_top_logprobs=request.top_logprobs,
801
                    tokenizer=tokenizer,
802
                    return_as_token_id=request.return_tokens_as_token_ids,
803
804
805
806
                )
            else:
                logprobs = None

807
808
809
            should_stream_with_reasoning_parsing = (
                self._should_stream_with_reasoning_parsing(request))

810
811
812
813
            # In the OpenAI API the finish_reason is "tools_called"
            # if the tool choice is auto and the model produced a tool
            # call. The same is not true for named function calls
            auto_tools_called = False
814

815
816
817
818
819
820
821
            if should_stream_with_reasoning_parsing and \
                self.reasoning_parser is not None:
                try:
                    reasoning_parser = self.reasoning_parser(tokenizer)
                except RuntimeError as e:
                    logger.exception("Error in reasoning parser creation.")
                    return self.create_error_response(str(e))
822
823
                # If the reasoning parser is enabled,
                # tool calls are extracted exclusively from the content.
824
825
826
                reasoning_content, content = (
                    reasoning_parser.extract_reasoning_content(
                        output.text, request=request))
827
828
829
            else:
                reasoning_content = None
                content = output.text
830

831
832
            # if auto tools are not enabled, and a named tool choice using
            #   outlines is not being used
833
834
835
836
837
838
839
            if (not self.enable_auto_tools
                    or not self.tool_parser) and not isinstance(
                        request.tool_choice,
                        ChatCompletionNamedToolChoiceParam):
                message = ChatMessage(role=role,
                                      reasoning_content=reasoning_content,
                                      content=content)
840
841
842

            # if the request uses tools and specified a tool choice
            elif request.tool_choice and type(
843
                    request.tool_choice) is ChatCompletionNamedToolChoiceParam:
844

845
846
                tool_call_class = MistralToolCall if isinstance(
                    tokenizer, MistralTokenizer) else ToolCall
847
848
                message = ChatMessage(
                    role=role,
849
                    reasoning_content=reasoning_content,
850
851
                    content="",
                    tool_calls=[
852
                        tool_call_class(function=FunctionCall(
853
                            name=request.tool_choice.function.name,
854
                            arguments=content))
855
                    ])
856
857
858

            # if the request doesn't use tool choice
            # OR specifies to not use a tool
859
            elif not request.tool_choice or request.tool_choice == "none":
860

861
862
863
                message = ChatMessage(role=role,
                                      reasoning_content=reasoning_content,
                                      content=content)
864
865
866
867
868
869
870

            # handle when there are tools and tool choice is auto
            elif request.tools and (
                    request.tool_choice == "auto"
                    or request.tool_choice is None) and self.enable_auto_tools \
                    and self.tool_parser:

871
872
873
                try:
                    tool_parser = self.tool_parser(tokenizer)
                except RuntimeError as e:
874
                    logger.exception("Error in tool parser creation.")
875
876
                    return self.create_error_response(str(e))

877
                tool_call_info = tool_parser.extract_tool_calls(
878
                    content if content is not None else "", request=request)
879
880
881
882
                # In the OpenAI API the finish_reason is "tools_called"
                # if the tool choice is auto and the model produced a tool
                # call. The same is not true for named function calls
                auto_tools_called = tool_call_info.tools_called
883
884
                if tool_call_info.tools_called:
                    message = ChatMessage(role=role,
885
                                          reasoning_content=reasoning_content,
886
887
888
889
890
891
                                          content=tool_call_info.content,
                                          tool_calls=tool_call_info.tool_calls)

                else:
                    # FOR NOW make it a chat message; we will have to detect
                    # the type to make it later.
892
893
894
                    message = ChatMessage(role=role,
                                          reasoning_content=reasoning_content,
                                          content=content)
895
896
897
898
899
900
901

            # undetermined case that is still important to handle
            else:
                logger.error(
                    "Error in chat_completion_full_generator - cannot determine"
                    " if tools should be extracted. Returning a standard chat "
                    "completion.")
902
903
904
                message = ChatMessage(role=role,
                                      reasoning_content=reasoning_content,
                                      content=content)
905

906
907
            choice_data = ChatCompletionResponseChoice(
                index=output.index,
908
                message=message,
909
                logprobs=logprobs,
910
                finish_reason="tool_calls" if auto_tools_called else
911
                output.finish_reason if output.finish_reason else "stop",
912
                stop_reason=output.stop_reason)
913
914
            choices.append(choice_data)

915
        if request.echo:
916
            last_msg_content: Union[str, list[dict[str, str]]] = ""
917
918
            if conversation and "content" in conversation[-1] and conversation[
                    -1].get("role") == role:
919
                last_msg_content = conversation[-1]["content"] or ""
920
921
922
            if isinstance(last_msg_content, list):
                last_msg_content = "\n".join(msg['text']
                                             for msg in last_msg_content)
923
924

            for choice in choices:
925
926
                full_message = last_msg_content + (choice.message.content
                                                   or "")
927
928
                choice.message.content = full_message

929
        assert final_res.prompt_token_ids is not None
930
        num_prompt_tokens = len(final_res.prompt_token_ids)
931
932
        if final_res.encoder_prompt_token_ids is not None:
            num_prompt_tokens += len(final_res.encoder_prompt_token_ids)
933
934
        num_generated_tokens = sum(
            len(output.token_ids) for output in final_res.outputs)
935
936
937
938
939
940
941
        usage = UsageInfo(prompt_tokens=num_prompt_tokens,
                          completion_tokens=num_generated_tokens,
                          total_tokens=num_prompt_tokens +
                          num_generated_tokens)
        if self.enable_prompt_tokens_details and final_res.num_cached_tokens:
            usage.prompt_tokens_details = PromptTokenUsageInfo(
                cached_tokens=final_res.num_cached_tokens)
942
943
944

        request_metadata.final_usage_info = usage

945
946
947
948
949
950
        response = ChatCompletionResponse(
            id=request_id,
            created=created_time,
            model=model_name,
            choices=choices,
            usage=usage,
951
            prompt_logprobs=clamp_prompt_logprobs(final_res.prompt_logprobs),
952
953
        )

954
        return response
955
956

    def _get_top_logprobs(
957
            self, logprobs: dict[int, Logprob], top_logprobs: Optional[int],
958
959
            tokenizer: AnyTokenizer,
            should_return_as_token_id: bool) -> list[ChatCompletionLogProb]:
960
        return [
961
962
963
964
            ChatCompletionLogProb(token=(token := self._get_decoded_token(
                p[1],
                p[0],
                tokenizer,
965
                return_as_token_id=should_return_as_token_id)),
966
967
968
                                  logprob=max(p[1].logprob, -9999.0),
                                  bytes=list(
                                      token.encode("utf-8", errors="replace")))
969
970
971
972
973
974
975
            for i, p in enumerate(logprobs.items())
            if top_logprobs and i < top_logprobs
        ]

    def _create_chat_logprobs(
        self,
        token_ids: GenericSequence[int],
976
        top_logprobs: GenericSequence[Optional[dict[int, Logprob]]],
977
        tokenizer: AnyTokenizer,
978
        num_output_top_logprobs: Optional[int] = None,
979
        return_as_token_id: Optional[bool] = None,
980
981
    ) -> ChatCompletionLogProbs:
        """Create OpenAI-style logprobs."""
982
        logprobs_content: list[ChatCompletionLogProbsContent] = []
983

984
985
        should_return_as_token_id = return_as_token_id if \
            return_as_token_id is not None else self.return_tokens_as_token_ids
986
987
988
        for i, token_id in enumerate(token_ids):
            step_top_logprobs = top_logprobs[i]
            if step_top_logprobs is None:
989
                token = tokenizer.decode(token_id)
990
                if should_return_as_token_id:
991
                    token = f"token_id:{token_id}"
992

993
994
                logprobs_content.append(
                    ChatCompletionLogProbsContent(
995
                        token=token,
996
997
                        bytes=list(token.encode("utf-8", errors="replace")),
                    ))
998
            else:
999
1000
1001
                step_token = step_top_logprobs[token_id]
                step_decoded = step_token.decoded_token

1002
1003
                logprobs_content.append(
                    ChatCompletionLogProbsContent(
1004
                        token=self._get_decoded_token(
1005
1006
1007
                            step_token,
                            token_id,
                            tokenizer,
1008
                            should_return_as_token_id,
1009
1010
1011
1012
                        ),
                        logprob=max(step_token.logprob, -9999.0),
                        bytes=None if step_decoded is None else list(
                            step_decoded.encode("utf-8", errors="replace")),
1013
                        top_logprobs=self._get_top_logprobs(
1014
1015
                            step_top_logprobs, num_output_top_logprobs,
                            tokenizer, should_return_as_token_id),
1016
                    ))
1017
1018

        return ChatCompletionLogProbs(content=logprobs_content)
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032

    def _should_stream_with_auto_tool_parsing(self,
                                              request: ChatCompletionRequest):
        """
        Utility function to check if streamed tokens should go through the tool
        call parser that was configured.

        We only want to do this IF user-provided tools are set, a tool parser
        is configured, "auto" tool choice is enabled, and the request's tool
        choice field indicates that "auto" tool choice should be used.
        """
        return (request.tools and self.tool_parser and self.enable_auto_tools
                and request.tool_choice in ['auto', None])

1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
    def _should_stream_with_reasoning_parsing(self,
                                              request: ChatCompletionRequest):
        """
            Utility function to check if streamed tokens should go through the
            reasoning parser that was configured.
    
            We only want to do this IF reasoning is enabled and a reasoning 
            parser is configured.
            """
        return self.enable_reasoning and self.reasoning_parser is not None

1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
    def _should_check_for_unstreamed_tool_arg_tokens(
        self,
        delta_message: Optional[DeltaMessage],
        output: CompletionOutput,
    ) -> bool:
        """
        Check to see if we should check for unstreamed tool arguments tokens.
        This is only applicable when auto tool parsing is enabled, the delta
        is a tool call with arguments.
        """

        # yapf: disable
        return bool(
            # if there is a delta message that includes tool calls which
            # include a function that has arguments
1059
1060
            output.finish_reason is not None
            and self.enable_auto_tools and self.tool_parser and delta_message
1061
1062
1063
1064
            and delta_message.tool_calls and delta_message.tool_calls[0]
            and delta_message.tool_calls[0].function
            and delta_message.tool_calls[0].function.arguments is not None
        )