entrypoints/openai/chat_completion/serving.py: OpenAIServingChat

        self.enable_auto_tools: bool = enable_auto_tools
        self._include_reasoning_tokens_details = bool(reasoning_parser)
        self.parser_cls = ParserManager.get_parser(
            tool_parser_name=tool_parser,
            reasoning_parser_name=reasoning_parser,
            enable_auto_tools=enable_auto_tools,
            tool_strict_level=tool_strict_level,
            model_name=self.model_config.model,
            is_harmony=self.model_config.hf_config.model_type == "gpt_oss",
            tokenizer=self.renderer.tokenizer,
        )
        self.exclude_tools_when_tool_choice_none = exclude_tools_when_tool_choice_none

        self.enable_prompt_tokens_details = enable_prompt_tokens_details
        self.enable_force_include_usage = enable_force_include_usage
        self.enable_per_request_metrics = enable_per_request_metrics
        self.default_sampling_params = self.model_config.get_diff_sampling_param()
        mc = self.model_config
        self.override_max_tokens = (
            self.default_sampling_params.get("max_tokens")
            if mc.generation_config not in ("auto", "vllm")
            else getattr(mc, "override_generation_config", {}).get("max_new_tokens")
        )

    async def render_chat_request(
        self,
        request: ChatCompletionRequest,
    ) -> tuple[list[ConversationMessage], list[EngineInput]] | ErrorResponse:
        """Validate the model and preprocess a chat completion request.

        Delegates preprocessing logic to OnlineRenderer, adding the
        engine-aware checks (LoRA model validation, engine health).

        Returns:
            A tuple of (conversation, engine_inputs) on success,
            or an ErrorResponse on failure.

        """
        error_check_ret = await self._check_model(request)
        if error_check_ret is not None:
            logger.error("Error with model %s", error_check_ret)
            return error_check_ret

        self._preflight(request.n or 1)

        return await self.online_renderer.render_chat(request)

        tokenizer = self.renderer.tokenizer
        assert tokenizer is not None
        chat_template_kwargs = self._effective_chat_template_kwargs(request)
        parser: Parser | None = None
        if self.parser_cls is not None:
            parser = self._make_parser(request, tokenizer, chat_template_kwargs)
        result = await self.render_chat_request(request)
        if isinstance(result, ErrorResponse):
            return result

        conversation, engine_inputs = result

        request_id = (
            f"chatcmpl-{self._base_request_id(raw_request, request.request_id)}"
        )

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

        lora_request = self._maybe_get_adapters(request, supports_default_mm_loras=True)

        model_name = self.models.model_name(lora_request)

        # Extract data_parallel_rank from header (router can inject it)
        data_parallel_rank = self._get_data_parallel_rank(raw_request)

        for i, engine_input in enumerate(engine_inputs):
            prompt_token_ids = self._extract_prompt_components(engine_input).token_ids
            mm_token_counts = _get_mm_token_counts(engine_input)

            # If we are creating sub requests for multiple prompts, ensure that they
            # have unique request ids.
            sub_request_id = (
                request_id if len(engine_inputs) == 1 else f"{request_id}_{i}"
            )

            max_tokens = get_max_tokens(
                max_model_len,
                request.max_completion_tokens
                if request.max_completion_tokens is not None
                else request.max_tokens,
                self._extract_prompt_len(engine_input),
                self.default_sampling_params,
                self.override_max_tokens,
                truncate_prompt_tokens=request.truncate_prompt_tokens,
            )

            sampling_params: SamplingParams | BeamSearchParams
            if request.use_beam_search:
                sampling_params = request.to_beam_search_params(
                    max_tokens, self.default_sampling_params
                )
            else:
                sampling_params = request.to_sampling_params(
                    max_tokens,
                    self.default_sampling_params,
                )

            else:
                generator = self.engine_client.generate(
                    engine_input,
                    sampling_params,
                    sub_request_id,
                    lora_request=lora_request,
                    trace_headers=trace_headers,
                    priority=self._get_priority(request, raw_request),
                    data_parallel_rank=data_parallel_rank,
                    session_id=session_id,
                    **self._engine_reasoning_kwargs(
                        request, parser, chat_template_kwargs, prompt_token_ids
                    ),
                )

            generators.append(generator)

        assert len(generators) == 1
        (result_generator,) = generators

        if request.stream:
            return self.chat_completion_stream_generator(
                request,
                result_generator,
                request_id,
                model_name,
                conversation,
                tokenizer,
                request_metadata,
                chat_template_kwargs=chat_template_kwargs,
                mm_token_counts=mm_token_counts,
            )

        return await self.chat_completion_full_generator(
            request,
            result_generator,
            request_id,
            model_name,
            conversation,
            tokenizer,
            request_metadata,
            parser=parser,
            mm_token_counts=mm_token_counts,
        )

        try:
            async for chunk in self.chat_completion_stream_generator_raw(
                request,
                result_generator,
                request_id,
                model_name,
                conversation,
                tokenizer,
                request_metadata,
                chat_template_kwargs=chat_template_kwargs,
                mm_token_counts=mm_token_counts,
            ):
                if isinstance(chunk, ErrorResponse):
                    data = chunk.model_dump_json()
                else:
                    data = chunk.model_dump_json(
                        exclude_unset=True, exclude_none=not chunk.choices
                    )
                yield f"data: {data}\n\n"
        except Exception as e:
            logger.exception("Error serializing chat completion stream.")
            data = self.create_streaming_error_response(e)
            yield f"data: {data}\n\n"
        yield "data: [DONE]\n\n"

                for output in res.outputs:
                    i = output.index
                    processor = processors[i]
                    parser = processor.parser
                    if finish_reason_sent[i]:
                        continue

                    self._raise_if_error(output.finish_reason, request_id)

                    if request.logprobs:
                        assert output.logprobs is not None, "Did not output logprobs"
                        logprobs = self._create_chat_logprobs(
                            token_ids=output.token_ids,
                            top_logprobs=output.logprobs,
                            tokenizer=tokenizer,
                            num_output_top_logprobs=request.top_logprobs,
                            logprob_token_ids=request.logprob_token_ids,
                            return_as_token_id=request.return_tokens_as_token_ids,
                        )
                    else:
                        logprobs = None

                    delta_text = output.text

                    if (
                        not delta_text
                        and not output.token_ids
                        and not processor.num_output_tokens
                        and output.finish_reason is None
                    ):
                        # Chunked prefill case, don't return empty chunks
                        continue

                    delta_message, finish_reason = processor.step(
                        delta_text,
                        as_list(output.token_ids),
                        output.finish_reason,
                        res.prompt_token_ids,
                    )

            # once the final token is handled, if stream_options.include_usage
            # is sent, send the usage
            if include_usage:
                completion_tokens = sum(p.num_output_tokens for p in processors)
                final_usage = UsageInfo(
                    prompt_tokens=num_prompt_tokens,
                    completion_tokens=completion_tokens,
                    total_tokens=num_prompt_tokens + completion_tokens,
                    completion_tokens_details=_make_completion_tokens_details(
                        sum(previous_reasoning_tokens)
                    )
                    if self._include_reasoning_tokens_details
                    else None,
                )
                final_usage.prompt_tokens_details = _make_prompt_tokens_details(
                    self.enable_prompt_tokens_details,
                    num_cached_tokens,
                    num_cache_creation_tokens,
                    mm_token_counts,
                )

        created_time = int(time.time())
        final_res: RequestOutput | None = None

        try:
            async for res in result_generator:
                final_res = res
        except asyncio.CancelledError:
            return self.create_error_response("Client disconnected")

        if final_res is None:
            return self.create_error_response(
                "No output received from the engine.",
                err_type="InternalServerError",
                status_code=HTTPStatus.INTERNAL_SERVER_ERROR,
            )

        choices: list[ChatCompletionResponseChoice] = []
        total_reasoning_tokens = 0

        role = self.get_chat_request_role(request)
        tool_parser_cls = (
            self.parser_cls.tool_parser_cls if self.parser_cls is not None else None
        )
        for i, output in enumerate(final_res.outputs):
            # check for error finish reason and raise GenerationError
            # finish_reason='error' indicates a retryable request-level internal error
            self._raise_if_error(output.finish_reason, request_id)
            if i > 0 and parser is not None:
                # Parsers are stateful: like streaming, use one per choice.
                parser = self._make_parser(
                    request, tokenizer, self._effective_chat_template_kwargs(request)
                )
            if parser is not None and final_res.prompt_token_ids is not None:
                parser.set_prompt_token_ids(final_res.prompt_token_ids)
            token_ids = output.token_ids
            out_logprobs = output.logprobs

            if request.logprobs:
                assert out_logprobs is not None, "Did not output logprobs"
                logprobs = self._create_chat_logprobs(
                    token_ids=token_ids,
                    top_logprobs=out_logprobs,
                    num_output_top_logprobs=request.top_logprobs,
                    logprob_token_ids=request.logprob_token_ids,
                    tokenizer=tokenizer,
                    return_as_token_id=request.return_tokens_as_token_ids,
                )
            else:
                logprobs = None

            if parser is not None:
                reasoning, content, tool_calls = parser.parse(
                    output.text,
                    request,
                    enable_auto_tools=self.enable_auto_tools,
                    model_output_token_ids=token_ids,
                )
                suppress_metadata = not request.include_reasoning and parser is not None
                if not request.include_reasoning:
                    reasoning = None
                total_reasoning_tokens += parser.count_reasoning_tokens(token_ids)
                if suppress_metadata:
                    logprobs = None
            else:
                reasoning = None
                content = output.text
                tool_calls = []
                suppress_metadata = False