serving_chat.py 70.2 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import asyncio
5
import json
6
import time
7
8
9
from collections.abc import AsyncGenerator, AsyncIterator
from collections.abc import Sequence as GenericSequence
from typing import Callable, Final, Optional, Union
10

11
import jinja2
12
import partial_json_parser
13
import regex as re
14
from fastapi import Request
15
from openai_harmony import Message as OpenAIMessage
16
from pydantic import TypeAdapter
17

18
from vllm.config import ModelConfig
19
from vllm.engine.protocol import EngineClient
20
from vllm.entrypoints.chat_utils import (ChatTemplateContentFormatOption,
21
                                         ConversationMessage,
22
23
                                         get_history_tool_calls_cnt,
                                         make_tool_call_id)
24
25
26
27
from vllm.entrypoints.harmony_utils import (
    get_developer_message, get_stop_tokens_for_assistant_actions,
    get_streamable_parser_for_assistant, get_system_message, parse_chat_input,
    parse_chat_output, render_for_completion)
28
from vllm.entrypoints.logger import RequestLogger
29
from vllm.entrypoints.openai.protocol import (
30
31
    ChatCompletionLogProb, ChatCompletionLogProbs,
    ChatCompletionLogProbsContent, ChatCompletionNamedToolChoiceParam,
32
    ChatCompletionRequest, ChatCompletionResponse,
33
    ChatCompletionResponseChoice, ChatCompletionResponseStreamChoice,
34
    ChatCompletionStreamResponse, ChatMessage, DeltaFunctionCall, DeltaMessage,
35
36
    DeltaToolCall, ErrorResponse, FunctionCall, FunctionDefinition,
    PromptTokenUsageInfo, RequestResponseMetadata, ToolCall, UsageInfo)
37
38
from vllm.entrypoints.openai.serving_engine import (OpenAIServing,
                                                    clamp_prompt_logprobs)
39
from vllm.entrypoints.openai.serving_models import OpenAIServingModels
40
from vllm.entrypoints.openai.tool_parsers import ToolParser, ToolParserManager
41
42
from vllm.entrypoints.openai.tool_parsers.mistral_tool_parser import (
    MistralToolCall)
43
from vllm.entrypoints.utils import get_max_tokens
44
from vllm.inputs.data import TokensPrompt as EngineTokensPrompt
45
from vllm.logger import init_logger
46
from vllm.outputs import CompletionOutput, RequestOutput
47
from vllm.reasoning import ReasoningParser, ReasoningParserManager
48
from vllm.sampling_params import BeamSearchParams, SamplingParams
49
from vllm.sequence import Logprob
50
from vllm.transformers_utils.tokenizer import AnyTokenizer, MistralTokenizer
51
from vllm.transformers_utils.tokenizers import (maybe_serialize_tool_calls,
52
53
                                                truncate_tool_call_ids,
                                                validate_request_params)
54
from vllm.utils import as_list
55
56
57
58
59
60

logger = init_logger(__name__)


class OpenAIServingChat(OpenAIServing):

61
62
63
64
    def __init__(
        self,
        engine_client: EngineClient,
        model_config: ModelConfig,
65
        models: OpenAIServingModels,
66
67
68
69
70
71
        response_role: str,
        *,
        request_logger: Optional[RequestLogger],
        chat_template: Optional[str],
        chat_template_content_format: ChatTemplateContentFormatOption,
        return_tokens_as_token_ids: bool = False,
72
        reasoning_parser: str = "",
73
        enable_auto_tools: bool = False,
74
        exclude_tools_when_tool_choice_none: bool = False,
75
76
        tool_parser: Optional[str] = None,
        enable_prompt_tokens_details: bool = False,
77
        enable_force_include_usage: bool = False,
78
        enable_log_outputs: bool = False,
79
        log_error_stack: bool = False,
80
    ) -> None:
81
        super().__init__(engine_client=engine_client,
82
                         model_config=model_config,
83
                         models=models,
84
                         request_logger=request_logger,
85
                         return_tokens_as_token_ids=return_tokens_as_token_ids,
86
87
                         enable_force_include_usage=enable_force_include_usage,
                         log_error_stack=log_error_stack)
88

89
        self.response_role = response_role
90
91
        self.chat_template = chat_template
        self.chat_template_content_format: Final = chat_template_content_format
92
        self.enable_log_outputs = enable_log_outputs
93

94
95
96
97
98
99
100
101
        # 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.")

102
103
        self.reasoning_parser: Optional[Callable[[AnyTokenizer],
                                                 ReasoningParser]] = None
104
        if reasoning_parser:
105
106
107
108
            try:
                self.reasoning_parser = (
                    ReasoningParserManager.get_reasoning_parser(
                        reasoning_parser))
109
                assert self.reasoning_parser is not None
110
            except Exception as e:
111
112
                raise TypeError(
                    f"{reasoning_parser=} has not been registered") from e
113
114
        self.tool_parser: Optional[Callable[[AnyTokenizer], ToolParser]] = None
        if self.enable_auto_tools:
115
            try:
116
117
118
119
120
                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")
121
122
123
                self.tool_parser = ToolParserManager.get_tool_parser(
                    tool_parser)
            except Exception as e:
124
                raise TypeError("Error: --enable-auto-tool-choice requires "
125
126
                                f"tool_parser:'{tool_parser}' which has not "
                                "been registered") from e
127
128
        self.exclude_tools_when_tool_choice_none = (
            exclude_tools_when_tool_choice_none)
129

130
        self.enable_prompt_tokens_details = enable_prompt_tokens_details
131
        self.enable_force_include_usage = enable_force_include_usage
132
133
134
        self.default_sampling_params = (
            self.model_config.get_diff_sampling_param())
        if self.default_sampling_params:
135
136
137
138
            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)
139
140
141
142
        if self.model_config.hf_config.model_type == 'kimi_k2':
            self.tool_call_id_type = 'kimi_k2'
        else:
            self.tool_call_id_type = 'random'
143

144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
        self.use_harmony = model_config.hf_config.model_type == "gpt_oss"
        if self.use_harmony:
            if "stop_token_ids" not in self.default_sampling_params:
                self.default_sampling_params["stop_token_ids"] = []
            self.default_sampling_params["stop_token_ids"].extend(
                get_stop_tokens_for_assistant_actions())

        # NOTE(woosuk): While OpenAI's chat completion API supports browsing
        # for some models, currently vLLM doesn't support it. Please use the
        # Responses API instead.
        self.supports_browsing = False
        self.browser_tool = None
        # NOTE(woosuk): Chat completion API does not support code interpreter.
        # Please use the Responses API instead.
        self.supports_code_interpreter = False
        self.python_tool = None

161
    async def create_chat_completion(
162
163
        self,
        request: ChatCompletionRequest,
164
165
166
        raw_request: Optional[Request] = None,
    ) -> Union[AsyncGenerator[str, None], ChatCompletionResponse,
               ErrorResponse]:
167
168
        """
        Chat Completion API similar to OpenAI's API.
169

170
171
        See https://platform.openai.com/docs/api-reference/chat/create
        for the API specification. This API mimics the OpenAI
172
        Chat Completion API.
173
174
175
        """
        error_check_ret = await self._check_model(request)
        if error_check_ret is not None:
176
            logger.error("Error with model %s", error_check_ret)
177
178
            return error_check_ret

179
180
181
182
183
184
        # 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

185
        try:
186
187
            lora_request = self._maybe_get_adapters(
                request, supports_default_mm_loras=True)
188

189
            model_name = self._get_model_name(request.model, lora_request)
190

191
            tokenizer = await self.engine_client.get_tokenizer(lora_request)
192

193
194
            tool_parser = self.tool_parser

195
            if isinstance(tokenizer, MistralTokenizer):
196
197
198
                # 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`
199
                maybe_serialize_tool_calls(request)
200
                truncate_tool_call_ids(request)
201
                validate_request_params(request)
202

203
204
            if (request.tool_choice == "auto" and
                    not (self.enable_auto_tools and tool_parser is not None)
205
206
                    and not isinstance(tokenizer, MistralTokenizer)
                    and not self.use_harmony):
207
208
209
210
211
212
                # 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"
                )
213

214
215
216
            if (request.tools is None
                    or (request.tool_choice == "none"
                        and self.exclude_tools_when_tool_choice_none)):
217
218
219
                tool_dicts = None
            else:
                tool_dicts = [tool.model_dump() for tool in request.tools]
220

221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
            if not self.use_harmony:
                # Common case.
                (
                    conversation,
                    request_prompts,
                    engine_prompts,
                ) = await self._preprocess_chat(
                    request,
                    tokenizer,
                    request.messages,
                    chat_template=request.chat_template or self.chat_template,
                    chat_template_content_format=self.
                    chat_template_content_format,
                    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,
                )
            else:
                # For GPT-OSS.
                (
                    conversation,
                    request_prompts,
                    engine_prompts,
                ) = self._make_request_with_harmony(request)
250
251
        except (ValueError, TypeError, RuntimeError,
                jinja2.TemplateError) as e:
252
            logger.exception("Error in preprocessing prompt inputs")
253
            return self.create_error_response(f"{e} {e.__cause__}")
254

255
256
        request_id = "chatcmpl-" \
                     f"{self._base_request_id(raw_request, request.request_id)}"
257
258
259
260
261

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

262
        # Schedule the request and get the result generator.
263
        generators: list[AsyncGenerator[RequestOutput, None]] = []
264
        try:
265
266
            for i, engine_prompt in enumerate(engine_prompts):
                sampling_params: Union[SamplingParams, BeamSearchParams]
267
268
269
270
271
272
273
274
275
276

                if self.default_sampling_params is None:
                    self.default_sampling_params = {}

                max_tokens = get_max_tokens(
                    max_model_len=self.max_model_len,
                    request=request,
                    input_length=len(engine_prompt["prompt_token_ids"]),
                    default_sampling_params=self.default_sampling_params)

277
278
                if request.use_beam_search:
                    sampling_params = request.to_beam_search_params(
279
                        max_tokens, self.default_sampling_params)
280
281
                else:
                    sampling_params = request.to_sampling_params(
282
                        max_tokens, self.model_config.logits_processor_pattern,
283
                        self.default_sampling_params)
284
285
286
287

                self._log_inputs(request_id,
                                 request_prompts[i],
                                 params=sampling_params,
288
                                 lora_request=lora_request)
289
290
291
292
293
294
295
296
297

                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,
298
                        lora_request=lora_request,
299
300
301
302
303
304
305
306
307
308
309
310
                    )
                else:
                    generator = self.engine_client.generate(
                        engine_prompt,
                        sampling_params,
                        request_id,
                        lora_request=lora_request,
                        trace_headers=trace_headers,
                        priority=request.priority,
                    )

                generators.append(generator)
311
        except ValueError as e:
312
            # TODO: Use a vllm-specific Validation Error
313
314
            return self.create_error_response(str(e))

315
316
317
        assert len(generators) == 1
        result_generator, = generators

318
319
320
        # Streaming response
        if request.stream:
            return self.chat_completion_stream_generator(
321
322
323
324
325
326
327
328
                request,
                result_generator,
                request_id,
                model_name,
                conversation,
                tokenizer,
                request_metadata,
                enable_force_include_usage=self.enable_force_include_usage)
329

330
331
        try:
            return await self.chat_completion_full_generator(
332
333
                request, result_generator, request_id, model_name,
                conversation, tokenizer, request_metadata)
334
335
336
        except ValueError as e:
            # TODO: Use a vllm-specific Validation Error
            return self.create_error_response(str(e))
337
338
339
340

    def get_chat_request_role(self, request: ChatCompletionRequest) -> str:
        if request.add_generation_prompt:
            return self.response_role
341
        return request.messages[-1]["role"]
342

343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
    @staticmethod
    def _bracket_level(s: str, opening='{', closing='}') -> int:
        """
        Calculate the current level of nested brackets in a given string.
        """
        level = 0
        for char in s:
            if char == opening:
                level += 1
            elif char == closing:
                level -= 1
        return level

    @staticmethod
    def _filter_delta_text(delta_text: str,
                           previous_text: str) -> tuple[str, bool]:
        # remove last '},' of the tool definition stemming from the
        # "name"/"parameters" outer object or closing ']' of the tool list
        # count occurrences of opening and closing curly braces and
        # once level 0 is reached stop outputting text
        # if 0 is reached while parsing the delta_text we know the current
        # tool will finish in this current iteration
        bracket_level = OpenAIServingChat._bracket_level(previous_text)
        updated_delta, passed_zero = "", False
        for c in delta_text:
            if c == '{':
                bracket_level += 1
                passed_zero = bracket_level == 0
            elif c == '}':
                bracket_level -= 1
                passed_zero = bracket_level == 0

            if bracket_level != 0:
                updated_delta += c
            else:
                # if a comma is reached at level 0 we can stop
                if c == ',':
                    break
        return updated_delta, passed_zero

    def extract_tool_call_required_streaming(
        self,
        previous_text: str,
386
        current_text: Optional[str],
387
388
        delta_text: str,
        function_name_returned: bool,
389
        tool_call_idx: Optional[int] = None
390
    ) -> tuple[Optional[DeltaMessage], bool]:
391
392
393
        if current_text is None or current_text == "":
            # if the current text is empty, we cannot parse it
            return None, function_name_returned
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
        try:
            obj = partial_json_parser.loads(current_text)
        except partial_json_parser.core.exceptions.MalformedJSON:
            logger.debug('not enough tokens to parse into JSON yet')
            obj = None

        # check if the current text is a valid array
        # containing a partial tool calling object
        # if not repeat
        if obj is None or not isinstance(obj, list) or not len(obj) > 0:
            function_name_returned = False
            delta_message = None
        else:
            _, finishes_previous_tool = OpenAIServingChat._filter_delta_text(
                delta_text, previous_text)
            # take the last tool call from the generated list
            current_tool_call = obj[-1]

            # once parameters have been generated the name is complete as well
            if not finishes_previous_tool and ("name" not in current_tool_call
                                               or "parameters"
                                               not in current_tool_call):
                function_name_returned = False
                delta_message = None
            else:
                if not function_name_returned:
                    # get partly generated arguments from the latest tool call
                    param_match = re.search(r'.*"parameters":\s*(.*)',
                                            current_text)
                    arguments = param_match.group(1) if param_match else ""
                    arguments, _ = OpenAIServingChat._filter_delta_text(
                        arguments, previous_text)

                    # if this iteration finishes a previous tool call but a
                    # new incomplete tool is already generated, take the
                    # previous from the list
                    if (finishes_previous_tool
                            and "parameters" not in current_tool_call):
                        current_tool_call = obj[-2]

                    function_name_returned = True
435
436
437
438
                    tool_call_id = make_tool_call_id(
                        id_type=self.tool_call_id_type,
                        func_name=current_tool_call["name"],
                        idx=tool_call_idx)
439
                    delta_message = DeltaMessage(tool_calls=[
440
                        DeltaToolCall(id=tool_call_id,
441
442
443
                                      function=DeltaFunctionCall(
                                          name=current_tool_call["name"],
                                          arguments=arguments),
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
                                      index=len(obj) - 1,
                                      type="function")
                    ])

                else:
                    delta_text, _ = OpenAIServingChat._filter_delta_text(
                        delta_text, previous_text)

                    if delta_text != "":
                        delta_message = DeltaMessage(tool_calls=[
                            DeltaToolCall(
                                function=DeltaFunctionCall(
                                    # OpenAI API returns None
                                    # instead of name every time
                                    name=None,
                                    arguments=delta_text),
460
                                index=len(obj) - 1)
461
462
463
464
465
466
                        ])
                    else:
                        delta_message = None

        return delta_message, function_name_returned

467
    async def chat_completion_stream_generator(
468
469
470
471
        self,
        request: ChatCompletionRequest,
        result_generator: AsyncIterator[RequestOutput],
        request_id: str,
472
        model_name: str,
473
        conversation: list[ConversationMessage],
474
        tokenizer: AnyTokenizer,
475
        request_metadata: RequestResponseMetadata,
476
        enable_force_include_usage: bool,
477
    ) -> AsyncGenerator[str, None]:
478
        created_time = int(time.time())
479
        chunk_object_type: Final = "chat.completion.chunk"
480
        first_iteration = True
481
482

        # Send response for each token for each request.n (index)
483
484
485
        num_choices = 1 if request.n is None else request.n
        previous_num_tokens = [0] * num_choices
        finish_reason_sent = [False] * num_choices
486
        num_prompt_tokens = 0
487
        num_cached_tokens = None
488
489
490
491
492
        if self.use_harmony:
            harmony_parsers = [
                get_streamable_parser_for_assistant()
                for _ in range(num_choices)
            ]
493
494
495
496
497
498
499
500
501
502
503

        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))

504
        all_previous_token_ids: Optional[list[list[int]]]
505
        function_name_returned = [False] * num_choices
506
507
508
509
        if self.tool_call_id_type == 'kimi_k2':
            history_tool_call_cnt = get_history_tool_calls_cnt(conversation)
        else:
            history_tool_call_cnt = 0
510

511
512
513
        # Always track previous_texts for comprehensive output logging
        previous_texts = [""] * num_choices

514
515
        # Only one of these will be used, thus previous_texts and
        # all_previous_token_ids will not be used twice in the same iteration.
516
        if tool_choice_auto or self.reasoning_parser:
517
518
            # These are only required in "auto" tool choice case
            all_previous_token_ids = [[]] * num_choices
519
520
521
            # For reasoning parser and tool call all enabled
            added_content_delta_arr = [False] * num_choices
            reasoning_end_arr = [False] * num_choices
522
523
        elif request.tool_choice == "required":
            all_previous_token_ids = None
524
        else:
525
            all_previous_token_ids = None
526

527
        try:
528
            if self.reasoning_parser:
529
530
531
532
533
534
535
                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
536
537
538
        # Prepare the tool parser if it's needed
        try:
            if tool_choice_auto and self.tool_parser:
539
                tool_parsers: list[Optional[ToolParser]] = [
540
541
542
543
                    self.tool_parser(tokenizer)
                ] * num_choices
            else:
                tool_parsers = [None] * num_choices
544
        except Exception as e:
545
            logger.exception("Error in tool parser creation.")
546
547
548
549
550
            data = self.create_streaming_error_response(str(e))
            yield f"data: {data}\n\n"
            yield "data: [DONE]\n\n"
            return

551
552
        stream_options = request.stream_options
        if stream_options:
553
554
            include_usage = stream_options.include_usage \
                            or enable_force_include_usage
555
556
557
558
559
            include_continuous_usage = include_usage and \
                                       stream_options.continuous_usage_stats
        else:
            include_usage, include_continuous_usage = False, False

560
561
        try:
            async for res in result_generator:
562
563
                if res.prompt_token_ids is not None:
                    num_prompt_tokens = len(res.prompt_token_ids)
564
565
                    if res.encoder_prompt_token_ids is not None:
                        num_prompt_tokens += len(res.encoder_prompt_token_ids)
566

567
568
569
570
                # 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:
571
                    num_cached_tokens = res.num_cached_tokens
572
573
                    # Send first response for each request.n (index) with
                    # the role
574
                    role = self.get_chat_request_role(request)
575
576
577

                    # NOTE num_choices defaults to 1 so this usually executes
                    # once per request
578
                    for i in range(num_choices):
579
580
                        choice_data = ChatCompletionResponseStreamChoice(
                            index=i,
581
582
583
584
                            delta=DeltaMessage(
                                role=role,
                                content="",
                            ),
585
586
                            logprobs=None,
                            finish_reason=None)
587
588

                        # return prompt_token_ids at the first chunk ever
589
590
591
592
593
                        chunk = ChatCompletionStreamResponse(
                            id=request_id,
                            object=chunk_object_type,
                            created=created_time,
                            choices=[choice_data],
594
595
596
597
                            model=model_name,
                            prompt_token_ids=(res.prompt_token_ids
                                              if request.return_token_ids else
                                              None))
598

599
600
601
602
603
604
                        # 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)
605

606
607
608
                        data = chunk.model_dump_json(exclude_unset=True)
                        yield f"data: {data}\n\n"

609
610
                    # Send response to echo the input portion of the
                    # last message
611
                    if request.echo:
612
                        last_msg_content: Union[str, list[dict[str, str]]] = ""
613
614
615
                        if conversation and "content" in conversation[
                                -1] and conversation[-1].get("role") == role:
                            last_msg_content = conversation[-1]["content"] or ""
616
617

                        if last_msg_content:
618
                            for i in range(num_choices):
619
620
621
622
623
                                choice_data = (
                                    ChatCompletionResponseStreamChoice(
                                        index=i,
                                        delta=DeltaMessage(
                                            content=last_msg_content),
624
                                        logprobs=None,
625
                                        finish_reason=None))
626
627
628
629
630
631
                                chunk = ChatCompletionStreamResponse(
                                    id=request_id,
                                    object=chunk_object_type,
                                    created=created_time,
                                    choices=[choice_data],
                                    model=model_name)
632
633
634
635
636
                                if include_continuous_usage:
                                    chunk.usage = UsageInfo(
                                        prompt_tokens=num_prompt_tokens,
                                        completion_tokens=0,
                                        total_tokens=num_prompt_tokens)
637

638
639
640
641
642
643
644
                                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
645
                    tool_parser = tool_parsers[i]
646
647
648
649

                    if finish_reason_sent[i]:
                        continue

650
                    if request.logprobs and request.top_logprobs is not None:
651
                        assert output.logprobs is not None, (
652
                            "Did not output logprobs")
653
                        logprobs = self._create_chat_logprobs(
654
655
                            token_ids=output.token_ids,
                            top_logprobs=output.logprobs,
656
                            tokenizer=tokenizer,
657
                            num_output_top_logprobs=request.top_logprobs,
658
659
                            return_as_token_id=request.
                            return_tokens_as_token_ids,
660
661
662
663
                        )
                    else:
                        logprobs = None

664
665
666
667
                    if self.use_harmony:
                        harmony_parser = harmony_parsers[i]
                        for token_id in output.token_ids:
                            harmony_parser.process(token_id)
668
669
670
                        is_reasoning = \
                            harmony_parser.current_channel == "analysis"
                        if not request.include_reasoning and is_reasoning:
671
672
673
674
675
                            # Skip the reasoning content.
                            continue
                        delta_text = harmony_parser.last_content_delta or ""
                    else:
                        delta_text = output.text
676
677
678
679
680
681

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

682
                    delta_message: Optional[DeltaMessage]
683

684
                    # just update previous_texts and previous_token_ids
685
686
                    if ((tool_choice_auto or self.reasoning_parser)
                            and not self.use_harmony):
687
688
689
690
691
                        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
692
693
                        # avoid the None + list error.
                        if previous_token_ids:
694
                            current_token_ids = previous_token_ids + as_list(
695
696
                                output.token_ids)
                        else:
697
                            current_token_ids = as_list(output.token_ids)
698

699
                    if self.use_harmony:
700
                        if is_reasoning:
701
702
                            delta_message = DeltaMessage(
                                reasoning_content=delta_text)
703
704
                        else:
                            delta_message = DeltaMessage(content=delta_text)
705
                    # handle streaming deltas for tools with named tool_choice
706
                    elif tool_choice_function_name:
707
                        if (self.reasoning_parser and not reasoning_end_arr[i]
708
709
710
711
712
713
714
715
716
717
718
719
720
                                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,
                                ))
721
722
723
724
725
                            # When encountering think end id in delta_token_ids
                            # or think end id in prompt_token_ids
                            # i.e {"enable_thinking": False},
                            # set reasoning status to end.
                            # Only keep 'content', remove 'reasoning_content'.
726
                            if reasoning_parser.is_reasoning_end(
727
728
729
730
                                    as_list(output.token_ids)) or (
                                        res.prompt_token_ids
                                        and reasoning_parser.is_reasoning_end(
                                            res.prompt_token_ids)):
731
                                reasoning_end_arr[i] = True
732
733
734
735
736
737
738
739
                                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`
740
                            if self.reasoning_parser:
741
742
743
                                delta_text = previous_text + delta_text
                                current_text = ""

744
745
746
747
748
749
750
                            if function_name_returned[i]:
                                delta_tool_call = DeltaToolCall(
                                    function=DeltaFunctionCall(
                                        arguments=delta_text),
                                    index=i)
                            else:
                                delta_tool_call = DeltaToolCall(
751
                                    id=make_tool_call_id(),
752
753
754
755
756
757
758
                                    type="function",
                                    function=DeltaFunctionCall(
                                        name=tool_choice_function_name,
                                        arguments=delta_text),
                                    index=i)
                                function_name_returned[i] = True

759
                            delta_message = DeltaMessage(tool_calls=[
760
                                delta_tool_call,
761
762
                            ])

763
764
765
766
767
768
                    elif request.tool_choice == "required":
                        assert previous_texts is not None
                        previous_text = previous_texts[i]
                        current_text = previous_text + delta_text
                        fn_name_returned = function_name_returned[i]

769
770
771
772
773
774
775
776
                        if self.reasoning_parser:
                            _, content = \
                                reasoning_parser.extract_reasoning_content(
                                    current_text,
                                    request
                                )
                        else:
                            content = current_text
777
778
779
                        delta_message, function_name_returned[i] = (
                            self.extract_tool_call_required_streaming(
                                previous_text=previous_text,
780
                                current_text=content,
781
                                delta_text=delta_text,
782
783
784
785
786
                                function_name_returned=fn_name_returned,
                                tool_call_idx=history_tool_call_cnt))
                        if (delta_message and delta_message.tool_calls and
                                delta_message.tool_calls[0].id is not None):
                            history_tool_call_cnt += 1
787
788
789
790

                        # update the previous values for the next iteration
                        previous_texts[i] = current_text

791
792
                    # handle streaming deltas for tools with "auto" tool choice
                    # and reasoning parser
793
                    elif tool_choice_auto and self.reasoning_parser:
794
795
796
797
                        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
798
                        output_token_ids = as_list(output.token_ids)
799
800
801
802
803
804
805
806
807
                        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,
808
                                    output_token_ids,
809
                                ))
810
811
812
813
814
815
816
                            # When encountering think end id in prompt_token_ids
                            # i.e {"enable_thinking": False},
                            # set reasoning status to end.
                            # Remove the text and token ids related
                            # to 'reasoning_content'.
                            if res.prompt_token_ids and \
                                reasoning_parser.is_reasoning_end(
817
                                    res.prompt_token_ids):
818
                                reasoning_end_arr[i] = True
819
                                current_token_ids = output_token_ids
820
821
822
823
824
                                if delta_message and delta_message.content:
                                    current_text = delta_message.content
                                    delta_message.content = None
                                else:
                                    current_text = ""
825
826
827
828
829
                            # 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(
830
                                    output_token_ids):
831
832
833
                                reasoning_end_arr[i] = True
                                current_token_ids =  \
                                    reasoning_parser.extract_content_ids(
834
                                        output_token_ids)
835
836
837
838
839
840
841
842
                                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:
843
                            delta_token_ids = output_token_ids
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
                            # 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
866
867
                        delta_message = (
                            tool_parser.extract_tool_calls_streaming(
868
869
                                previous_text=previous_text,
                                current_text=current_text,
870
                                delta_text=delta_text,
871
872
                                previous_token_ids=previous_token_ids,
                                current_token_ids=current_token_ids,
873
874
                                delta_token_ids=output.token_ids,
                                request=request))
875

876
                    # when only reasoning
877
                    elif self.reasoning_parser:
878
879
880
881
882
883
884
885
886
                        delta_message = (reasoning_parser.
                                         extract_reasoning_content_streaming(
                                             previous_text,
                                             current_text,
                                             delta_text,
                                             previous_token_ids,
                                             current_token_ids,
                                             output.token_ids,
                                         ))
887
                    # handle streaming just a content delta
888
889
890
                    else:
                        delta_message = DeltaMessage(content=delta_text)

891
                    # update the previous values for the next iteration
892
893
                    if ((tool_choice_auto or self.reasoning_parser)
                            and not self.use_harmony):
894
895
896
897
                        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
898
899
900
901
                    else:
                        # Update for comprehensive logging even in simple case
                        assert previous_texts is not None
                        previous_texts[i] += delta_text
902

903
                    # set the previous values for the next iteration
904
                    previous_num_tokens[i] += len(output.token_ids)
905
906
907
908
909
910
911
912

                    # 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

913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
                    # Log streaming delta if output logging is enabled
                    if self.enable_log_outputs and self.request_logger:
                        delta_content = ""
                        if delta_message.content:
                            delta_content = delta_message.content
                        elif delta_message.tool_calls:
                            delta_content = "".join(
                                tc.function.arguments
                                for tc in delta_message.tool_calls
                                if tc.function and tc.function.arguments)

                        if delta_content:
                            self.request_logger.log_outputs(
                                request_id=request_id,
                                outputs=delta_content,
928
                                output_token_ids=as_list(output.token_ids),
929
930
931
932
933
                                finish_reason=output.finish_reason,
                                is_streaming=True,
                                delta=True,
                            )

934
935
936
937
                    if output.finish_reason is None:
                        # Send token-by-token response for each request.n
                        choice_data = ChatCompletionResponseStreamChoice(
                            index=i,
938
                            delta=delta_message,
939
                            logprobs=logprobs,
940
941
942
                            finish_reason=None,
                            token_ids=(as_list(output.token_ids)
                                       if request.return_token_ids else None))
943
944

                    # if the model is finished generating
945
                    else:
946
947
948
949
                        # 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
950
                        auto_tools_called = False
951
                        if tool_parser:
952
953
954
955
                            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
956
957
958
959
960
                        else:
                            index = 0

                        if self._should_check_for_unstreamed_tool_arg_tokens(
                                delta_message, output) and tool_parser:
961
962
963
964
965
966
967
968
969
970
                            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)

971
972
973
974
                            # 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(
975
976
                                    "arguments", {}),
                                ensure_ascii=False)
977

978
                            # get what we've streamed so far for arguments
979
980
981
                            # for the current tool
                            actual_call = tool_parser.streamed_args_for_tool[
                                index]
982
983
                            if (latest_delta_len > 0):
                                actual_call = actual_call[:-latest_delta_len]
984
985
986
987
988
989
990
991
992
993
994
995

                            # 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))
                            ])

996
997
998
                        # Send the finish response for each request.n only once
                        choice_data = ChatCompletionResponseStreamChoice(
                            index=i,
999
                            delta=delta_message,
1000
                            logprobs=logprobs,
1001
                            finish_reason=output.finish_reason
1002
                            if not auto_tools_called else "tool_calls",
1003
1004
1005
                            stop_reason=output.stop_reason,
                            token_ids=(as_list(output.token_ids)
                                       if request.return_token_ids else None))
1006

1007
                        finish_reason_sent[i] = True
1008

1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
                    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,
                        )

1025
                    data = chunk.model_dump_json(exclude_unset=True)
1026
1027
                    yield f"data: {data}\n\n"

1028
1029
            # once the final token is handled, if stream_options.include_usage
            # is sent, send the usage
1030
1031
            if include_usage:
                completion_tokens = sum(previous_num_tokens)
1032
1033
1034
1035
1036
1037
1038
                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)
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049

                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"
1050

1051
1052
1053
1054
1055
            # 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,
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
                total_tokens=num_prompt_tokens + num_completion_tokens,
            )

            # Log complete streaming response if output logging is enabled
            if self.enable_log_outputs and self.request_logger:
                # Log the complete response for each choice
                for i in range(num_choices):
                    full_text = (
                        previous_texts[i]
                        if previous_texts and i < len(previous_texts) else
                        f"<streaming_complete: {previous_num_tokens[i]} tokens>"
                    )
                    self.request_logger.log_outputs(
                        request_id=request_id,
                        outputs=full_text,
                        output_token_ids=
                        None,  # Consider also logging all token IDs
                        finish_reason="streaming_complete",
                        is_streaming=True,
                        delta=False,
                    )
1077

1078
        except Exception as e:
1079
            # TODO: Use a vllm-specific Validation Error
1080
            logger.exception("Error in chat completion stream generator.")
1081
1082
            data = self.create_streaming_error_response(str(e))
            yield f"data: {data}\n\n"
1083
1084
1085
1086
        # Send the final done message after all response.n are finished
        yield "data: [DONE]\n\n"

    async def chat_completion_full_generator(
1087
1088
1089
1090
        self,
        request: ChatCompletionRequest,
        result_generator: AsyncIterator[RequestOutput],
        request_id: str,
1091
        model_name: str,
1092
        conversation: list[ConversationMessage],
1093
        tokenizer: AnyTokenizer,
1094
        request_metadata: RequestResponseMetadata,
1095
    ) -> Union[ErrorResponse, ChatCompletionResponse]:
1096

1097
        created_time = int(time.time())
1098
        final_res: Optional[RequestOutput] = None
1099

1100
1101
1102
1103
1104
        try:
            async for res in result_generator:
                final_res = res
        except asyncio.CancelledError:
            return self.create_error_response("Client disconnected")
1105
1106
1107
        except ValueError as e:
            # TODO: Use a vllm-specific Validation Error
            return self.create_error_response(str(e))
1108

1109
1110
        assert final_res is not None

1111
        choices: list[ChatCompletionResponseChoice] = []
1112
1113
1114
1115
        if self.tool_call_id_type == 'kimi_k2':
            history_tool_call_cnt = get_history_tool_calls_cnt(conversation)
        else:
            history_tool_call_cnt = 0
1116

1117
1118
        role = self.get_chat_request_role(request)
        for output in final_res.outputs:
1119
            token_ids = output.token_ids
1120
            out_logprobs = output.logprobs
1121

1122
1123
            if request.logprobs and request.top_logprobs is not None:
                assert out_logprobs is not None, "Did not output logprobs"
1124
                logprobs = self._create_chat_logprobs(
1125
                    token_ids=token_ids,
1126
                    top_logprobs=out_logprobs,
1127
                    num_output_top_logprobs=request.top_logprobs,
1128
                    tokenizer=tokenizer,
1129
                    return_as_token_id=request.return_tokens_as_token_ids,
1130
1131
1132
                )
            else:
                logprobs = None
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164

            if self.use_harmony:
                reasoning_content, final_content, is_tool_call = (
                    parse_chat_output(token_ids))
                if not request.include_reasoning:
                    reasoning_content = None

                if is_tool_call:
                    # TODO(woosuk): Implement tool call for gpt-oss.
                    # For now, only Responses API supports tool call for
                    # gpt-oss.
                    raise NotImplementedError(
                        "Tool call in Chat Completion API is not supported "
                        "for gpt-oss yet. Please use Responses API instead.")
                else:
                    # Normal message
                    message = ChatMessage(
                        role=role,
                        reasoning_content=reasoning_content,
                        content=final_content,
                    )

                choice_data = ChatCompletionResponseChoice(
                    index=output.index,
                    message=message,
                    logprobs=logprobs,
                    finish_reason="tool_calls" if is_tool_call else
                    output.finish_reason if output.finish_reason else "stop",
                    stop_reason=output.stop_reason,
                )
                choices.append(choice_data)
                continue
1165

1166
            if self.reasoning_parser:
1167
1168
1169
1170
1171
                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))
1172
1173
                # If the reasoning parser is enabled,
                # tool calls are extracted exclusively from the content.
1174
1175
1176
                reasoning_content, content = (
                    reasoning_parser.extract_reasoning_content(
                        output.text, request=request))
1177
1178
                if not request.include_reasoning:
                    reasoning_content = None
1179
1180
1181
            else:
                reasoning_content = None
                content = output.text
1182

1183
            auto_tools_called = False
1184
1185
            # if auto tools are not enabled, and a named tool choice using
            #   outlines is not being used
1186
1187
1188
1189
            if (not self.enable_auto_tools or not self.tool_parser) and \
                (not isinstance(request.tool_choice,
                                ChatCompletionNamedToolChoiceParam
                                ) and request.tool_choice != "required"):
1190
1191
1192
                message = ChatMessage(role=role,
                                      reasoning_content=reasoning_content,
                                      content=content)
1193
1194
1195

            # if the request uses tools and specified a tool choice
            elif request.tool_choice and type(
1196
                    request.tool_choice) is ChatCompletionNamedToolChoiceParam:
1197

1198
1199
                tool_call_class = MistralToolCall if isinstance(
                    tokenizer, MistralTokenizer) else ToolCall
1200
1201
                message = ChatMessage(
                    role=role,
1202
                    reasoning_content=reasoning_content,
1203
1204
                    content="",
                    tool_calls=[
1205
                        tool_call_class(function=FunctionCall(
1206
                            name=request.tool_choice.function.name,
1207
1208
1209
1210
                            arguments=content,
                        ))
                    ],
                )
1211

1212
1213
1214
1215
1216
1217
            elif request.tool_choice and request.tool_choice == "required":
                tool_call_class = MistralToolCall if isinstance(
                    tokenizer, MistralTokenizer) else ToolCall

                # the fields of FunctionDefinition are a superset of the
                # tool call outputs and can be used for parsing
1218
                assert content is not None
1219
                tool_calls = TypeAdapter(
1220
                    list[FunctionDefinition]).validate_json(content)
1221
1222
1223
1224
1225
1226
1227
                tool_call_ids = []
                for tool_call in tool_calls:
                    tool_call_ids.append(
                        make_tool_call_id(id_type=self.tool_call_id_type,
                                          func_name=tool_call.name,
                                          idx=history_tool_call_cnt))
                    history_tool_call_cnt += 1
1228
1229
1230
1231
                message = ChatMessage(
                    role=role,
                    content="",
                    tool_calls=[
1232
1233
1234
1235
1236
1237
1238
1239
1240
                        tool_call_class(id=tool_call_ids[i],
                                        function=FunctionCall(
                                            name=tool_call.name,
                                            arguments=json.dumps(
                                                tool_call.parameters,
                                                ensure_ascii=False)))
                        for i, tool_call in enumerate(tool_calls)
                    ],
                    reasoning_content=reasoning_content)
1241

1242
1243
            # if the request doesn't use tool choice
            # OR specifies to not use a tool
1244
            elif not request.tool_choice or request.tool_choice == "none":
1245

1246
1247
1248
                message = ChatMessage(role=role,
                                      reasoning_content=reasoning_content,
                                      content=content)
1249
1250
1251
1252
1253
1254
1255

            # 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:

1256
1257
1258
                try:
                    tool_parser = self.tool_parser(tokenizer)
                except RuntimeError as e:
1259
                    logger.exception("Error in tool parser creation.")
1260
1261
                    return self.create_error_response(str(e))

1262
                tool_call_info = tool_parser.extract_tool_calls(
1263
                    content if content is not None else "", request=request)
1264
1265
1266
1267
                # 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
1268
1269
                if tool_call_info.tools_called:
                    message = ChatMessage(role=role,
1270
                                          reasoning_content=reasoning_content,
1271
1272
1273
1274
1275
1276
                                          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.
1277
1278
1279
1280
1281
1282
1283
                    ret_content = content

                    # try to use content return from tool parser first,
                    # tool parser may do some modify for the content.
                    if (tool_call_info.content
                            and len(tool_call_info.content) > 0):
                        ret_content = tool_call_info.content
1284
1285
                    message = ChatMessage(role=role,
                                          reasoning_content=reasoning_content,
1286
                                          content=ret_content)
1287
1288
1289
1290
1291
1292
1293

            # 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.")
1294
1295
1296
                message = ChatMessage(role=role,
                                      reasoning_content=reasoning_content,
                                      content=content)
1297

1298
1299
            choice_data = ChatCompletionResponseChoice(
                index=output.index,
1300
                message=message,
1301
                logprobs=logprobs,
1302
                finish_reason="tool_calls" if auto_tools_called else
1303
                output.finish_reason if output.finish_reason else "stop",
1304
1305
1306
1307
                stop_reason=output.stop_reason,
                token_ids=(as_list(output.token_ids)
                           if request.return_token_ids else None),
            )
1308

1309
1310
            choices.append(choice_data)

1311
        if request.echo:
1312
            last_msg_content: Union[str, list[dict[str, str]]] = ""
1313
1314
            if (conversation and "content" in conversation[-1]
                    and conversation[-1].get("role") == role):
1315
                last_msg_content = conversation[-1]["content"] or ""
1316
1317
1318
            if isinstance(last_msg_content, list):
                last_msg_content = "\n".join(msg['text']
                                             for msg in last_msg_content)
1319
1320

            for choice in choices:
1321
1322
                full_message = last_msg_content + (choice.message.content
                                                   or "")
1323
1324
                choice.message.content = full_message

1325
        assert final_res.prompt_token_ids is not None
1326
        num_prompt_tokens = len(final_res.prompt_token_ids)
1327
1328
        if final_res.encoder_prompt_token_ids is not None:
            num_prompt_tokens += len(final_res.encoder_prompt_token_ids)
1329
1330
        num_generated_tokens = sum(
            len(output.token_ids) for output in final_res.outputs)
1331
1332
1333
1334
1335
1336
1337
        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)
1338
1339
1340

        request_metadata.final_usage_info = usage

1341
1342
1343
1344
1345
1346
        response = ChatCompletionResponse(
            id=request_id,
            created=created_time,
            model=model_name,
            choices=choices,
            usage=usage,
1347
            prompt_logprobs=clamp_prompt_logprobs(final_res.prompt_logprobs),
1348
1349
            prompt_token_ids=(final_res.prompt_token_ids
                              if request.return_token_ids else None),
Robert Shaw's avatar
Robert Shaw committed
1350
            kv_transfer_params=final_res.kv_transfer_params,
1351
1352
        )

1353
1354
1355
1356
1357
1358
1359
1360
1361
        # Log complete response if output logging is enabled
        if self.enable_log_outputs and self.request_logger:
            for choice in choices:
                output_text = ""
                if choice.message.content:
                    output_text = choice.message.content
                elif choice.message.tool_calls:
                    # For tool calls, log the function name and arguments
                    tool_call_descriptions = []
1362
1363
1364
                    for tc in choice.message.tool_calls:
                        if hasattr(tc.function, "name") and hasattr(
                                tc.function, "arguments"):
1365
                            tool_call_descriptions.append(
1366
                                f"{tc.function.name}({tc.function.arguments})")
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
                    tool_calls_str = ", ".join(tool_call_descriptions)
                    output_text = f"[tool_calls: {tool_calls_str}]"

                if output_text:
                    # Get the corresponding output token IDs
                    output_token_ids = None
                    if choice.index < len(final_res.outputs):
                        output_token_ids = final_res.outputs[
                            choice.index].token_ids

                    self.request_logger.log_outputs(
                        request_id=request_id,
                        outputs=output_text,
                        output_token_ids=output_token_ids,
                        finish_reason=choice.finish_reason,
                        is_streaming=False,
                        delta=False,
                    )

1386
        return response
1387
1388

    def _get_top_logprobs(
1389
            self, logprobs: dict[int, Logprob], top_logprobs: Optional[int],
1390
1391
            tokenizer: AnyTokenizer,
            should_return_as_token_id: bool) -> list[ChatCompletionLogProb]:
1392
        return [
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
            ChatCompletionLogProb(
                token=(token := self._get_decoded_token(
                    p[1],
                    p[0],
                    tokenizer,
                    return_as_token_id=should_return_as_token_id,
                )),
                logprob=max(p[1].logprob, -9999.0),
                bytes=list(token.encode("utf-8", errors="replace")),
            ) for i, p in enumerate(logprobs.items())
1403
1404
1405
1406
1407
1408
            if top_logprobs and i < top_logprobs
        ]

    def _create_chat_logprobs(
        self,
        token_ids: GenericSequence[int],
1409
        top_logprobs: GenericSequence[Optional[dict[int, Logprob]]],
1410
        tokenizer: AnyTokenizer,
1411
        num_output_top_logprobs: Optional[int] = None,
1412
        return_as_token_id: Optional[bool] = None,
1413
1414
    ) -> ChatCompletionLogProbs:
        """Create OpenAI-style logprobs."""
1415
        logprobs_content: list[ChatCompletionLogProbsContent] = []
1416

1417
1418
        should_return_as_token_id = return_as_token_id if \
            return_as_token_id is not None else self.return_tokens_as_token_ids
1419
1420
        for i, token_id in enumerate(token_ids):
            step_top_logprobs = top_logprobs[i]
1421
1422
            if step_top_logprobs is None or step_top_logprobs.get(
                    token_id) is None:
1423
                token = tokenizer.decode(token_id)
1424
                if should_return_as_token_id:
1425
                    token = f"token_id:{token_id}"
1426

1427
1428
                logprobs_content.append(
                    ChatCompletionLogProbsContent(
1429
                        token=token,
1430
1431
                        bytes=list(token.encode("utf-8", errors="replace")),
                    ))
1432
            else:
1433
1434
1435
                step_token = step_top_logprobs[token_id]
                step_decoded = step_token.decoded_token

1436
1437
                logprobs_content.append(
                    ChatCompletionLogProbsContent(
1438
                        token=self._get_decoded_token(
1439
1440
1441
                            step_token,
                            token_id,
                            tokenizer,
1442
                            should_return_as_token_id,
1443
1444
1445
1446
                        ),
                        logprob=max(step_token.logprob, -9999.0),
                        bytes=None if step_decoded is None else list(
                            step_decoded.encode("utf-8", errors="replace")),
1447
                        top_logprobs=self._get_top_logprobs(
1448
1449
                            step_top_logprobs, num_output_top_logprobs,
                            tokenizer, should_return_as_token_id),
1450
                    ))
1451
1452

        return ChatCompletionLogProbs(content=logprobs_content)
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481

    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])

    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
1482
1483
            output.finish_reason is not None
            and self.enable_auto_tools and self.tool_parser and delta_message
1484
1485
1486
1487
            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
        )
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516

    def _make_request_with_harmony(
        self,
        request: ChatCompletionRequest,
    ):
        messages: list[OpenAIMessage] = []

        # Add system message.
        # NOTE: In Chat Completion API, browsing is enabled by default
        # if the model supports it. TODO: Support browsing.
        assert not self.supports_browsing
        assert not self.supports_code_interpreter
        sys_msg = get_system_message(
            reasoning_effort=request.reasoning_effort,
            browser_description=None,
            python_description=None)
        messages.append(sys_msg)

        # Add developer message.
        dev_msg = get_developer_message()
        messages.append(dev_msg)

        # Add user message.
        for chat_msg in request.messages:
            messages.append(parse_chat_input(chat_msg))

        # Render prompt token ids.
        prompt_token_ids = render_for_completion(messages)
        engine_prompt = EngineTokensPrompt(prompt_token_ids=prompt_token_ids)
1517
1518
1519
1520
1521

        # Add cache_salt if provided in the request
        if request.cache_salt is not None:
            engine_prompt["cache_salt"] = request.cache_salt

1522
        return messages, [prompt_token_ids], [engine_prompt]